This is a live mirror of the Perl 5 development currently hosted at https://github.com/perl/perl5
regen/regcharclass.pl: Generate macros for multi-char fold sequences
[perl5.git] / regcomp.c
CommitLineData
a0d0e21e
LW
1/* regcomp.c
2 */
3
4/*
4ac71550
TC
5 * 'A fair jaw-cracker dwarf-language must be.' --Samwise Gamgee
6 *
7 * [p.285 of _The Lord of the Rings_, II/iii: "The Ring Goes South"]
a0d0e21e
LW
8 */
9
61296642
DM
10/* This file contains functions for compiling a regular expression. See
11 * also regexec.c which funnily enough, contains functions for executing
166f8a29 12 * a regular expression.
e4a054ea
DM
13 *
14 * This file is also copied at build time to ext/re/re_comp.c, where
15 * it's built with -DPERL_EXT_RE_BUILD -DPERL_EXT_RE_DEBUG -DPERL_EXT.
16 * This causes the main functions to be compiled under new names and with
17 * debugging support added, which makes "use re 'debug'" work.
166f8a29
DM
18 */
19
a687059c
LW
20/* NOTE: this is derived from Henry Spencer's regexp code, and should not
21 * confused with the original package (see point 3 below). Thanks, Henry!
22 */
23
24/* Additional note: this code is very heavily munged from Henry's version
25 * in places. In some spots I've traded clarity for efficiency, so don't
26 * blame Henry for some of the lack of readability.
27 */
28
e50aee73 29/* The names of the functions have been changed from regcomp and
3b753521 30 * regexec to pregcomp and pregexec in order to avoid conflicts
e50aee73
AD
31 * with the POSIX routines of the same names.
32*/
33
b9d5759e 34#ifdef PERL_EXT_RE_BUILD
54df2634 35#include "re_top.h"
b81d288d 36#endif
56953603 37
a687059c 38/*
e50aee73 39 * pregcomp and pregexec -- regsub and regerror are not used in perl
a687059c
LW
40 *
41 * Copyright (c) 1986 by University of Toronto.
42 * Written by Henry Spencer. Not derived from licensed software.
43 *
44 * Permission is granted to anyone to use this software for any
45 * purpose on any computer system, and to redistribute it freely,
46 * subject to the following restrictions:
47 *
48 * 1. The author is not responsible for the consequences of use of
49 * this software, no matter how awful, even if they arise
50 * from defects in it.
51 *
52 * 2. The origin of this software must not be misrepresented, either
53 * by explicit claim or by omission.
54 *
55 * 3. Altered versions must be plainly marked as such, and must not
56 * be misrepresented as being the original software.
57 *
58 *
59 **** Alterations to Henry's code are...
60 ****
4bb101f2 61 **** Copyright (C) 1991, 1992, 1993, 1994, 1995, 1996, 1997, 1998, 1999,
1129b882
NC
62 **** 2000, 2001, 2002, 2003, 2004, 2005, 2006, 2007, 2008
63 **** by Larry Wall and others
a687059c 64 ****
9ef589d8
LW
65 **** You may distribute under the terms of either the GNU General Public
66 **** License or the Artistic License, as specified in the README file.
67
a687059c
LW
68 *
69 * Beware that some of this code is subtly aware of the way operator
70 * precedence is structured in regular expressions. Serious changes in
71 * regular-expression syntax might require a total rethink.
72 */
73#include "EXTERN.h"
864dbfa3 74#define PERL_IN_REGCOMP_C
a687059c 75#include "perl.h"
d06ea78c 76
acfe0abc 77#ifndef PERL_IN_XSUB_RE
d06ea78c
GS
78# include "INTERN.h"
79#endif
c277df42
IZ
80
81#define REG_COMP_C
54df2634
NC
82#ifdef PERL_IN_XSUB_RE
83# include "re_comp.h"
40b6423c 84extern const struct regexp_engine my_reg_engine;
54df2634
NC
85#else
86# include "regcomp.h"
87#endif
a687059c 88
04e98a4d 89#include "dquote_static.c"
26faadbd 90#include "charclass_invlists.h"
81e983c1 91#include "inline_invlist.c"
1b0f46bf 92#include "unicode_constants.h"
04e98a4d 93
963df1c9
KW
94#ifdef HAS_ISBLANK
95# define hasISBLANK 1
96#else
97# define hasISBLANK 0
98#endif
99
94dc5c2d 100#define HAS_NONLATIN1_FOLD_CLOSURE(i) _HAS_NONLATIN1_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)
26faadbd 101#define IS_NON_FINAL_FOLD(c) _IS_NON_FINAL_FOLD_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
94dc5c2d 102
d4cce5f1 103#ifdef op
11343788 104#undef op
d4cce5f1 105#endif /* op */
11343788 106
fe14fcc3 107#ifdef MSDOS
7e4e8c89 108# if defined(BUGGY_MSC6)
fe14fcc3 109 /* MSC 6.00A breaks on op/regexp.t test 85 unless we turn this off */
7e4e8c89 110# pragma optimize("a",off)
fe14fcc3 111 /* But MSC 6.00A is happy with 'w', for aliases only across function calls*/
7e4e8c89
NC
112# pragma optimize("w",on )
113# endif /* BUGGY_MSC6 */
fe14fcc3
LW
114#endif /* MSDOS */
115
a687059c
LW
116#ifndef STATIC
117#define STATIC static
118#endif
119
b1603ef8 120
830247a4 121typedef struct RExC_state_t {
514a91f1
DM
122 U32 flags; /* RXf_* are we folding, multilining? */
123 U32 pm_flags; /* PMf_* stuff from the calling PMOP */
830247a4 124 char *precomp; /* uncompiled string. */
288b8c02 125 REGEXP *rx_sv; /* The SV that is the regexp. */
f8fc2ecf
YO
126 regexp *rx; /* perl core regexp structure */
127 regexp_internal *rxi; /* internal data for regexp object pprivate field */
fac92740 128 char *start; /* Start of input for compile */
830247a4
IZ
129 char *end; /* End of input for compile */
130 char *parse; /* Input-scan pointer. */
131 I32 whilem_seen; /* number of WHILEM in this expr */
fac92740 132 regnode *emit_start; /* Start of emitted-code area */
3b57cd43 133 regnode *emit_bound; /* First regnode outside of the allocated space */
ffc61ed2 134 regnode *emit; /* Code-emit pointer; &regdummy = don't = compiling */
830247a4
IZ
135 I32 naughty; /* How bad is this pattern? */
136 I32 sawback; /* Did we see \1, ...? */
137 U32 seen;
138 I32 size; /* Code size. */
c74340f9
YO
139 I32 npar; /* Capture buffer count, (OPEN). */
140 I32 cpar; /* Capture buffer count, (CLOSE). */
e2e6a0f1 141 I32 nestroot; /* root parens we are in - used by accept */
830247a4
IZ
142 I32 extralen;
143 I32 seen_zerolen;
40d049e4
YO
144 regnode **open_parens; /* pointers to open parens */
145 regnode **close_parens; /* pointers to close parens */
146 regnode *opend; /* END node in program */
02daf0ab
YO
147 I32 utf8; /* whether the pattern is utf8 or not */
148 I32 orig_utf8; /* whether the pattern was originally in utf8 */
149 /* XXX use this for future optimisation of case
150 * where pattern must be upgraded to utf8. */
e40e74fe
KW
151 I32 uni_semantics; /* If a d charset modifier should use unicode
152 rules, even if the pattern is not in
153 utf8 */
81714fb9 154 HV *paren_names; /* Paren names */
1f1031fe 155
40d049e4
YO
156 regnode **recurse; /* Recurse regops */
157 I32 recurse_count; /* Number of recurse regops */
b57e4118 158 I32 in_lookbehind;
4624b182 159 I32 contains_locale;
bb3f3ed2 160 I32 override_recoding;
3d2bd50a 161 struct reg_code_block *code_blocks; /* positions of literal (?{})
68e2671b 162 within pattern */
b1603ef8
DM
163 int num_code_blocks; /* size of code_blocks[] */
164 int code_index; /* next code_blocks[] slot */
830247a4
IZ
165#if ADD_TO_REGEXEC
166 char *starttry; /* -Dr: where regtry was called. */
167#define RExC_starttry (pRExC_state->starttry)
168#endif
d24ca0c5 169 SV *runtime_code_qr; /* qr with the runtime code blocks */
3dab1dad 170#ifdef DEBUGGING
be8e71aa 171 const char *lastparse;
3dab1dad 172 I32 lastnum;
1f1031fe 173 AV *paren_name_list; /* idx -> name */
3dab1dad
YO
174#define RExC_lastparse (pRExC_state->lastparse)
175#define RExC_lastnum (pRExC_state->lastnum)
1f1031fe 176#define RExC_paren_name_list (pRExC_state->paren_name_list)
3dab1dad 177#endif
830247a4
IZ
178} RExC_state_t;
179
e2509266 180#define RExC_flags (pRExC_state->flags)
514a91f1 181#define RExC_pm_flags (pRExC_state->pm_flags)
830247a4 182#define RExC_precomp (pRExC_state->precomp)
288b8c02 183#define RExC_rx_sv (pRExC_state->rx_sv)
830247a4 184#define RExC_rx (pRExC_state->rx)
f8fc2ecf 185#define RExC_rxi (pRExC_state->rxi)
fac92740 186#define RExC_start (pRExC_state->start)
830247a4
IZ
187#define RExC_end (pRExC_state->end)
188#define RExC_parse (pRExC_state->parse)
189#define RExC_whilem_seen (pRExC_state->whilem_seen)
7122b237
YO
190#ifdef RE_TRACK_PATTERN_OFFSETS
191#define RExC_offsets (pRExC_state->rxi->u.offsets) /* I am not like the others */
192#endif
830247a4 193#define RExC_emit (pRExC_state->emit)
fac92740 194#define RExC_emit_start (pRExC_state->emit_start)
3b57cd43 195#define RExC_emit_bound (pRExC_state->emit_bound)
830247a4
IZ
196#define RExC_naughty (pRExC_state->naughty)
197#define RExC_sawback (pRExC_state->sawback)
198#define RExC_seen (pRExC_state->seen)
199#define RExC_size (pRExC_state->size)
200#define RExC_npar (pRExC_state->npar)
e2e6a0f1 201#define RExC_nestroot (pRExC_state->nestroot)
830247a4
IZ
202#define RExC_extralen (pRExC_state->extralen)
203#define RExC_seen_zerolen (pRExC_state->seen_zerolen)
1aa99e6b 204#define RExC_utf8 (pRExC_state->utf8)
e40e74fe 205#define RExC_uni_semantics (pRExC_state->uni_semantics)
02daf0ab 206#define RExC_orig_utf8 (pRExC_state->orig_utf8)
40d049e4
YO
207#define RExC_open_parens (pRExC_state->open_parens)
208#define RExC_close_parens (pRExC_state->close_parens)
209#define RExC_opend (pRExC_state->opend)
81714fb9 210#define RExC_paren_names (pRExC_state->paren_names)
40d049e4
YO
211#define RExC_recurse (pRExC_state->recurse)
212#define RExC_recurse_count (pRExC_state->recurse_count)
b57e4118 213#define RExC_in_lookbehind (pRExC_state->in_lookbehind)
4624b182 214#define RExC_contains_locale (pRExC_state->contains_locale)
bb3f3ed2 215#define RExC_override_recoding (pRExC_state->override_recoding)
830247a4 216
cde0cee5 217
a687059c
LW
218#define ISMULT1(c) ((c) == '*' || (c) == '+' || (c) == '?')
219#define ISMULT2(s) ((*s) == '*' || (*s) == '+' || (*s) == '?' || \
220 ((*s) == '{' && regcurly(s)))
a687059c 221
35c8bce7
LW
222#ifdef SPSTART
223#undef SPSTART /* dratted cpp namespace... */
224#endif
a687059c
LW
225/*
226 * Flags to be passed up and down.
227 */
a687059c 228#define WORST 0 /* Worst case. */
a3b492c3 229#define HASWIDTH 0x01 /* Known to match non-null strings. */
fda99bee 230
18890cbd 231/* Simple enough to be STAR/PLUS operand; in an EXACT node must be a single
cf8c372d
KW
232 * character, and if utf8, must be invariant. Note that this is not the same
233 * thing as REGNODE_SIMPLE */
fda99bee 234#define SIMPLE 0x02
a3b492c3
YO
235#define SPSTART 0x04 /* Starts with * or +. */
236#define TRYAGAIN 0x08 /* Weeded out a declaration. */
237#define POSTPONED 0x10 /* (?1),(?&name), (??{...}) or similar */
a687059c 238
3dab1dad
YO
239#define REG_NODE_NUM(x) ((x) ? (int)((x)-RExC_emit_start) : -1)
240
07be1b83
YO
241/* whether trie related optimizations are enabled */
242#if PERL_ENABLE_EXTENDED_TRIE_OPTIMISATION
243#define TRIE_STUDY_OPT
786e8c11 244#define FULL_TRIE_STUDY
07be1b83
YO
245#define TRIE_STCLASS
246#endif
1de06328
YO
247
248
40d049e4
YO
249
250#define PBYTE(u8str,paren) ((U8*)(u8str))[(paren) >> 3]
251#define PBITVAL(paren) (1 << ((paren) & 7))
252#define PAREN_TEST(u8str,paren) ( PBYTE(u8str,paren) & PBITVAL(paren))
253#define PAREN_SET(u8str,paren) PBYTE(u8str,paren) |= PBITVAL(paren)
254#define PAREN_UNSET(u8str,paren) PBYTE(u8str,paren) &= (~PBITVAL(paren))
255
bbd61b5f
KW
256/* If not already in utf8, do a longjmp back to the beginning */
257#define UTF8_LONGJMP 42 /* Choose a value not likely to ever conflict */
258#define REQUIRE_UTF8 STMT_START { \
259 if (! UTF) JMPENV_JUMP(UTF8_LONGJMP); \
260 } STMT_END
40d049e4 261
1de06328
YO
262/* About scan_data_t.
263
264 During optimisation we recurse through the regexp program performing
265 various inplace (keyhole style) optimisations. In addition study_chunk
266 and scan_commit populate this data structure with information about
267 what strings MUST appear in the pattern. We look for the longest
3b753521 268 string that must appear at a fixed location, and we look for the
1de06328
YO
269 longest string that may appear at a floating location. So for instance
270 in the pattern:
271
272 /FOO[xX]A.*B[xX]BAR/
273
274 Both 'FOO' and 'A' are fixed strings. Both 'B' and 'BAR' are floating
275 strings (because they follow a .* construct). study_chunk will identify
276 both FOO and BAR as being the longest fixed and floating strings respectively.
277
278 The strings can be composites, for instance
279
280 /(f)(o)(o)/
281
282 will result in a composite fixed substring 'foo'.
283
284 For each string some basic information is maintained:
285
286 - offset or min_offset
287 This is the position the string must appear at, or not before.
288 It also implicitly (when combined with minlenp) tells us how many
3b753521
FN
289 characters must match before the string we are searching for.
290 Likewise when combined with minlenp and the length of the string it
1de06328
YO
291 tells us how many characters must appear after the string we have
292 found.
293
294 - max_offset
295 Only used for floating strings. This is the rightmost point that
3b753521 296 the string can appear at. If set to I32 max it indicates that the
1de06328
YO
297 string can occur infinitely far to the right.
298
299 - minlenp
2d608413
KW
300 A pointer to the minimum number of characters of the pattern that the
301 string was found inside. This is important as in the case of positive
1de06328
YO
302 lookahead or positive lookbehind we can have multiple patterns
303 involved. Consider
304
305 /(?=FOO).*F/
306
307 The minimum length of the pattern overall is 3, the minimum length
308 of the lookahead part is 3, but the minimum length of the part that
309 will actually match is 1. So 'FOO's minimum length is 3, but the
310 minimum length for the F is 1. This is important as the minimum length
311 is used to determine offsets in front of and behind the string being
312 looked for. Since strings can be composites this is the length of the
486ec47a 313 pattern at the time it was committed with a scan_commit. Note that
1de06328
YO
314 the length is calculated by study_chunk, so that the minimum lengths
315 are not known until the full pattern has been compiled, thus the
316 pointer to the value.
317
318 - lookbehind
319
320 In the case of lookbehind the string being searched for can be
321 offset past the start point of the final matching string.
322 If this value was just blithely removed from the min_offset it would
323 invalidate some of the calculations for how many chars must match
324 before or after (as they are derived from min_offset and minlen and
325 the length of the string being searched for).
326 When the final pattern is compiled and the data is moved from the
327 scan_data_t structure into the regexp structure the information
328 about lookbehind is factored in, with the information that would
329 have been lost precalculated in the end_shift field for the
330 associated string.
331
332 The fields pos_min and pos_delta are used to store the minimum offset
333 and the delta to the maximum offset at the current point in the pattern.
334
335*/
2c2d71f5
JH
336
337typedef struct scan_data_t {
1de06328
YO
338 /*I32 len_min; unused */
339 /*I32 len_delta; unused */
2c2d71f5
JH
340 I32 pos_min;
341 I32 pos_delta;
342 SV *last_found;
1de06328 343 I32 last_end; /* min value, <0 unless valid. */
2c2d71f5
JH
344 I32 last_start_min;
345 I32 last_start_max;
1de06328
YO
346 SV **longest; /* Either &l_fixed, or &l_float. */
347 SV *longest_fixed; /* longest fixed string found in pattern */
348 I32 offset_fixed; /* offset where it starts */
486ec47a 349 I32 *minlen_fixed; /* pointer to the minlen relevant to the string */
1de06328
YO
350 I32 lookbehind_fixed; /* is the position of the string modfied by LB */
351 SV *longest_float; /* longest floating string found in pattern */
352 I32 offset_float_min; /* earliest point in string it can appear */
353 I32 offset_float_max; /* latest point in string it can appear */
486ec47a 354 I32 *minlen_float; /* pointer to the minlen relevant to the string */
1de06328 355 I32 lookbehind_float; /* is the position of the string modified by LB */
2c2d71f5
JH
356 I32 flags;
357 I32 whilem_c;
cb434fcc 358 I32 *last_closep;
653099ff 359 struct regnode_charclass_class *start_class;
2c2d71f5
JH
360} scan_data_t;
361
a687059c 362/*
e50aee73 363 * Forward declarations for pregcomp()'s friends.
a687059c 364 */
a0d0e21e 365
27da23d5 366static const scan_data_t zero_scan_data =
1de06328 367 { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ,0};
c277df42
IZ
368
369#define SF_BEFORE_EOL (SF_BEFORE_SEOL|SF_BEFORE_MEOL)
07be1b83
YO
370#define SF_BEFORE_SEOL 0x0001
371#define SF_BEFORE_MEOL 0x0002
c277df42
IZ
372#define SF_FIX_BEFORE_EOL (SF_FIX_BEFORE_SEOL|SF_FIX_BEFORE_MEOL)
373#define SF_FL_BEFORE_EOL (SF_FL_BEFORE_SEOL|SF_FL_BEFORE_MEOL)
374
09b7f37c
CB
375#ifdef NO_UNARY_PLUS
376# define SF_FIX_SHIFT_EOL (0+2)
377# define SF_FL_SHIFT_EOL (0+4)
378#else
379# define SF_FIX_SHIFT_EOL (+2)
380# define SF_FL_SHIFT_EOL (+4)
381#endif
c277df42
IZ
382
383#define SF_FIX_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FIX_SHIFT_EOL)
384#define SF_FIX_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FIX_SHIFT_EOL)
385
386#define SF_FL_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FL_SHIFT_EOL)
387#define SF_FL_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FL_SHIFT_EOL) /* 0x20 */
07be1b83
YO
388#define SF_IS_INF 0x0040
389#define SF_HAS_PAR 0x0080
390#define SF_IN_PAR 0x0100
391#define SF_HAS_EVAL 0x0200
392#define SCF_DO_SUBSTR 0x0400
653099ff
GS
393#define SCF_DO_STCLASS_AND 0x0800
394#define SCF_DO_STCLASS_OR 0x1000
395#define SCF_DO_STCLASS (SCF_DO_STCLASS_AND|SCF_DO_STCLASS_OR)
e1901655 396#define SCF_WHILEM_VISITED_POS 0x2000
c277df42 397
786e8c11 398#define SCF_TRIE_RESTUDY 0x4000 /* Do restudy? */
e2e6a0f1 399#define SCF_SEEN_ACCEPT 0x8000
07be1b83 400
43fead97 401#define UTF cBOOL(RExC_utf8)
00b27cfc
KW
402
403/* The enums for all these are ordered so things work out correctly */
a62b1201 404#define LOC (get_regex_charset(RExC_flags) == REGEX_LOCALE_CHARSET)
cfaf538b 405#define DEPENDS_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_DEPENDS_CHARSET)
00b27cfc 406#define UNI_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_UNICODE_CHARSET)
cfaf538b
KW
407#define AT_LEAST_UNI_SEMANTICS (get_regex_charset(RExC_flags) >= REGEX_UNICODE_CHARSET)
408#define ASCII_RESTRICTED (get_regex_charset(RExC_flags) == REGEX_ASCII_RESTRICTED_CHARSET)
2f7f8cb1 409#define AT_LEAST_ASCII_RESTRICTED (get_regex_charset(RExC_flags) >= REGEX_ASCII_RESTRICTED_CHARSET)
a725e29c 410#define ASCII_FOLD_RESTRICTED (get_regex_charset(RExC_flags) == REGEX_ASCII_MORE_RESTRICTED_CHARSET)
a62b1201 411
43fead97 412#define FOLD cBOOL(RExC_flags & RXf_PMf_FOLD)
a0ed51b3 413
93733859 414#define OOB_NAMEDCLASS -1
b8c5462f 415
8e661ac5
KW
416/* There is no code point that is out-of-bounds, so this is problematic. But
417 * its only current use is to initialize a variable that is always set before
418 * looked at. */
419#define OOB_UNICODE 0xDEADBEEF
420
a0ed51b3
LW
421#define CHR_SVLEN(sv) (UTF ? sv_len_utf8(sv) : SvCUR(sv))
422#define CHR_DIST(a,b) (UTF ? utf8_distance(a,b) : a - b)
423
8615cb43 424
b45f050a
JF
425/* length of regex to show in messages that don't mark a position within */
426#define RegexLengthToShowInErrorMessages 127
427
428/*
429 * If MARKER[12] are adjusted, be sure to adjust the constants at the top
430 * of t/op/regmesg.t, the tests in t/op/re_tests, and those in
431 * op/pragma/warn/regcomp.
432 */
7253e4e3
RK
433#define MARKER1 "<-- HERE" /* marker as it appears in the description */
434#define MARKER2 " <-- HERE " /* marker as it appears within the regex */
b81d288d 435
7253e4e3 436#define REPORT_LOCATION " in regex; marked by " MARKER1 " in m/%.*s" MARKER2 "%s/"
b45f050a
JF
437
438/*
439 * Calls SAVEDESTRUCTOR_X if needed, then calls Perl_croak with the given
440 * arg. Show regex, up to a maximum length. If it's too long, chop and add
441 * "...".
442 */
58e23c8d 443#define _FAIL(code) STMT_START { \
bfed75c6 444 const char *ellipses = ""; \
ccb2c380
MP
445 IV len = RExC_end - RExC_precomp; \
446 \
447 if (!SIZE_ONLY) \
288b8c02 448 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv); \
ccb2c380
MP
449 if (len > RegexLengthToShowInErrorMessages) { \
450 /* chop 10 shorter than the max, to ensure meaning of "..." */ \
451 len = RegexLengthToShowInErrorMessages - 10; \
452 ellipses = "..."; \
453 } \
58e23c8d 454 code; \
ccb2c380 455} STMT_END
8615cb43 456
58e23c8d
YO
457#define FAIL(msg) _FAIL( \
458 Perl_croak(aTHX_ "%s in regex m/%.*s%s/", \
459 msg, (int)len, RExC_precomp, ellipses))
460
461#define FAIL2(msg,arg) _FAIL( \
462 Perl_croak(aTHX_ msg " in regex m/%.*s%s/", \
463 arg, (int)len, RExC_precomp, ellipses))
464
b45f050a 465/*
b45f050a
JF
466 * Simple_vFAIL -- like FAIL, but marks the current location in the scan
467 */
ccb2c380 468#define Simple_vFAIL(m) STMT_START { \
a28509cc 469 const IV offset = RExC_parse - RExC_precomp; \
ccb2c380
MP
470 Perl_croak(aTHX_ "%s" REPORT_LOCATION, \
471 m, (int)offset, RExC_precomp, RExC_precomp + offset); \
472} STMT_END
b45f050a
JF
473
474/*
475 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL()
476 */
ccb2c380
MP
477#define vFAIL(m) STMT_START { \
478 if (!SIZE_ONLY) \
288b8c02 479 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv); \
ccb2c380
MP
480 Simple_vFAIL(m); \
481} STMT_END
b45f050a
JF
482
483/*
484 * Like Simple_vFAIL(), but accepts two arguments.
485 */
ccb2c380 486#define Simple_vFAIL2(m,a1) STMT_START { \
a28509cc 487 const IV offset = RExC_parse - RExC_precomp; \
ccb2c380
MP
488 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, \
489 (int)offset, RExC_precomp, RExC_precomp + offset); \
490} STMT_END
b45f050a
JF
491
492/*
493 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL2().
494 */
ccb2c380
MP
495#define vFAIL2(m,a1) STMT_START { \
496 if (!SIZE_ONLY) \
288b8c02 497 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv); \
ccb2c380
MP
498 Simple_vFAIL2(m, a1); \
499} STMT_END
b45f050a
JF
500
501
502/*
503 * Like Simple_vFAIL(), but accepts three arguments.
504 */
ccb2c380 505#define Simple_vFAIL3(m, a1, a2) STMT_START { \
a28509cc 506 const IV offset = RExC_parse - RExC_precomp; \
ccb2c380
MP
507 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, a2, \
508 (int)offset, RExC_precomp, RExC_precomp + offset); \
509} STMT_END
b45f050a
JF
510
511/*
512 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL3().
513 */
ccb2c380
MP
514#define vFAIL3(m,a1,a2) STMT_START { \
515 if (!SIZE_ONLY) \
288b8c02 516 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv); \
ccb2c380
MP
517 Simple_vFAIL3(m, a1, a2); \
518} STMT_END
b45f050a
JF
519
520/*
521 * Like Simple_vFAIL(), but accepts four arguments.
522 */
ccb2c380 523#define Simple_vFAIL4(m, a1, a2, a3) STMT_START { \
a28509cc 524 const IV offset = RExC_parse - RExC_precomp; \
ccb2c380
MP
525 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, a2, a3, \
526 (int)offset, RExC_precomp, RExC_precomp + offset); \
527} STMT_END
b45f050a 528
668c081a 529#define ckWARNreg(loc,m) STMT_START { \
a28509cc 530 const IV offset = loc - RExC_precomp; \
f10f4c18
NC
531 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
532 (int)offset, RExC_precomp, RExC_precomp + offset); \
ccb2c380
MP
533} STMT_END
534
668c081a 535#define ckWARNregdep(loc,m) STMT_START { \
a28509cc 536 const IV offset = loc - RExC_precomp; \
d1d15184 537 Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
f10f4c18
NC
538 m REPORT_LOCATION, \
539 (int)offset, RExC_precomp, RExC_precomp + offset); \
ccb2c380
MP
540} STMT_END
541
2335b3d3
KW
542#define ckWARN2regdep(loc,m, a1) STMT_START { \
543 const IV offset = loc - RExC_precomp; \
544 Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
545 m REPORT_LOCATION, \
546 a1, (int)offset, RExC_precomp, RExC_precomp + offset); \
547} STMT_END
548
668c081a 549#define ckWARN2reg(loc, m, a1) STMT_START { \
a28509cc 550 const IV offset = loc - RExC_precomp; \
668c081a 551 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
ccb2c380
MP
552 a1, (int)offset, RExC_precomp, RExC_precomp + offset); \
553} STMT_END
554
555#define vWARN3(loc, m, a1, a2) STMT_START { \
a28509cc 556 const IV offset = loc - RExC_precomp; \
ccb2c380
MP
557 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
558 a1, a2, (int)offset, RExC_precomp, RExC_precomp + offset); \
559} STMT_END
560
668c081a
NC
561#define ckWARN3reg(loc, m, a1, a2) STMT_START { \
562 const IV offset = loc - RExC_precomp; \
563 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
564 a1, a2, (int)offset, RExC_precomp, RExC_precomp + offset); \
565} STMT_END
566
ccb2c380 567#define vWARN4(loc, m, a1, a2, a3) STMT_START { \
a28509cc 568 const IV offset = loc - RExC_precomp; \
ccb2c380
MP
569 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
570 a1, a2, a3, (int)offset, RExC_precomp, RExC_precomp + offset); \
571} STMT_END
572
668c081a
NC
573#define ckWARN4reg(loc, m, a1, a2, a3) STMT_START { \
574 const IV offset = loc - RExC_precomp; \
575 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
576 a1, a2, a3, (int)offset, RExC_precomp, RExC_precomp + offset); \
577} STMT_END
578
ccb2c380 579#define vWARN5(loc, m, a1, a2, a3, a4) STMT_START { \
a28509cc 580 const IV offset = loc - RExC_precomp; \
ccb2c380
MP
581 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
582 a1, a2, a3, a4, (int)offset, RExC_precomp, RExC_precomp + offset); \
583} STMT_END
9d1d55b5 584
8615cb43 585
cd439c50 586/* Allow for side effects in s */
ccb2c380
MP
587#define REGC(c,s) STMT_START { \
588 if (!SIZE_ONLY) *(s) = (c); else (void)(s); \
589} STMT_END
cd439c50 590
fac92740
MJD
591/* Macros for recording node offsets. 20001227 mjd@plover.com
592 * Nodes are numbered 1, 2, 3, 4. Node #n's position is recorded in
593 * element 2*n-1 of the array. Element #2n holds the byte length node #n.
594 * Element 0 holds the number n.
07be1b83 595 * Position is 1 indexed.
fac92740 596 */
7122b237
YO
597#ifndef RE_TRACK_PATTERN_OFFSETS
598#define Set_Node_Offset_To_R(node,byte)
599#define Set_Node_Offset(node,byte)
600#define Set_Cur_Node_Offset
601#define Set_Node_Length_To_R(node,len)
602#define Set_Node_Length(node,len)
603#define Set_Node_Cur_Length(node)
604#define Node_Offset(n)
605#define Node_Length(n)
606#define Set_Node_Offset_Length(node,offset,len)
607#define ProgLen(ri) ri->u.proglen
608#define SetProgLen(ri,x) ri->u.proglen = x
609#else
610#define ProgLen(ri) ri->u.offsets[0]
611#define SetProgLen(ri,x) ri->u.offsets[0] = x
ccb2c380
MP
612#define Set_Node_Offset_To_R(node,byte) STMT_START { \
613 if (! SIZE_ONLY) { \
614 MJD_OFFSET_DEBUG(("** (%d) offset of node %d is %d.\n", \
2a49f0f5 615 __LINE__, (int)(node), (int)(byte))); \
ccb2c380 616 if((node) < 0) { \
551405c4 617 Perl_croak(aTHX_ "value of node is %d in Offset macro", (int)(node)); \
ccb2c380
MP
618 } else { \
619 RExC_offsets[2*(node)-1] = (byte); \
620 } \
621 } \
622} STMT_END
623
624#define Set_Node_Offset(node,byte) \
625 Set_Node_Offset_To_R((node)-RExC_emit_start, (byte)-RExC_start)
626#define Set_Cur_Node_Offset Set_Node_Offset(RExC_emit, RExC_parse)
627
628#define Set_Node_Length_To_R(node,len) STMT_START { \
629 if (! SIZE_ONLY) { \
630 MJD_OFFSET_DEBUG(("** (%d) size of node %d is %d.\n", \
551405c4 631 __LINE__, (int)(node), (int)(len))); \
ccb2c380 632 if((node) < 0) { \
551405c4 633 Perl_croak(aTHX_ "value of node is %d in Length macro", (int)(node)); \
ccb2c380
MP
634 } else { \
635 RExC_offsets[2*(node)] = (len); \
636 } \
637 } \
638} STMT_END
639
640#define Set_Node_Length(node,len) \
641 Set_Node_Length_To_R((node)-RExC_emit_start, len)
642#define Set_Cur_Node_Length(len) Set_Node_Length(RExC_emit, len)
643#define Set_Node_Cur_Length(node) \
644 Set_Node_Length(node, RExC_parse - parse_start)
fac92740
MJD
645
646/* Get offsets and lengths */
647#define Node_Offset(n) (RExC_offsets[2*((n)-RExC_emit_start)-1])
648#define Node_Length(n) (RExC_offsets[2*((n)-RExC_emit_start)])
649
07be1b83
YO
650#define Set_Node_Offset_Length(node,offset,len) STMT_START { \
651 Set_Node_Offset_To_R((node)-RExC_emit_start, (offset)); \
652 Set_Node_Length_To_R((node)-RExC_emit_start, (len)); \
653} STMT_END
7122b237 654#endif
07be1b83
YO
655
656#if PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS
657#define EXPERIMENTAL_INPLACESCAN
f427392e 658#endif /*PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS*/
07be1b83 659
304ee84b
YO
660#define DEBUG_STUDYDATA(str,data,depth) \
661DEBUG_OPTIMISE_MORE_r(if(data){ \
1de06328 662 PerlIO_printf(Perl_debug_log, \
304ee84b
YO
663 "%*s" str "Pos:%"IVdf"/%"IVdf \
664 " Flags: 0x%"UVXf" Whilem_c: %"IVdf" Lcp: %"IVdf" %s", \
1de06328
YO
665 (int)(depth)*2, "", \
666 (IV)((data)->pos_min), \
667 (IV)((data)->pos_delta), \
304ee84b 668 (UV)((data)->flags), \
1de06328 669 (IV)((data)->whilem_c), \
304ee84b
YO
670 (IV)((data)->last_closep ? *((data)->last_closep) : -1), \
671 is_inf ? "INF " : "" \
1de06328
YO
672 ); \
673 if ((data)->last_found) \
674 PerlIO_printf(Perl_debug_log, \
675 "Last:'%s' %"IVdf":%"IVdf"/%"IVdf" %sFixed:'%s' @ %"IVdf \
676 " %sFloat: '%s' @ %"IVdf"/%"IVdf"", \
677 SvPVX_const((data)->last_found), \
678 (IV)((data)->last_end), \
679 (IV)((data)->last_start_min), \
680 (IV)((data)->last_start_max), \
681 ((data)->longest && \
682 (data)->longest==&((data)->longest_fixed)) ? "*" : "", \
683 SvPVX_const((data)->longest_fixed), \
684 (IV)((data)->offset_fixed), \
685 ((data)->longest && \
686 (data)->longest==&((data)->longest_float)) ? "*" : "", \
687 SvPVX_const((data)->longest_float), \
688 (IV)((data)->offset_float_min), \
689 (IV)((data)->offset_float_max) \
690 ); \
691 PerlIO_printf(Perl_debug_log,"\n"); \
692});
693
acfe0abc 694static void clear_re(pTHX_ void *r);
4327152a 695
653099ff 696/* Mark that we cannot extend a found fixed substring at this point.
786e8c11 697 Update the longest found anchored substring and the longest found
653099ff
GS
698 floating substrings if needed. */
699
4327152a 700STATIC void
304ee84b 701S_scan_commit(pTHX_ const RExC_state_t *pRExC_state, scan_data_t *data, I32 *minlenp, int is_inf)
c277df42 702{
e1ec3a88
AL
703 const STRLEN l = CHR_SVLEN(data->last_found);
704 const STRLEN old_l = CHR_SVLEN(*data->longest);
1de06328 705 GET_RE_DEBUG_FLAGS_DECL;
b81d288d 706
7918f24d
NC
707 PERL_ARGS_ASSERT_SCAN_COMMIT;
708
c277df42 709 if ((l >= old_l) && ((l > old_l) || (data->flags & SF_BEFORE_EOL))) {
6b43b216 710 SvSetMagicSV(*data->longest, data->last_found);
c277df42
IZ
711 if (*data->longest == data->longest_fixed) {
712 data->offset_fixed = l ? data->last_start_min : data->pos_min;
713 if (data->flags & SF_BEFORE_EOL)
b81d288d 714 data->flags
c277df42
IZ
715 |= ((data->flags & SF_BEFORE_EOL) << SF_FIX_SHIFT_EOL);
716 else
717 data->flags &= ~SF_FIX_BEFORE_EOL;
686b73d4 718 data->minlen_fixed=minlenp;
1de06328 719 data->lookbehind_fixed=0;
a0ed51b3 720 }
304ee84b 721 else { /* *data->longest == data->longest_float */
c277df42 722 data->offset_float_min = l ? data->last_start_min : data->pos_min;
b81d288d
AB
723 data->offset_float_max = (l
724 ? data->last_start_max
c277df42 725 : data->pos_min + data->pos_delta);
304ee84b 726 if (is_inf || (U32)data->offset_float_max > (U32)I32_MAX)
9051bda5 727 data->offset_float_max = I32_MAX;
c277df42 728 if (data->flags & SF_BEFORE_EOL)
b81d288d 729 data->flags
c277df42
IZ
730 |= ((data->flags & SF_BEFORE_EOL) << SF_FL_SHIFT_EOL);
731 else
732 data->flags &= ~SF_FL_BEFORE_EOL;
1de06328
YO
733 data->minlen_float=minlenp;
734 data->lookbehind_float=0;
c277df42
IZ
735 }
736 }
737 SvCUR_set(data->last_found, 0);
0eda9292 738 {
a28509cc 739 SV * const sv = data->last_found;
097eb12c
AL
740 if (SvUTF8(sv) && SvMAGICAL(sv)) {
741 MAGIC * const mg = mg_find(sv, PERL_MAGIC_utf8);
742 if (mg)
743 mg->mg_len = 0;
744 }
0eda9292 745 }
c277df42
IZ
746 data->last_end = -1;
747 data->flags &= ~SF_BEFORE_EOL;
bcdf7404 748 DEBUG_STUDYDATA("commit: ",data,0);
c277df42
IZ
749}
750
653099ff
GS
751/* Can match anything (initialization) */
752STATIC void
3fffb88a 753S_cl_anything(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
653099ff 754{
7918f24d
NC
755 PERL_ARGS_ASSERT_CL_ANYTHING;
756
f8bef550 757 ANYOF_BITMAP_SETALL(cl);
dd58aee1 758 cl->flags = ANYOF_CLASS|ANYOF_EOS|ANYOF_UNICODE_ALL
3ad98780 759 |ANYOF_LOC_NONBITMAP_FOLD|ANYOF_NON_UTF8_LATIN1_ALL;
3fffb88a
KW
760
761 /* If any portion of the regex is to operate under locale rules,
762 * initialization includes it. The reason this isn't done for all regexes
763 * is that the optimizer was written under the assumption that locale was
764 * all-or-nothing. Given the complexity and lack of documentation in the
765 * optimizer, and that there are inadequate test cases for locale, so many
766 * parts of it may not work properly, it is safest to avoid locale unless
767 * necessary. */
768 if (RExC_contains_locale) {
9d7a1e63 769 ANYOF_CLASS_SETALL(cl); /* /l uses class */
3fffb88a
KW
770 cl->flags |= ANYOF_LOCALE;
771 }
9d7a1e63
KW
772 else {
773 ANYOF_CLASS_ZERO(cl); /* Only /l uses class now */
774 }
653099ff
GS
775}
776
777/* Can match anything (initialization) */
778STATIC int
5f66b61c 779S_cl_is_anything(const struct regnode_charclass_class *cl)
653099ff
GS
780{
781 int value;
782
7918f24d
NC
783 PERL_ARGS_ASSERT_CL_IS_ANYTHING;
784
aaa51d5e 785 for (value = 0; value <= ANYOF_MAX; value += 2)
653099ff
GS
786 if (ANYOF_CLASS_TEST(cl, value) && ANYOF_CLASS_TEST(cl, value + 1))
787 return 1;
1aa99e6b
IH
788 if (!(cl->flags & ANYOF_UNICODE_ALL))
789 return 0;
10edeb5d 790 if (!ANYOF_BITMAP_TESTALLSET((const void*)cl))
f8bef550 791 return 0;
653099ff
GS
792 return 1;
793}
794
795/* Can match anything (initialization) */
796STATIC void
e755fd73 797S_cl_init(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
653099ff 798{
7918f24d
NC
799 PERL_ARGS_ASSERT_CL_INIT;
800
8ecf7187 801 Zero(cl, 1, struct regnode_charclass_class);
653099ff 802 cl->type = ANYOF;
3fffb88a 803 cl_anything(pRExC_state, cl);
1411dba4 804 ARG_SET(cl, ANYOF_NONBITMAP_EMPTY);
653099ff
GS
805}
806
1051e1c4
KW
807/* These two functions currently do the exact same thing */
808#define cl_init_zero S_cl_init
653099ff 809
dd58aee1
KW
810/* 'AND' a given class with another one. Can create false positives. 'cl'
811 * should not be inverted. 'and_with->flags & ANYOF_CLASS' should be 0 if
812 * 'and_with' is a regnode_charclass instead of a regnode_charclass_class. */
653099ff 813STATIC void
5f66b61c 814S_cl_and(struct regnode_charclass_class *cl,
a28509cc 815 const struct regnode_charclass_class *and_with)
653099ff 816{
7918f24d 817 PERL_ARGS_ASSERT_CL_AND;
40d049e4
YO
818
819 assert(and_with->type == ANYOF);
1e6ade67 820
c6b76537 821 /* I (khw) am not sure all these restrictions are necessary XXX */
1e6ade67
KW
822 if (!(ANYOF_CLASS_TEST_ANY_SET(and_with))
823 && !(ANYOF_CLASS_TEST_ANY_SET(cl))
653099ff 824 && (and_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
39065660
KW
825 && !(and_with->flags & ANYOF_LOC_NONBITMAP_FOLD)
826 && !(cl->flags & ANYOF_LOC_NONBITMAP_FOLD)) {
653099ff
GS
827 int i;
828
829 if (and_with->flags & ANYOF_INVERT)
830 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
831 cl->bitmap[i] &= ~and_with->bitmap[i];
832 else
833 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
834 cl->bitmap[i] &= and_with->bitmap[i];
835 } /* XXXX: logic is complicated otherwise, leave it along for a moment. */
1aa99e6b 836
c6b76537 837 if (and_with->flags & ANYOF_INVERT) {
8951c461 838
c6b76537
KW
839 /* Here, the and'ed node is inverted. Get the AND of the flags that
840 * aren't affected by the inversion. Those that are affected are
841 * handled individually below */
842 U8 affected_flags = cl->flags & ~INVERSION_UNAFFECTED_FLAGS;
843 cl->flags &= (and_with->flags & INVERSION_UNAFFECTED_FLAGS);
844 cl->flags |= affected_flags;
845
846 /* We currently don't know how to deal with things that aren't in the
847 * bitmap, but we know that the intersection is no greater than what
848 * is already in cl, so let there be false positives that get sorted
849 * out after the synthetic start class succeeds, and the node is
850 * matched for real. */
851
852 /* The inversion of these two flags indicate that the resulting
853 * intersection doesn't have them */
854 if (and_with->flags & ANYOF_UNICODE_ALL) {
4713bfe1
KW
855 cl->flags &= ~ANYOF_UNICODE_ALL;
856 }
c6b76537
KW
857 if (and_with->flags & ANYOF_NON_UTF8_LATIN1_ALL) {
858 cl->flags &= ~ANYOF_NON_UTF8_LATIN1_ALL;
137165a6 859 }
1aa99e6b 860 }
c6b76537 861 else { /* and'd node is not inverted */
3ad98780
KW
862 U8 outside_bitmap_but_not_utf8; /* Temp variable */
863
137165a6 864 if (! ANYOF_NONBITMAP(and_with)) {
c6b76537
KW
865
866 /* Here 'and_with' doesn't match anything outside the bitmap
867 * (except possibly ANYOF_UNICODE_ALL), which means the
868 * intersection can't either, except for ANYOF_UNICODE_ALL, in
869 * which case we don't know what the intersection is, but it's no
870 * greater than what cl already has, so can just leave it alone,
871 * with possible false positives */
872 if (! (and_with->flags & ANYOF_UNICODE_ALL)) {
873 ARG_SET(cl, ANYOF_NONBITMAP_EMPTY);
871d0d1a 874 cl->flags &= ~ANYOF_NONBITMAP_NON_UTF8;
c6b76537 875 }
137165a6 876 }
c6b76537
KW
877 else if (! ANYOF_NONBITMAP(cl)) {
878
879 /* Here, 'and_with' does match something outside the bitmap, and cl
880 * doesn't have a list of things to match outside the bitmap. If
881 * cl can match all code points above 255, the intersection will
3ad98780
KW
882 * be those above-255 code points that 'and_with' matches. If cl
883 * can't match all Unicode code points, it means that it can't
884 * match anything outside the bitmap (since the 'if' that got us
885 * into this block tested for that), so we leave the bitmap empty.
886 */
c6b76537
KW
887 if (cl->flags & ANYOF_UNICODE_ALL) {
888 ARG_SET(cl, ARG(and_with));
3ad98780
KW
889
890 /* and_with's ARG may match things that don't require UTF8.
891 * And now cl's will too, in spite of this being an 'and'. See
892 * the comments below about the kludge */
893 cl->flags |= and_with->flags & ANYOF_NONBITMAP_NON_UTF8;
c6b76537
KW
894 }
895 }
896 else {
897 /* Here, both 'and_with' and cl match something outside the
898 * bitmap. Currently we do not do the intersection, so just match
899 * whatever cl had at the beginning. */
900 }
901
902
3ad98780
KW
903 /* Take the intersection of the two sets of flags. However, the
904 * ANYOF_NONBITMAP_NON_UTF8 flag is treated as an 'or'. This is a
905 * kludge around the fact that this flag is not treated like the others
906 * which are initialized in cl_anything(). The way the optimizer works
907 * is that the synthetic start class (SSC) is initialized to match
908 * anything, and then the first time a real node is encountered, its
909 * values are AND'd with the SSC's with the result being the values of
910 * the real node. However, there are paths through the optimizer where
911 * the AND never gets called, so those initialized bits are set
912 * inappropriately, which is not usually a big deal, as they just cause
913 * false positives in the SSC, which will just mean a probably
914 * imperceptible slow down in execution. However this bit has a
915 * higher false positive consequence in that it can cause utf8.pm,
916 * utf8_heavy.pl ... to be loaded when not necessary, which is a much
917 * bigger slowdown and also causes significant extra memory to be used.
918 * In order to prevent this, the code now takes a different tack. The
919 * bit isn't set unless some part of the regular expression needs it,
920 * but once set it won't get cleared. This means that these extra
921 * modules won't get loaded unless there was some path through the
922 * pattern that would have required them anyway, and so any false
923 * positives that occur by not ANDing them out when they could be
924 * aren't as severe as they would be if we treated this bit like all
925 * the others */
926 outside_bitmap_but_not_utf8 = (cl->flags | and_with->flags)
927 & ANYOF_NONBITMAP_NON_UTF8;
c6b76537 928 cl->flags &= and_with->flags;
3ad98780 929 cl->flags |= outside_bitmap_but_not_utf8;
137165a6 930 }
653099ff
GS
931}
932
dd58aee1
KW
933/* 'OR' a given class with another one. Can create false positives. 'cl'
934 * should not be inverted. 'or_with->flags & ANYOF_CLASS' should be 0 if
935 * 'or_with' is a regnode_charclass instead of a regnode_charclass_class. */
653099ff 936STATIC void
3fffb88a 937S_cl_or(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl, const struct regnode_charclass_class *or_with)
653099ff 938{
7918f24d
NC
939 PERL_ARGS_ASSERT_CL_OR;
940
653099ff 941 if (or_with->flags & ANYOF_INVERT) {
c6b76537
KW
942
943 /* Here, the or'd node is to be inverted. This means we take the
944 * complement of everything not in the bitmap, but currently we don't
945 * know what that is, so give up and match anything */
946 if (ANYOF_NONBITMAP(or_with)) {
3fffb88a 947 cl_anything(pRExC_state, cl);
c6b76537 948 }
653099ff
GS
949 /* We do not use
950 * (B1 | CL1) | (!B2 & !CL2) = (B1 | !B2 & !CL2) | (CL1 | (!B2 & !CL2))
951 * <= (B1 | !B2) | (CL1 | !CL2)
952 * which is wasteful if CL2 is small, but we ignore CL2:
953 * (B1 | CL1) | (!B2 & !CL2) <= (B1 | CL1) | !B2 = (B1 | !B2) | CL1
954 * XXXX Can we handle case-fold? Unclear:
955 * (OK1(i) | OK1(i')) | !(OK1(i) | OK1(i')) =
956 * (OK1(i) | OK1(i')) | (!OK1(i) & !OK1(i'))
957 */
c6b76537 958 else if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
39065660
KW
959 && !(or_with->flags & ANYOF_LOC_NONBITMAP_FOLD)
960 && !(cl->flags & ANYOF_LOC_NONBITMAP_FOLD) ) {
653099ff
GS
961 int i;
962
963 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
964 cl->bitmap[i] |= ~or_with->bitmap[i];
965 } /* XXXX: logic is complicated otherwise */
966 else {
3fffb88a 967 cl_anything(pRExC_state, cl);
653099ff 968 }
c6b76537
KW
969
970 /* And, we can just take the union of the flags that aren't affected
971 * by the inversion */
972 cl->flags |= or_with->flags & INVERSION_UNAFFECTED_FLAGS;
973
974 /* For the remaining flags:
975 ANYOF_UNICODE_ALL and inverted means to not match anything above
976 255, which means that the union with cl should just be
977 what cl has in it, so can ignore this flag
978 ANYOF_NON_UTF8_LATIN1_ALL and inverted means if not utf8 and ord
979 is 127-255 to match them, but then invert that, so the
980 union with cl should just be what cl has in it, so can
981 ignore this flag
982 */
983 } else { /* 'or_with' is not inverted */
653099ff
GS
984 /* (B1 | CL1) | (B2 | CL2) = (B1 | B2) | (CL1 | CL2)) */
985 if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
39065660
KW
986 && (!(or_with->flags & ANYOF_LOC_NONBITMAP_FOLD)
987 || (cl->flags & ANYOF_LOC_NONBITMAP_FOLD)) ) {
653099ff
GS
988 int i;
989
990 /* OR char bitmap and class bitmap separately */
991 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
992 cl->bitmap[i] |= or_with->bitmap[i];
1e6ade67 993 if (ANYOF_CLASS_TEST_ANY_SET(or_with)) {
653099ff
GS
994 for (i = 0; i < ANYOF_CLASSBITMAP_SIZE; i++)
995 cl->classflags[i] |= or_with->classflags[i];
996 cl->flags |= ANYOF_CLASS;
997 }
998 }
999 else { /* XXXX: logic is complicated, leave it along for a moment. */
3fffb88a 1000 cl_anything(pRExC_state, cl);
653099ff 1001 }
9826f543 1002
c6b76537
KW
1003 if (ANYOF_NONBITMAP(or_with)) {
1004
1005 /* Use the added node's outside-the-bit-map match if there isn't a
1006 * conflict. If there is a conflict (both nodes match something
1007 * outside the bitmap, but what they match outside is not the same
1008 * pointer, and hence not easily compared until XXX we extend
1009 * inversion lists this far), give up and allow the start class to
d94b1d13
KW
1010 * match everything outside the bitmap. If that stuff is all above
1011 * 255, can just set UNICODE_ALL, otherwise caould be anything. */
c6b76537
KW
1012 if (! ANYOF_NONBITMAP(cl)) {
1013 ARG_SET(cl, ARG(or_with));
1014 }
1015 else if (ARG(cl) != ARG(or_with)) {
d94b1d13
KW
1016
1017 if ((or_with->flags & ANYOF_NONBITMAP_NON_UTF8)) {
1018 cl_anything(pRExC_state, cl);
1019 }
1020 else {
1021 cl->flags |= ANYOF_UNICODE_ALL;
1022 }
c6b76537 1023 }
4c34a693 1024 }
0b9668ee
KW
1025
1026 /* Take the union */
1027 cl->flags |= or_with->flags;
1aa99e6b 1028 }
653099ff
GS
1029}
1030
a3621e74
YO
1031#define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
1032#define TRIE_LIST_CUR(state) ( TRIE_LIST_ITEM( state, 0 ).forid )
1033#define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
1034#define TRIE_LIST_USED(idx) ( trie->states[state].trans.list ? (TRIE_LIST_CUR( idx ) - 1) : 0 )
1035
3dab1dad
YO
1036
1037#ifdef DEBUGGING
07be1b83 1038/*
2b8b4781
NC
1039 dump_trie(trie,widecharmap,revcharmap)
1040 dump_trie_interim_list(trie,widecharmap,revcharmap,next_alloc)
1041 dump_trie_interim_table(trie,widecharmap,revcharmap,next_alloc)
3dab1dad
YO
1042
1043 These routines dump out a trie in a somewhat readable format.
07be1b83
YO
1044 The _interim_ variants are used for debugging the interim
1045 tables that are used to generate the final compressed
1046 representation which is what dump_trie expects.
1047
486ec47a 1048 Part of the reason for their existence is to provide a form
3dab1dad 1049 of documentation as to how the different representations function.
07be1b83
YO
1050
1051*/
3dab1dad
YO
1052
1053/*
3dab1dad
YO
1054 Dumps the final compressed table form of the trie to Perl_debug_log.
1055 Used for debugging make_trie().
1056*/
b9a59e08 1057
3dab1dad 1058STATIC void
2b8b4781
NC
1059S_dump_trie(pTHX_ const struct _reg_trie_data *trie, HV *widecharmap,
1060 AV *revcharmap, U32 depth)
3dab1dad
YO
1061{
1062 U32 state;
ab3bbdeb 1063 SV *sv=sv_newmortal();
55eed653 1064 int colwidth= widecharmap ? 6 : 4;
2e64971a 1065 U16 word;
3dab1dad
YO
1066 GET_RE_DEBUG_FLAGS_DECL;
1067
7918f24d 1068 PERL_ARGS_ASSERT_DUMP_TRIE;
ab3bbdeb 1069
3dab1dad
YO
1070 PerlIO_printf( Perl_debug_log, "%*sChar : %-6s%-6s%-4s ",
1071 (int)depth * 2 + 2,"",
1072 "Match","Base","Ofs" );
1073
1074 for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
2b8b4781 1075 SV ** const tmp = av_fetch( revcharmap, state, 0);
3dab1dad 1076 if ( tmp ) {
ab3bbdeb
YO
1077 PerlIO_printf( Perl_debug_log, "%*s",
1078 colwidth,
ddc5bc0f 1079 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
ab3bbdeb
YO
1080 PL_colors[0], PL_colors[1],
1081 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1082 PERL_PV_ESCAPE_FIRSTCHAR
1083 )
1084 );
3dab1dad
YO
1085 }
1086 }
1087 PerlIO_printf( Perl_debug_log, "\n%*sState|-----------------------",
1088 (int)depth * 2 + 2,"");
1089
1090 for( state = 0 ; state < trie->uniquecharcount ; state++ )
ab3bbdeb 1091 PerlIO_printf( Perl_debug_log, "%.*s", colwidth, "--------");
3dab1dad
YO
1092 PerlIO_printf( Perl_debug_log, "\n");
1093
1e2e3d02 1094 for( state = 1 ; state < trie->statecount ; state++ ) {
be8e71aa 1095 const U32 base = trie->states[ state ].trans.base;
3dab1dad
YO
1096
1097 PerlIO_printf( Perl_debug_log, "%*s#%4"UVXf"|", (int)depth * 2 + 2,"", (UV)state);
1098
1099 if ( trie->states[ state ].wordnum ) {
1100 PerlIO_printf( Perl_debug_log, " W%4X", trie->states[ state ].wordnum );
1101 } else {
1102 PerlIO_printf( Perl_debug_log, "%6s", "" );
1103 }
1104
1105 PerlIO_printf( Perl_debug_log, " @%4"UVXf" ", (UV)base );
1106
1107 if ( base ) {
1108 U32 ofs = 0;
1109
1110 while( ( base + ofs < trie->uniquecharcount ) ||
1111 ( base + ofs - trie->uniquecharcount < trie->lasttrans
1112 && trie->trans[ base + ofs - trie->uniquecharcount ].check != state))
1113 ofs++;
1114
1115 PerlIO_printf( Perl_debug_log, "+%2"UVXf"[ ", (UV)ofs);
1116
1117 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
1118 if ( ( base + ofs >= trie->uniquecharcount ) &&
1119 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
1120 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
1121 {
ab3bbdeb
YO
1122 PerlIO_printf( Perl_debug_log, "%*"UVXf,
1123 colwidth,
3dab1dad
YO
1124 (UV)trie->trans[ base + ofs - trie->uniquecharcount ].next );
1125 } else {
ab3bbdeb 1126 PerlIO_printf( Perl_debug_log, "%*s",colwidth," ." );
3dab1dad
YO
1127 }
1128 }
1129
1130 PerlIO_printf( Perl_debug_log, "]");
1131
1132 }
1133 PerlIO_printf( Perl_debug_log, "\n" );
1134 }
2e64971a
DM
1135 PerlIO_printf(Perl_debug_log, "%*sword_info N:(prev,len)=", (int)depth*2, "");
1136 for (word=1; word <= trie->wordcount; word++) {
1137 PerlIO_printf(Perl_debug_log, " %d:(%d,%d)",
1138 (int)word, (int)(trie->wordinfo[word].prev),
1139 (int)(trie->wordinfo[word].len));
1140 }
1141 PerlIO_printf(Perl_debug_log, "\n" );
3dab1dad
YO
1142}
1143/*
3dab1dad
YO
1144 Dumps a fully constructed but uncompressed trie in list form.
1145 List tries normally only are used for construction when the number of
1146 possible chars (trie->uniquecharcount) is very high.
1147 Used for debugging make_trie().
1148*/
1149STATIC void
55eed653 1150S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie,
2b8b4781
NC
1151 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1152 U32 depth)
3dab1dad
YO
1153{
1154 U32 state;
ab3bbdeb 1155 SV *sv=sv_newmortal();
55eed653 1156 int colwidth= widecharmap ? 6 : 4;
3dab1dad 1157 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
1158
1159 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_LIST;
1160
3dab1dad 1161 /* print out the table precompression. */
ab3bbdeb
YO
1162 PerlIO_printf( Perl_debug_log, "%*sState :Word | Transition Data\n%*s%s",
1163 (int)depth * 2 + 2,"", (int)depth * 2 + 2,"",
1164 "------:-----+-----------------\n" );
3dab1dad
YO
1165
1166 for( state=1 ; state < next_alloc ; state ++ ) {
1167 U16 charid;
1168
ab3bbdeb 1169 PerlIO_printf( Perl_debug_log, "%*s %4"UVXf" :",
3dab1dad
YO
1170 (int)depth * 2 + 2,"", (UV)state );
1171 if ( ! trie->states[ state ].wordnum ) {
1172 PerlIO_printf( Perl_debug_log, "%5s| ","");
1173 } else {
1174 PerlIO_printf( Perl_debug_log, "W%4x| ",
1175 trie->states[ state ].wordnum
1176 );
1177 }
1178 for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
2b8b4781 1179 SV ** const tmp = av_fetch( revcharmap, TRIE_LIST_ITEM(state,charid).forid, 0);
ab3bbdeb
YO
1180 if ( tmp ) {
1181 PerlIO_printf( Perl_debug_log, "%*s:%3X=%4"UVXf" | ",
1182 colwidth,
ddc5bc0f 1183 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
ab3bbdeb
YO
1184 PL_colors[0], PL_colors[1],
1185 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1186 PERL_PV_ESCAPE_FIRSTCHAR
1187 ) ,
1e2e3d02
YO
1188 TRIE_LIST_ITEM(state,charid).forid,
1189 (UV)TRIE_LIST_ITEM(state,charid).newstate
1190 );
1191 if (!(charid % 10))
664e119d
RGS
1192 PerlIO_printf(Perl_debug_log, "\n%*s| ",
1193 (int)((depth * 2) + 14), "");
1e2e3d02 1194 }
ab3bbdeb
YO
1195 }
1196 PerlIO_printf( Perl_debug_log, "\n");
3dab1dad
YO
1197 }
1198}
1199
1200/*
3dab1dad
YO
1201 Dumps a fully constructed but uncompressed trie in table form.
1202 This is the normal DFA style state transition table, with a few
1203 twists to facilitate compression later.
1204 Used for debugging make_trie().
1205*/
1206STATIC void
55eed653 1207S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie,
2b8b4781
NC
1208 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1209 U32 depth)
3dab1dad
YO
1210{
1211 U32 state;
1212 U16 charid;
ab3bbdeb 1213 SV *sv=sv_newmortal();
55eed653 1214 int colwidth= widecharmap ? 6 : 4;
3dab1dad 1215 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
1216
1217 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_TABLE;
3dab1dad
YO
1218
1219 /*
1220 print out the table precompression so that we can do a visual check
1221 that they are identical.
1222 */
1223
1224 PerlIO_printf( Perl_debug_log, "%*sChar : ",(int)depth * 2 + 2,"" );
1225
1226 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
2b8b4781 1227 SV ** const tmp = av_fetch( revcharmap, charid, 0);
3dab1dad 1228 if ( tmp ) {
ab3bbdeb
YO
1229 PerlIO_printf( Perl_debug_log, "%*s",
1230 colwidth,
ddc5bc0f 1231 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
ab3bbdeb
YO
1232 PL_colors[0], PL_colors[1],
1233 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1234 PERL_PV_ESCAPE_FIRSTCHAR
1235 )
1236 );
3dab1dad
YO
1237 }
1238 }
1239
1240 PerlIO_printf( Perl_debug_log, "\n%*sState+-",(int)depth * 2 + 2,"" );
1241
1242 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
ab3bbdeb 1243 PerlIO_printf( Perl_debug_log, "%.*s", colwidth,"--------");
3dab1dad
YO
1244 }
1245
1246 PerlIO_printf( Perl_debug_log, "\n" );
1247
1248 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
1249
1250 PerlIO_printf( Perl_debug_log, "%*s%4"UVXf" : ",
1251 (int)depth * 2 + 2,"",
1252 (UV)TRIE_NODENUM( state ) );
1253
1254 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
ab3bbdeb
YO
1255 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
1256 if (v)
1257 PerlIO_printf( Perl_debug_log, "%*"UVXf, colwidth, v );
1258 else
1259 PerlIO_printf( Perl_debug_log, "%*s", colwidth, "." );
3dab1dad
YO
1260 }
1261 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
1262 PerlIO_printf( Perl_debug_log, " (%4"UVXf")\n", (UV)trie->trans[ state ].check );
1263 } else {
1264 PerlIO_printf( Perl_debug_log, " (%4"UVXf") W%4X\n", (UV)trie->trans[ state ].check,
1265 trie->states[ TRIE_NODENUM( state ) ].wordnum );
1266 }
1267 }
07be1b83 1268}
3dab1dad
YO
1269
1270#endif
1271
2e64971a 1272
786e8c11
YO
1273/* make_trie(startbranch,first,last,tail,word_count,flags,depth)
1274 startbranch: the first branch in the whole branch sequence
1275 first : start branch of sequence of branch-exact nodes.
1276 May be the same as startbranch
1277 last : Thing following the last branch.
1278 May be the same as tail.
1279 tail : item following the branch sequence
1280 count : words in the sequence
1281 flags : currently the OP() type we will be building one of /EXACT(|F|Fl)/
1282 depth : indent depth
3dab1dad 1283
786e8c11 1284Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
07be1b83 1285
786e8c11
YO
1286A trie is an N'ary tree where the branches are determined by digital
1287decomposition of the key. IE, at the root node you look up the 1st character and
1288follow that branch repeat until you find the end of the branches. Nodes can be
1289marked as "accepting" meaning they represent a complete word. Eg:
07be1b83 1290
786e8c11 1291 /he|she|his|hers/
72f13be8 1292
786e8c11
YO
1293would convert into the following structure. Numbers represent states, letters
1294following numbers represent valid transitions on the letter from that state, if
1295the number is in square brackets it represents an accepting state, otherwise it
1296will be in parenthesis.
07be1b83 1297
786e8c11
YO
1298 +-h->+-e->[3]-+-r->(8)-+-s->[9]
1299 | |
1300 | (2)
1301 | |
1302 (1) +-i->(6)-+-s->[7]
1303 |
1304 +-s->(3)-+-h->(4)-+-e->[5]
07be1b83 1305
786e8c11
YO
1306 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
1307
1308This shows that when matching against the string 'hers' we will begin at state 1
1309read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
1310then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
1311is also accepting. Thus we know that we can match both 'he' and 'hers' with a
1312single traverse. We store a mapping from accepting to state to which word was
1313matched, and then when we have multiple possibilities we try to complete the
1314rest of the regex in the order in which they occured in the alternation.
1315
1316The only prior NFA like behaviour that would be changed by the TRIE support is
1317the silent ignoring of duplicate alternations which are of the form:
1318
1319 / (DUPE|DUPE) X? (?{ ... }) Y /x
1320
4b714af6 1321Thus EVAL blocks following a trie may be called a different number of times with
786e8c11 1322and without the optimisation. With the optimisations dupes will be silently
486ec47a 1323ignored. This inconsistent behaviour of EVAL type nodes is well established as
786e8c11
YO
1324the following demonstrates:
1325
1326 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
1327
1328which prints out 'word' three times, but
1329
1330 'words'=~/(word|word|word)(?{ print $1 })S/
1331
1332which doesnt print it out at all. This is due to other optimisations kicking in.
1333
1334Example of what happens on a structural level:
1335
486ec47a 1336The regexp /(ac|ad|ab)+/ will produce the following debug output:
786e8c11
YO
1337
1338 1: CURLYM[1] {1,32767}(18)
1339 5: BRANCH(8)
1340 6: EXACT <ac>(16)
1341 8: BRANCH(11)
1342 9: EXACT <ad>(16)
1343 11: BRANCH(14)
1344 12: EXACT <ab>(16)
1345 16: SUCCEED(0)
1346 17: NOTHING(18)
1347 18: END(0)
1348
1349This would be optimizable with startbranch=5, first=5, last=16, tail=16
1350and should turn into:
1351
1352 1: CURLYM[1] {1,32767}(18)
1353 5: TRIE(16)
1354 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
1355 <ac>
1356 <ad>
1357 <ab>
1358 16: SUCCEED(0)
1359 17: NOTHING(18)
1360 18: END(0)
1361
1362Cases where tail != last would be like /(?foo|bar)baz/:
1363
1364 1: BRANCH(4)
1365 2: EXACT <foo>(8)
1366 4: BRANCH(7)
1367 5: EXACT <bar>(8)
1368 7: TAIL(8)
1369 8: EXACT <baz>(10)
1370 10: END(0)
1371
1372which would be optimizable with startbranch=1, first=1, last=7, tail=8
1373and would end up looking like:
1374
1375 1: TRIE(8)
1376 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
1377 <foo>
1378 <bar>
1379 7: TAIL(8)
1380 8: EXACT <baz>(10)
1381 10: END(0)
1382
1383 d = uvuni_to_utf8_flags(d, uv, 0);
1384
1385is the recommended Unicode-aware way of saying
1386
1387 *(d++) = uv;
1388*/
1389
fab2782b 1390#define TRIE_STORE_REVCHAR(val) \
786e8c11 1391 STMT_START { \
73031816 1392 if (UTF) { \
fab2782b 1393 SV *zlopp = newSV(7); /* XXX: optimize me */ \
88c9ea1e 1394 unsigned char *flrbbbbb = (unsigned char *) SvPVX(zlopp); \
fab2782b 1395 unsigned const char *const kapow = uvuni_to_utf8(flrbbbbb, val); \
73031816
NC
1396 SvCUR_set(zlopp, kapow - flrbbbbb); \
1397 SvPOK_on(zlopp); \
1398 SvUTF8_on(zlopp); \
1399 av_push(revcharmap, zlopp); \
1400 } else { \
fab2782b 1401 char ooooff = (char)val; \
73031816
NC
1402 av_push(revcharmap, newSVpvn(&ooooff, 1)); \
1403 } \
1404 } STMT_END
786e8c11 1405
fab2782b
YO
1406#define TRIE_READ_CHAR STMT_START { \
1407 wordlen++; \
1408 if ( UTF ) { \
1409 /* if it is UTF then it is either already folded, or does not need folding */ \
1410 uvc = utf8n_to_uvuni( (const U8*) uc, UTF8_MAXLEN, &len, uniflags); \
1411 } \
1412 else if (folder == PL_fold_latin1) { \
1413 /* if we use this folder we have to obey unicode rules on latin-1 data */ \
1414 if ( foldlen > 0 ) { \
1415 uvc = utf8n_to_uvuni( (const U8*) scan, UTF8_MAXLEN, &len, uniflags ); \
1416 foldlen -= len; \
1417 scan += len; \
1418 len = 0; \
1419 } else { \
1420 len = 1; \
1421 uvc = _to_fold_latin1( (U8) *uc, foldbuf, &foldlen, 1); \
1422 skiplen = UNISKIP(uvc); \
1423 foldlen -= skiplen; \
1424 scan = foldbuf + skiplen; \
1425 } \
1426 } else { \
1427 /* raw data, will be folded later if needed */ \
1428 uvc = (U32)*uc; \
1429 len = 1; \
1430 } \
786e8c11
YO
1431} STMT_END
1432
1433
1434
1435#define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
1436 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
f9003953
NC
1437 U32 ging = TRIE_LIST_LEN( state ) *= 2; \
1438 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
786e8c11
YO
1439 } \
1440 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
1441 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
1442 TRIE_LIST_CUR( state )++; \
1443} STMT_END
07be1b83 1444
786e8c11
YO
1445#define TRIE_LIST_NEW(state) STMT_START { \
1446 Newxz( trie->states[ state ].trans.list, \
1447 4, reg_trie_trans_le ); \
1448 TRIE_LIST_CUR( state ) = 1; \
1449 TRIE_LIST_LEN( state ) = 4; \
1450} STMT_END
07be1b83 1451
786e8c11
YO
1452#define TRIE_HANDLE_WORD(state) STMT_START { \
1453 U16 dupe= trie->states[ state ].wordnum; \
1454 regnode * const noper_next = regnext( noper ); \
1455 \
786e8c11
YO
1456 DEBUG_r({ \
1457 /* store the word for dumping */ \
1458 SV* tmp; \
1459 if (OP(noper) != NOTHING) \
740cce10 1460 tmp = newSVpvn_utf8(STRING(noper), STR_LEN(noper), UTF); \
786e8c11 1461 else \
740cce10 1462 tmp = newSVpvn_utf8( "", 0, UTF ); \
2b8b4781 1463 av_push( trie_words, tmp ); \
786e8c11
YO
1464 }); \
1465 \
1466 curword++; \
2e64971a
DM
1467 trie->wordinfo[curword].prev = 0; \
1468 trie->wordinfo[curword].len = wordlen; \
1469 trie->wordinfo[curword].accept = state; \
786e8c11
YO
1470 \
1471 if ( noper_next < tail ) { \
1472 if (!trie->jump) \
c944940b 1473 trie->jump = (U16 *) PerlMemShared_calloc( word_count + 1, sizeof(U16) ); \
7f69552c 1474 trie->jump[curword] = (U16)(noper_next - convert); \
786e8c11
YO
1475 if (!jumper) \
1476 jumper = noper_next; \
1477 if (!nextbranch) \
1478 nextbranch= regnext(cur); \
1479 } \
1480 \
1481 if ( dupe ) { \
2e64971a
DM
1482 /* It's a dupe. Pre-insert into the wordinfo[].prev */\
1483 /* chain, so that when the bits of chain are later */\
1484 /* linked together, the dups appear in the chain */\
1485 trie->wordinfo[curword].prev = trie->wordinfo[dupe].prev; \
1486 trie->wordinfo[dupe].prev = curword; \
786e8c11
YO
1487 } else { \
1488 /* we haven't inserted this word yet. */ \
1489 trie->states[ state ].wordnum = curword; \
1490 } \
1491} STMT_END
07be1b83 1492
3dab1dad 1493
786e8c11
YO
1494#define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
1495 ( ( base + charid >= ucharcount \
1496 && base + charid < ubound \
1497 && state == trie->trans[ base - ucharcount + charid ].check \
1498 && trie->trans[ base - ucharcount + charid ].next ) \
1499 ? trie->trans[ base - ucharcount + charid ].next \
1500 : ( state==1 ? special : 0 ) \
1501 )
3dab1dad 1502
786e8c11
YO
1503#define MADE_TRIE 1
1504#define MADE_JUMP_TRIE 2
1505#define MADE_EXACT_TRIE 4
3dab1dad 1506
a3621e74 1507STATIC I32
786e8c11 1508S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch, regnode *first, regnode *last, regnode *tail, U32 word_count, U32 flags, U32 depth)
a3621e74 1509{
27da23d5 1510 dVAR;
a3621e74
YO
1511 /* first pass, loop through and scan words */
1512 reg_trie_data *trie;
55eed653 1513 HV *widecharmap = NULL;
2b8b4781 1514 AV *revcharmap = newAV();
a3621e74 1515 regnode *cur;
9f7f3913 1516 const U32 uniflags = UTF8_ALLOW_DEFAULT;
a3621e74
YO
1517 STRLEN len = 0;
1518 UV uvc = 0;
1519 U16 curword = 0;
1520 U32 next_alloc = 0;
786e8c11
YO
1521 regnode *jumper = NULL;
1522 regnode *nextbranch = NULL;
7f69552c 1523 regnode *convert = NULL;
2e64971a 1524 U32 *prev_states; /* temp array mapping each state to previous one */
a3621e74 1525 /* we just use folder as a flag in utf8 */
1e696034 1526 const U8 * folder = NULL;
a3621e74 1527
2b8b4781
NC
1528#ifdef DEBUGGING
1529 const U32 data_slot = add_data( pRExC_state, 4, "tuuu" );
1530 AV *trie_words = NULL;
1531 /* along with revcharmap, this only used during construction but both are
1532 * useful during debugging so we store them in the struct when debugging.
8e11feef 1533 */
2b8b4781
NC
1534#else
1535 const U32 data_slot = add_data( pRExC_state, 2, "tu" );
3dab1dad 1536 STRLEN trie_charcount=0;
3dab1dad 1537#endif
2b8b4781 1538 SV *re_trie_maxbuff;
a3621e74 1539 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
1540
1541 PERL_ARGS_ASSERT_MAKE_TRIE;
72f13be8
YO
1542#ifndef DEBUGGING
1543 PERL_UNUSED_ARG(depth);
1544#endif
a3621e74 1545
1e696034 1546 switch (flags) {
c46d03cf 1547 case EXACT: break;
2f7f8cb1 1548 case EXACTFA:
fab2782b
YO
1549 case EXACTFU_SS:
1550 case EXACTFU_TRICKYFOLD:
1e696034
KW
1551 case EXACTFU: folder = PL_fold_latin1; break;
1552 case EXACTF: folder = PL_fold; break;
1553 case EXACTFL: folder = PL_fold_locale; break;
fab2782b 1554 default: Perl_croak( aTHX_ "panic! In trie construction, unknown node type %u %s", (unsigned) flags, PL_reg_name[flags] );
1e696034
KW
1555 }
1556
c944940b 1557 trie = (reg_trie_data *) PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
a3621e74 1558 trie->refcount = 1;
3dab1dad 1559 trie->startstate = 1;
786e8c11 1560 trie->wordcount = word_count;
f8fc2ecf 1561 RExC_rxi->data->data[ data_slot ] = (void*)trie;
c944940b 1562 trie->charmap = (U16 *) PerlMemShared_calloc( 256, sizeof(U16) );
fab2782b 1563 if (flags == EXACT)
c944940b 1564 trie->bitmap = (char *) PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
2e64971a
DM
1565 trie->wordinfo = (reg_trie_wordinfo *) PerlMemShared_calloc(
1566 trie->wordcount+1, sizeof(reg_trie_wordinfo));
1567
a3621e74 1568 DEBUG_r({
2b8b4781 1569 trie_words = newAV();
a3621e74 1570 });
a3621e74 1571
0111c4fd 1572 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
a3621e74 1573 if (!SvIOK(re_trie_maxbuff)) {
0111c4fd 1574 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
a3621e74 1575 }
df826430 1576 DEBUG_TRIE_COMPILE_r({
3dab1dad 1577 PerlIO_printf( Perl_debug_log,
786e8c11 1578 "%*smake_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
3dab1dad
YO
1579 (int)depth * 2 + 2, "",
1580 REG_NODE_NUM(startbranch),REG_NODE_NUM(first),
786e8c11 1581 REG_NODE_NUM(last), REG_NODE_NUM(tail),
85c3142d 1582 (int)depth);
3dab1dad 1583 });
7f69552c
YO
1584
1585 /* Find the node we are going to overwrite */
1586 if ( first == startbranch && OP( last ) != BRANCH ) {
1587 /* whole branch chain */
1588 convert = first;
1589 } else {
1590 /* branch sub-chain */
1591 convert = NEXTOPER( first );
1592 }
1593
a3621e74
YO
1594 /* -- First loop and Setup --
1595
1596 We first traverse the branches and scan each word to determine if it
1597 contains widechars, and how many unique chars there are, this is
1598 important as we have to build a table with at least as many columns as we
1599 have unique chars.
1600
1601 We use an array of integers to represent the character codes 0..255
38a44b82 1602 (trie->charmap) and we use a an HV* to store Unicode characters. We use the
a3621e74
YO
1603 native representation of the character value as the key and IV's for the
1604 coded index.
1605
1606 *TODO* If we keep track of how many times each character is used we can
1607 remap the columns so that the table compression later on is more
3b753521 1608 efficient in terms of memory by ensuring the most common value is in the
a3621e74
YO
1609 middle and the least common are on the outside. IMO this would be better
1610 than a most to least common mapping as theres a decent chance the most
1611 common letter will share a node with the least common, meaning the node
486ec47a 1612 will not be compressible. With a middle is most common approach the worst
a3621e74
YO
1613 case is when we have the least common nodes twice.
1614
1615 */
1616
a3621e74 1617 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
df826430 1618 regnode *noper = NEXTOPER( cur );
e1ec3a88 1619 const U8 *uc = (U8*)STRING( noper );
df826430 1620 const U8 *e = uc + STR_LEN( noper );
a3621e74
YO
1621 STRLEN foldlen = 0;
1622 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
fab2782b 1623 STRLEN skiplen = 0;
2af232bd 1624 const U8 *scan = (U8*)NULL;
07be1b83 1625 U32 wordlen = 0; /* required init */
02daf0ab
YO
1626 STRLEN chars = 0;
1627 bool set_bit = trie->bitmap ? 1 : 0; /*store the first char in the bitmap?*/
a3621e74 1628
3dab1dad 1629 if (OP(noper) == NOTHING) {
df826430
YO
1630 regnode *noper_next= regnext(noper);
1631 if (noper_next != tail && OP(noper_next) == flags) {
1632 noper = noper_next;
1633 uc= (U8*)STRING(noper);
1634 e= uc + STR_LEN(noper);
1635 trie->minlen= STR_LEN(noper);
1636 } else {
1637 trie->minlen= 0;
1638 continue;
1639 }
3dab1dad 1640 }
df826430 1641
fab2782b 1642 if ( set_bit ) { /* bitmap only alloced when !(UTF&&Folding) */
02daf0ab
YO
1643 TRIE_BITMAP_SET(trie,*uc); /* store the raw first byte
1644 regardless of encoding */
fab2782b
YO
1645 if (OP( noper ) == EXACTFU_SS) {
1646 /* false positives are ok, so just set this */
1647 TRIE_BITMAP_SET(trie,0xDF);
1648 }
1649 }
a3621e74 1650 for ( ; uc < e ; uc += len ) {
3dab1dad 1651 TRIE_CHARCOUNT(trie)++;
a3621e74 1652 TRIE_READ_CHAR;
3dab1dad 1653 chars++;
a3621e74 1654 if ( uvc < 256 ) {
fab2782b
YO
1655 if ( folder ) {
1656 U8 folded= folder[ (U8) uvc ];
1657 if ( !trie->charmap[ folded ] ) {
1658 trie->charmap[ folded ]=( ++trie->uniquecharcount );
1659 TRIE_STORE_REVCHAR( folded );
1660 }
1661 }
a3621e74
YO
1662 if ( !trie->charmap[ uvc ] ) {
1663 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
fab2782b 1664 TRIE_STORE_REVCHAR( uvc );
a3621e74 1665 }
02daf0ab 1666 if ( set_bit ) {
62012aee
KW
1667 /* store the codepoint in the bitmap, and its folded
1668 * equivalent. */
fab2782b 1669 TRIE_BITMAP_SET(trie, uvc);
0921ee73
T
1670
1671 /* store the folded codepoint */
fab2782b 1672 if ( folder ) TRIE_BITMAP_SET(trie, folder[(U8) uvc ]);
0921ee73
T
1673
1674 if ( !UTF ) {
1675 /* store first byte of utf8 representation of
acdf4139
KW
1676 variant codepoints */
1677 if (! UNI_IS_INVARIANT(uvc)) {
1678 TRIE_BITMAP_SET(trie, UTF8_TWO_BYTE_HI(uvc));
0921ee73
T
1679 }
1680 }
02daf0ab
YO
1681 set_bit = 0; /* We've done our bit :-) */
1682 }
a3621e74
YO
1683 } else {
1684 SV** svpp;
55eed653
NC
1685 if ( !widecharmap )
1686 widecharmap = newHV();
a3621e74 1687
55eed653 1688 svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );
a3621e74
YO
1689
1690 if ( !svpp )
e4584336 1691 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%"UVXf, uvc );
a3621e74
YO
1692
1693 if ( !SvTRUE( *svpp ) ) {
1694 sv_setiv( *svpp, ++trie->uniquecharcount );
fab2782b 1695 TRIE_STORE_REVCHAR(uvc);
a3621e74
YO
1696 }
1697 }
1698 }
3dab1dad 1699 if( cur == first ) {
fab2782b
YO
1700 trie->minlen = chars;
1701 trie->maxlen = chars;
3dab1dad 1702 } else if (chars < trie->minlen) {
fab2782b 1703 trie->minlen = chars;
3dab1dad 1704 } else if (chars > trie->maxlen) {
fab2782b
YO
1705 trie->maxlen = chars;
1706 }
1707 if (OP( noper ) == EXACTFU_SS) {
1708 /* XXX: workaround - 'ss' could match "\x{DF}" so minlen could be 1 and not 2*/
1709 if (trie->minlen > 1)
1710 trie->minlen= 1;
1711 }
1712 if (OP( noper ) == EXACTFU_TRICKYFOLD) {
1713 /* XXX: workround - things like "\x{1FBE}\x{0308}\x{0301}" can match "\x{0390}"
1714 * - We assume that any such sequence might match a 2 byte string */
1715 if (trie->minlen > 2 )
1716 trie->minlen= 2;
3dab1dad
YO
1717 }
1718
a3621e74
YO
1719 } /* end first pass */
1720 DEBUG_TRIE_COMPILE_r(
3dab1dad
YO
1721 PerlIO_printf( Perl_debug_log, "%*sTRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
1722 (int)depth * 2 + 2,"",
55eed653 1723 ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
be8e71aa
YO
1724 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
1725 (int)trie->minlen, (int)trie->maxlen )
a3621e74 1726 );
a3621e74
YO
1727
1728 /*
1729 We now know what we are dealing with in terms of unique chars and
1730 string sizes so we can calculate how much memory a naive
0111c4fd
RGS
1731 representation using a flat table will take. If it's over a reasonable
1732 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
a3621e74
YO
1733 conservative but potentially much slower representation using an array
1734 of lists.
1735
1736 At the end we convert both representations into the same compressed
1737 form that will be used in regexec.c for matching with. The latter
1738 is a form that cannot be used to construct with but has memory
1739 properties similar to the list form and access properties similar
1740 to the table form making it both suitable for fast searches and
1741 small enough that its feasable to store for the duration of a program.
1742
1743 See the comment in the code where the compressed table is produced
1744 inplace from the flat tabe representation for an explanation of how
1745 the compression works.
1746
1747 */
1748
1749
2e64971a
DM
1750 Newx(prev_states, TRIE_CHARCOUNT(trie) + 2, U32);
1751 prev_states[1] = 0;
1752
3dab1dad 1753 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1) > SvIV(re_trie_maxbuff) ) {
a3621e74
YO
1754 /*
1755 Second Pass -- Array Of Lists Representation
1756
1757 Each state will be represented by a list of charid:state records
1758 (reg_trie_trans_le) the first such element holds the CUR and LEN
1759 points of the allocated array. (See defines above).
1760
1761 We build the initial structure using the lists, and then convert
1762 it into the compressed table form which allows faster lookups
1763 (but cant be modified once converted).
a3621e74
YO
1764 */
1765
a3621e74
YO
1766 STRLEN transcount = 1;
1767
1e2e3d02
YO
1768 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
1769 "%*sCompiling trie using list compiler\n",
1770 (int)depth * 2 + 2, ""));
686b73d4 1771
c944940b
JH
1772 trie->states = (reg_trie_state *)
1773 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
1774 sizeof(reg_trie_state) );
a3621e74
YO
1775 TRIE_LIST_NEW(1);
1776 next_alloc = 2;
1777
1778 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1779
df826430 1780 regnode *noper = NEXTOPER( cur );
c445ea15 1781 U8 *uc = (U8*)STRING( noper );
df826430 1782 const U8 *e = uc + STR_LEN( noper );
c445ea15
AL
1783 U32 state = 1; /* required init */
1784 U16 charid = 0; /* sanity init */
1785 U8 *scan = (U8*)NULL; /* sanity init */
1786 STRLEN foldlen = 0; /* required init */
07be1b83 1787 U32 wordlen = 0; /* required init */
c445ea15 1788 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
fab2782b 1789 STRLEN skiplen = 0;
c445ea15 1790
df826430
YO
1791 if (OP(noper) == NOTHING) {
1792 regnode *noper_next= regnext(noper);
1793 if (noper_next != tail && OP(noper_next) == flags) {
1794 noper = noper_next;
1795 uc= (U8*)STRING(noper);
1796 e= uc + STR_LEN(noper);
1797 }
1798 }
1799
3dab1dad 1800 if (OP(noper) != NOTHING) {
786e8c11 1801 for ( ; uc < e ; uc += len ) {
c445ea15 1802
786e8c11 1803 TRIE_READ_CHAR;
c445ea15 1804
786e8c11
YO
1805 if ( uvc < 256 ) {
1806 charid = trie->charmap[ uvc ];
c445ea15 1807 } else {
55eed653 1808 SV** const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
786e8c11
YO
1809 if ( !svpp ) {
1810 charid = 0;
1811 } else {
1812 charid=(U16)SvIV( *svpp );
1813 }
c445ea15 1814 }
786e8c11
YO
1815 /* charid is now 0 if we dont know the char read, or nonzero if we do */
1816 if ( charid ) {
a3621e74 1817
786e8c11
YO
1818 U16 check;
1819 U32 newstate = 0;
a3621e74 1820
786e8c11
YO
1821 charid--;
1822 if ( !trie->states[ state ].trans.list ) {
1823 TRIE_LIST_NEW( state );
c445ea15 1824 }
786e8c11
YO
1825 for ( check = 1; check <= TRIE_LIST_USED( state ); check++ ) {
1826 if ( TRIE_LIST_ITEM( state, check ).forid == charid ) {
1827 newstate = TRIE_LIST_ITEM( state, check ).newstate;
1828 break;
1829 }
1830 }
1831 if ( ! newstate ) {
1832 newstate = next_alloc++;
2e64971a 1833 prev_states[newstate] = state;
786e8c11
YO
1834 TRIE_LIST_PUSH( state, charid, newstate );
1835 transcount++;
1836 }
1837 state = newstate;
1838 } else {
1839 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
c445ea15 1840 }
a28509cc 1841 }
c445ea15 1842 }
3dab1dad 1843 TRIE_HANDLE_WORD(state);
a3621e74
YO
1844
1845 } /* end second pass */
1846
1e2e3d02
YO
1847 /* next alloc is the NEXT state to be allocated */
1848 trie->statecount = next_alloc;
c944940b
JH
1849 trie->states = (reg_trie_state *)
1850 PerlMemShared_realloc( trie->states,
1851 next_alloc
1852 * sizeof(reg_trie_state) );
a3621e74 1853
3dab1dad 1854 /* and now dump it out before we compress it */
2b8b4781
NC
1855 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
1856 revcharmap, next_alloc,
1857 depth+1)
1e2e3d02 1858 );
a3621e74 1859
c944940b
JH
1860 trie->trans = (reg_trie_trans *)
1861 PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
a3621e74
YO
1862 {
1863 U32 state;
a3621e74
YO
1864 U32 tp = 0;
1865 U32 zp = 0;
1866
1867
1868 for( state=1 ; state < next_alloc ; state ++ ) {
1869 U32 base=0;
1870
1871 /*
1872 DEBUG_TRIE_COMPILE_MORE_r(
1873 PerlIO_printf( Perl_debug_log, "tp: %d zp: %d ",tp,zp)
1874 );
1875 */
1876
1877 if (trie->states[state].trans.list) {
1878 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
1879 U16 maxid=minid;
a28509cc 1880 U16 idx;
a3621e74
YO
1881
1882 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
c445ea15
AL
1883 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
1884 if ( forid < minid ) {
1885 minid=forid;
1886 } else if ( forid > maxid ) {
1887 maxid=forid;
1888 }
a3621e74
YO
1889 }
1890 if ( transcount < tp + maxid - minid + 1) {
1891 transcount *= 2;
c944940b
JH
1892 trie->trans = (reg_trie_trans *)
1893 PerlMemShared_realloc( trie->trans,
446bd890
NC
1894 transcount
1895 * sizeof(reg_trie_trans) );
a3621e74
YO
1896 Zero( trie->trans + (transcount / 2), transcount / 2 , reg_trie_trans );
1897 }
1898 base = trie->uniquecharcount + tp - minid;
1899 if ( maxid == minid ) {
1900 U32 set = 0;
1901 for ( ; zp < tp ; zp++ ) {
1902 if ( ! trie->trans[ zp ].next ) {
1903 base = trie->uniquecharcount + zp - minid;
1904 trie->trans[ zp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1905 trie->trans[ zp ].check = state;
1906 set = 1;
1907 break;
1908 }
1909 }
1910 if ( !set ) {
1911 trie->trans[ tp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1912 trie->trans[ tp ].check = state;
1913 tp++;
1914 zp = tp;
1915 }
1916 } else {
1917 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
c445ea15 1918 const U32 tid = base - trie->uniquecharcount + TRIE_LIST_ITEM( state, idx ).forid;
a3621e74
YO
1919 trie->trans[ tid ].next = TRIE_LIST_ITEM( state, idx ).newstate;
1920 trie->trans[ tid ].check = state;
1921 }
1922 tp += ( maxid - minid + 1 );
1923 }
1924 Safefree(trie->states[ state ].trans.list);
1925 }
1926 /*
1927 DEBUG_TRIE_COMPILE_MORE_r(
1928 PerlIO_printf( Perl_debug_log, " base: %d\n",base);
1929 );
1930 */
1931 trie->states[ state ].trans.base=base;
1932 }
cc601c31 1933 trie->lasttrans = tp + 1;
a3621e74
YO
1934 }
1935 } else {
1936 /*
1937 Second Pass -- Flat Table Representation.
1938
1939 we dont use the 0 slot of either trans[] or states[] so we add 1 to each.
1940 We know that we will need Charcount+1 trans at most to store the data
1941 (one row per char at worst case) So we preallocate both structures
1942 assuming worst case.
1943
1944 We then construct the trie using only the .next slots of the entry
1945 structs.
1946
3b753521 1947 We use the .check field of the first entry of the node temporarily to
a3621e74
YO
1948 make compression both faster and easier by keeping track of how many non
1949 zero fields are in the node.
1950
1951 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
1952 transition.
1953
1954 There are two terms at use here: state as a TRIE_NODEIDX() which is a
1955 number representing the first entry of the node, and state as a
1956 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1) and
1957 TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3) if there
1958 are 2 entrys per node. eg:
1959
1960 A B A B
1961 1. 2 4 1. 3 7
1962 2. 0 3 3. 0 5
1963 3. 0 0 5. 0 0
1964 4. 0 0 7. 0 0
1965
1966 The table is internally in the right hand, idx form. However as we also
1967 have to deal with the states array which is indexed by nodenum we have to
1968 use TRIE_NODENUM() to convert.
1969
1970 */
1e2e3d02
YO
1971 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
1972 "%*sCompiling trie using table compiler\n",
1973 (int)depth * 2 + 2, ""));
3dab1dad 1974
c944940b
JH
1975 trie->trans = (reg_trie_trans *)
1976 PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
1977 * trie->uniquecharcount + 1,
1978 sizeof(reg_trie_trans) );
1979 trie->states = (reg_trie_state *)
1980 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
1981 sizeof(reg_trie_state) );
a3621e74
YO
1982 next_alloc = trie->uniquecharcount + 1;
1983
3dab1dad 1984
a3621e74
YO
1985 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1986
df826430 1987 regnode *noper = NEXTOPER( cur );
a28509cc 1988 const U8 *uc = (U8*)STRING( noper );
df826430 1989 const U8 *e = uc + STR_LEN( noper );
a3621e74
YO
1990
1991 U32 state = 1; /* required init */
1992
1993 U16 charid = 0; /* sanity init */
1994 U32 accept_state = 0; /* sanity init */
1995 U8 *scan = (U8*)NULL; /* sanity init */
1996
1997 STRLEN foldlen = 0; /* required init */
07be1b83 1998 U32 wordlen = 0; /* required init */
fab2782b 1999 STRLEN skiplen = 0;
a3621e74
YO
2000 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
2001
df826430
YO
2002 if (OP(noper) == NOTHING) {
2003 regnode *noper_next= regnext(noper);
2004 if (noper_next != tail && OP(noper_next) == flags) {
2005 noper = noper_next;
2006 uc= (U8*)STRING(noper);
2007 e= uc + STR_LEN(noper);
2008 }
2009 }
fab2782b 2010
3dab1dad 2011 if ( OP(noper) != NOTHING ) {
786e8c11 2012 for ( ; uc < e ; uc += len ) {
a3621e74 2013
786e8c11 2014 TRIE_READ_CHAR;
a3621e74 2015
786e8c11
YO
2016 if ( uvc < 256 ) {
2017 charid = trie->charmap[ uvc ];
2018 } else {
55eed653 2019 SV* const * const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
786e8c11 2020 charid = svpp ? (U16)SvIV(*svpp) : 0;
a3621e74 2021 }
786e8c11
YO
2022 if ( charid ) {
2023 charid--;
2024 if ( !trie->trans[ state + charid ].next ) {
2025 trie->trans[ state + charid ].next = next_alloc;
2026 trie->trans[ state ].check++;
2e64971a
DM
2027 prev_states[TRIE_NODENUM(next_alloc)]
2028 = TRIE_NODENUM(state);
786e8c11
YO
2029 next_alloc += trie->uniquecharcount;
2030 }
2031 state = trie->trans[ state + charid ].next;
2032 } else {
2033 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
2034 }
2035 /* charid is now 0 if we dont know the char read, or nonzero if we do */
a3621e74 2036 }
a3621e74 2037 }
3dab1dad
YO
2038 accept_state = TRIE_NODENUM( state );
2039 TRIE_HANDLE_WORD(accept_state);
a3621e74
YO
2040
2041 } /* end second pass */
2042
3dab1dad 2043 /* and now dump it out before we compress it */
2b8b4781
NC
2044 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
2045 revcharmap,
2046 next_alloc, depth+1));
a3621e74 2047
a3621e74
YO
2048 {
2049 /*
2050 * Inplace compress the table.*
2051
2052 For sparse data sets the table constructed by the trie algorithm will
2053 be mostly 0/FAIL transitions or to put it another way mostly empty.
2054 (Note that leaf nodes will not contain any transitions.)
2055
2056 This algorithm compresses the tables by eliminating most such
2057 transitions, at the cost of a modest bit of extra work during lookup:
2058
2059 - Each states[] entry contains a .base field which indicates the
2060 index in the state[] array wheres its transition data is stored.
2061
3b753521 2062 - If .base is 0 there are no valid transitions from that node.
a3621e74
YO
2063
2064 - If .base is nonzero then charid is added to it to find an entry in
2065 the trans array.
2066
2067 -If trans[states[state].base+charid].check!=state then the
2068 transition is taken to be a 0/Fail transition. Thus if there are fail
2069 transitions at the front of the node then the .base offset will point
2070 somewhere inside the previous nodes data (or maybe even into a node
2071 even earlier), but the .check field determines if the transition is
2072 valid.
2073
786e8c11 2074 XXX - wrong maybe?
a3621e74 2075 The following process inplace converts the table to the compressed
3b753521 2076 table: We first do not compress the root node 1,and mark all its
a3621e74 2077 .check pointers as 1 and set its .base pointer as 1 as well. This
3b753521
FN
2078 allows us to do a DFA construction from the compressed table later,
2079 and ensures that any .base pointers we calculate later are greater
2080 than 0.
a3621e74
YO
2081
2082 - We set 'pos' to indicate the first entry of the second node.
2083
2084 - We then iterate over the columns of the node, finding the first and
2085 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
2086 and set the .check pointers accordingly, and advance pos
2087 appropriately and repreat for the next node. Note that when we copy
2088 the next pointers we have to convert them from the original
2089 NODEIDX form to NODENUM form as the former is not valid post
2090 compression.
2091
2092 - If a node has no transitions used we mark its base as 0 and do not
2093 advance the pos pointer.
2094
2095 - If a node only has one transition we use a second pointer into the
2096 structure to fill in allocated fail transitions from other states.
2097 This pointer is independent of the main pointer and scans forward
2098 looking for null transitions that are allocated to a state. When it
2099 finds one it writes the single transition into the "hole". If the
786e8c11 2100 pointer doesnt find one the single transition is appended as normal.
a3621e74
YO
2101
2102 - Once compressed we can Renew/realloc the structures to release the
2103 excess space.
2104
2105 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
2106 specifically Fig 3.47 and the associated pseudocode.
2107
2108 demq
2109 */
a3b680e6 2110 const U32 laststate = TRIE_NODENUM( next_alloc );
a28509cc 2111 U32 state, charid;
a3621e74 2112 U32 pos = 0, zp=0;
1e2e3d02 2113 trie->statecount = laststate;
a3621e74
YO
2114
2115 for ( state = 1 ; state < laststate ; state++ ) {
2116 U8 flag = 0;
a28509cc
AL
2117 const U32 stateidx = TRIE_NODEIDX( state );
2118 const U32 o_used = trie->trans[ stateidx ].check;
2119 U32 used = trie->trans[ stateidx ].check;
a3621e74
YO
2120 trie->trans[ stateidx ].check = 0;
2121
2122 for ( charid = 0 ; used && charid < trie->uniquecharcount ; charid++ ) {
2123 if ( flag || trie->trans[ stateidx + charid ].next ) {
2124 if ( trie->trans[ stateidx + charid ].next ) {
2125 if (o_used == 1) {
2126 for ( ; zp < pos ; zp++ ) {
2127 if ( ! trie->trans[ zp ].next ) {
2128 break;
2129 }
2130 }
2131 trie->states[ state ].trans.base = zp + trie->uniquecharcount - charid ;
2132 trie->trans[ zp ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
2133 trie->trans[ zp ].check = state;
2134 if ( ++zp > pos ) pos = zp;
2135 break;
2136 }
2137 used--;
2138 }
2139 if ( !flag ) {
2140 flag = 1;
2141 trie->states[ state ].trans.base = pos + trie->uniquecharcount - charid ;
2142 }
2143 trie->trans[ pos ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
2144 trie->trans[ pos ].check = state;
2145 pos++;
2146 }
2147 }
2148 }
cc601c31 2149 trie->lasttrans = pos + 1;
c944940b
JH
2150 trie->states = (reg_trie_state *)
2151 PerlMemShared_realloc( trie->states, laststate
2152 * sizeof(reg_trie_state) );
a3621e74 2153 DEBUG_TRIE_COMPILE_MORE_r(
e4584336 2154 PerlIO_printf( Perl_debug_log,
3dab1dad
YO
2155 "%*sAlloc: %d Orig: %"IVdf" elements, Final:%"IVdf". Savings of %%%5.2f\n",
2156 (int)depth * 2 + 2,"",
2157 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1 ),
5d7488b2
AL
2158 (IV)next_alloc,
2159 (IV)pos,
a3621e74
YO
2160 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
2161 );
2162
2163 } /* end table compress */
2164 }
1e2e3d02
YO
2165 DEBUG_TRIE_COMPILE_MORE_r(
2166 PerlIO_printf(Perl_debug_log, "%*sStatecount:%"UVxf" Lasttrans:%"UVxf"\n",
2167 (int)depth * 2 + 2, "",
2168 (UV)trie->statecount,
2169 (UV)trie->lasttrans)
2170 );
cc601c31 2171 /* resize the trans array to remove unused space */
c944940b
JH
2172 trie->trans = (reg_trie_trans *)
2173 PerlMemShared_realloc( trie->trans, trie->lasttrans
2174 * sizeof(reg_trie_trans) );
a3621e74 2175
3b753521 2176 { /* Modify the program and insert the new TRIE node */
3dab1dad
YO
2177 U8 nodetype =(U8)(flags & 0xFF);
2178 char *str=NULL;
786e8c11 2179
07be1b83 2180#ifdef DEBUGGING
e62cc96a 2181 regnode *optimize = NULL;
7122b237
YO
2182#ifdef RE_TRACK_PATTERN_OFFSETS
2183
b57a0404
JH
2184 U32 mjd_offset = 0;
2185 U32 mjd_nodelen = 0;
7122b237
YO
2186#endif /* RE_TRACK_PATTERN_OFFSETS */
2187#endif /* DEBUGGING */
a3621e74 2188 /*
3dab1dad
YO
2189 This means we convert either the first branch or the first Exact,
2190 depending on whether the thing following (in 'last') is a branch
2191 or not and whther first is the startbranch (ie is it a sub part of
2192 the alternation or is it the whole thing.)
3b753521 2193 Assuming its a sub part we convert the EXACT otherwise we convert
3dab1dad 2194 the whole branch sequence, including the first.
a3621e74 2195 */
3dab1dad 2196 /* Find the node we are going to overwrite */
7f69552c 2197 if ( first != startbranch || OP( last ) == BRANCH ) {
07be1b83 2198 /* branch sub-chain */
3dab1dad 2199 NEXT_OFF( first ) = (U16)(last - first);
7122b237 2200#ifdef RE_TRACK_PATTERN_OFFSETS
07be1b83
YO
2201 DEBUG_r({
2202 mjd_offset= Node_Offset((convert));
2203 mjd_nodelen= Node_Length((convert));
2204 });
7122b237 2205#endif
7f69552c 2206 /* whole branch chain */
7122b237
YO
2207 }
2208#ifdef RE_TRACK_PATTERN_OFFSETS
2209 else {
7f69552c
YO
2210 DEBUG_r({
2211 const regnode *nop = NEXTOPER( convert );
2212 mjd_offset= Node_Offset((nop));
2213 mjd_nodelen= Node_Length((nop));
2214 });
07be1b83
YO
2215 }
2216 DEBUG_OPTIMISE_r(
2217 PerlIO_printf(Perl_debug_log, "%*sMJD offset:%"UVuf" MJD length:%"UVuf"\n",
2218 (int)depth * 2 + 2, "",
786e8c11 2219 (UV)mjd_offset, (UV)mjd_nodelen)
07be1b83 2220 );
7122b237 2221#endif
3dab1dad
YO
2222 /* But first we check to see if there is a common prefix we can
2223 split out as an EXACT and put in front of the TRIE node. */
2224 trie->startstate= 1;
55eed653 2225 if ( trie->bitmap && !widecharmap && !trie->jump ) {
3dab1dad 2226 U32 state;
1e2e3d02 2227 for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
a3621e74 2228 U32 ofs = 0;
8e11feef
RGS
2229 I32 idx = -1;
2230 U32 count = 0;
2231 const U32 base = trie->states[ state ].trans.base;
a3621e74 2232
3dab1dad 2233 if ( trie->states[state].wordnum )
8e11feef 2234 count = 1;
a3621e74 2235
8e11feef 2236 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
cc601c31
YO
2237 if ( ( base + ofs >= trie->uniquecharcount ) &&
2238 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
a3621e74
YO
2239 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
2240 {
3dab1dad 2241 if ( ++count > 1 ) {
2b8b4781 2242 SV **tmp = av_fetch( revcharmap, ofs, 0);
07be1b83 2243 const U8 *ch = (U8*)SvPV_nolen_const( *tmp );
8e11feef 2244 if ( state == 1 ) break;
3dab1dad
YO
2245 if ( count == 2 ) {
2246 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
2247 DEBUG_OPTIMISE_r(
8e11feef
RGS
2248 PerlIO_printf(Perl_debug_log,
2249 "%*sNew Start State=%"UVuf" Class: [",
2250 (int)depth * 2 + 2, "",
786e8c11 2251 (UV)state));
be8e71aa 2252 if (idx >= 0) {
2b8b4781 2253 SV ** const tmp = av_fetch( revcharmap, idx, 0);
be8e71aa 2254 const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );
8e11feef 2255
3dab1dad 2256 TRIE_BITMAP_SET(trie,*ch);
8e11feef
RGS
2257 if ( folder )
2258 TRIE_BITMAP_SET(trie, folder[ *ch ]);
3dab1dad 2259 DEBUG_OPTIMISE_r(
f1f66076 2260 PerlIO_printf(Perl_debug_log, "%s", (char*)ch)
3dab1dad 2261 );
8e11feef
RGS
2262 }
2263 }
2264 TRIE_BITMAP_SET(trie,*ch);
2265 if ( folder )
2266 TRIE_BITMAP_SET(trie,folder[ *ch ]);
2267 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"%s", ch));
2268 }
2269 idx = ofs;
2270 }
3dab1dad
YO
2271 }
2272 if ( count == 1 ) {
2b8b4781 2273 SV **tmp = av_fetch( revcharmap, idx, 0);
c490c714
YO
2274 STRLEN len;
2275 char *ch = SvPV( *tmp, len );
de734bd5
A
2276 DEBUG_OPTIMISE_r({
2277 SV *sv=sv_newmortal();
8e11feef
RGS
2278 PerlIO_printf( Perl_debug_log,
2279 "%*sPrefix State: %"UVuf" Idx:%"UVuf" Char='%s'\n",
2280 (int)depth * 2 + 2, "",
de734bd5
A
2281 (UV)state, (UV)idx,
2282 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), 6,
2283 PL_colors[0], PL_colors[1],
2284 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2285 PERL_PV_ESCAPE_FIRSTCHAR
2286 )
2287 );
2288 });
3dab1dad
YO
2289 if ( state==1 ) {
2290 OP( convert ) = nodetype;
2291 str=STRING(convert);
2292 STR_LEN(convert)=0;
2293 }
c490c714
YO
2294 STR_LEN(convert) += len;
2295 while (len--)
de734bd5 2296 *str++ = *ch++;
8e11feef 2297 } else {
f9049ba1 2298#ifdef DEBUGGING
8e11feef
RGS
2299 if (state>1)
2300 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"]\n"));
f9049ba1 2301#endif
8e11feef
RGS
2302 break;
2303 }
2304 }
2e64971a 2305 trie->prefixlen = (state-1);
3dab1dad 2306 if (str) {
8e11feef 2307 regnode *n = convert+NODE_SZ_STR(convert);
07be1b83 2308 NEXT_OFF(convert) = NODE_SZ_STR(convert);
8e11feef 2309 trie->startstate = state;
07be1b83
YO
2310 trie->minlen -= (state - 1);
2311 trie->maxlen -= (state - 1);
33809eae
JH
2312#ifdef DEBUGGING
2313 /* At least the UNICOS C compiler choked on this
2314 * being argument to DEBUG_r(), so let's just have
2315 * it right here. */
2316 if (
2317#ifdef PERL_EXT_RE_BUILD
2318 1
2319#else
2320 DEBUG_r_TEST
2321#endif
2322 ) {
2323 regnode *fix = convert;
2324 U32 word = trie->wordcount;
2325 mjd_nodelen++;
2326 Set_Node_Offset_Length(convert, mjd_offset, state - 1);
2327 while( ++fix < n ) {
2328 Set_Node_Offset_Length(fix, 0, 0);
2329 }
2330 while (word--) {
2331 SV ** const tmp = av_fetch( trie_words, word, 0 );
2332 if (tmp) {
2333 if ( STR_LEN(convert) <= SvCUR(*tmp) )
2334 sv_chop(*tmp, SvPV_nolen(*tmp) + STR_LEN(convert));
2335 else
2336 sv_chop(*tmp, SvPV_nolen(*tmp) + SvCUR(*tmp));
2337 }
2338 }
2339 }
2340#endif
8e11feef
RGS
2341 if (trie->maxlen) {
2342 convert = n;
2343 } else {
3dab1dad 2344 NEXT_OFF(convert) = (U16)(tail - convert);
a5ca303d 2345 DEBUG_r(optimize= n);
3dab1dad
YO
2346 }
2347 }
2348 }
a5ca303d
YO
2349 if (!jumper)
2350 jumper = last;
3dab1dad 2351 if ( trie->maxlen ) {
8e11feef
RGS
2352 NEXT_OFF( convert ) = (U16)(tail - convert);
2353 ARG_SET( convert, data_slot );
786e8c11
YO
2354 /* Store the offset to the first unabsorbed branch in
2355 jump[0], which is otherwise unused by the jump logic.
2356 We use this when dumping a trie and during optimisation. */
2357 if (trie->jump)
7f69552c 2358 trie->jump[0] = (U16)(nextbranch - convert);
a5ca303d 2359
6c48061a
YO
2360 /* If the start state is not accepting (meaning there is no empty string/NOTHING)
2361 * and there is a bitmap
2362 * and the first "jump target" node we found leaves enough room
2363 * then convert the TRIE node into a TRIEC node, with the bitmap
2364 * embedded inline in the opcode - this is hypothetically faster.
2365 */
2366 if ( !trie->states[trie->startstate].wordnum
2367 && trie->bitmap
2368 && ( (char *)jumper - (char *)convert) >= (int)sizeof(struct regnode_charclass) )
786e8c11
YO
2369 {
2370 OP( convert ) = TRIEC;
2371 Copy(trie->bitmap, ((struct regnode_charclass *)convert)->bitmap, ANYOF_BITMAP_SIZE, char);
446bd890 2372 PerlMemShared_free(trie->bitmap);
786e8c11
YO
2373 trie->bitmap= NULL;
2374 } else
2375 OP( convert ) = TRIE;
a3621e74 2376
3dab1dad
YO
2377 /* store the type in the flags */
2378 convert->flags = nodetype;
a5ca303d
YO
2379 DEBUG_r({
2380 optimize = convert
2381 + NODE_STEP_REGNODE
2382 + regarglen[ OP( convert ) ];
2383 });
2384 /* XXX We really should free up the resource in trie now,
2385 as we won't use them - (which resources?) dmq */
3dab1dad 2386 }
a3621e74 2387 /* needed for dumping*/
e62cc96a 2388 DEBUG_r(if (optimize) {
07be1b83 2389 regnode *opt = convert;
bcdf7404 2390
e62cc96a 2391 while ( ++opt < optimize) {
07be1b83
YO
2392 Set_Node_Offset_Length(opt,0,0);
2393 }
786e8c11
YO
2394 /*
2395 Try to clean up some of the debris left after the
2396 optimisation.
a3621e74 2397 */
786e8c11 2398 while( optimize < jumper ) {
07be1b83 2399 mjd_nodelen += Node_Length((optimize));
a3621e74 2400 OP( optimize ) = OPTIMIZED;
07be1b83 2401 Set_Node_Offset_Length(optimize,0,0);
a3621e74
YO
2402 optimize++;
2403 }
07be1b83 2404 Set_Node_Offset_Length(convert,mjd_offset,mjd_nodelen);
a3621e74
YO
2405 });
2406 } /* end node insert */
2e64971a
DM
2407
2408 /* Finish populating the prev field of the wordinfo array. Walk back
2409 * from each accept state until we find another accept state, and if
2410 * so, point the first word's .prev field at the second word. If the
2411 * second already has a .prev field set, stop now. This will be the
2412 * case either if we've already processed that word's accept state,
3b753521
FN
2413 * or that state had multiple words, and the overspill words were
2414 * already linked up earlier.
2e64971a
DM
2415 */
2416 {
2417 U16 word;
2418 U32 state;
2419 U16 prev;
2420
2421 for (word=1; word <= trie->wordcount; word++) {
2422 prev = 0;
2423 if (trie->wordinfo[word].prev)
2424 continue;
2425 state = trie->wordinfo[word].accept;
2426 while (state) {
2427 state = prev_states[state];
2428 if (!state)
2429 break;
2430 prev = trie->states[state].wordnum;
2431 if (prev)
2432 break;
2433 }
2434 trie->wordinfo[word].prev = prev;
2435 }
2436 Safefree(prev_states);
2437 }
2438
2439
2440 /* and now dump out the compressed format */
2441 DEBUG_TRIE_COMPILE_r(dump_trie(trie, widecharmap, revcharmap, depth+1));
2442
55eed653 2443 RExC_rxi->data->data[ data_slot + 1 ] = (void*)widecharmap;
2b8b4781
NC
2444#ifdef DEBUGGING
2445 RExC_rxi->data->data[ data_slot + TRIE_WORDS_OFFSET ] = (void*)trie_words;
2446 RExC_rxi->data->data[ data_slot + 3 ] = (void*)revcharmap;
2447#else
2448 SvREFCNT_dec(revcharmap);
07be1b83 2449#endif
786e8c11
YO
2450 return trie->jump
2451 ? MADE_JUMP_TRIE
2452 : trie->startstate>1
2453 ? MADE_EXACT_TRIE
2454 : MADE_TRIE;
2455}
2456
2457STATIC void
2458S_make_trie_failtable(pTHX_ RExC_state_t *pRExC_state, regnode *source, regnode *stclass, U32 depth)
2459{
3b753521 2460/* The Trie is constructed and compressed now so we can build a fail array if it's needed
786e8c11
YO
2461
2462 This is basically the Aho-Corasick algorithm. Its from exercise 3.31 and 3.32 in the
2463 "Red Dragon" -- Compilers, principles, techniques, and tools. Aho, Sethi, Ullman 1985/88
2464 ISBN 0-201-10088-6
2465
2466 We find the fail state for each state in the trie, this state is the longest proper
3b753521
FN
2467 suffix of the current state's 'word' that is also a proper prefix of another word in our
2468 trie. State 1 represents the word '' and is thus the default fail state. This allows
786e8c11
YO
2469 the DFA not to have to restart after its tried and failed a word at a given point, it
2470 simply continues as though it had been matching the other word in the first place.
2471 Consider
2472 'abcdgu'=~/abcdefg|cdgu/
2473 When we get to 'd' we are still matching the first word, we would encounter 'g' which would
3b753521
FN
2474 fail, which would bring us to the state representing 'd' in the second word where we would
2475 try 'g' and succeed, proceeding to match 'cdgu'.
786e8c11
YO
2476 */
2477 /* add a fail transition */
3251b653
NC
2478 const U32 trie_offset = ARG(source);
2479 reg_trie_data *trie=(reg_trie_data *)RExC_rxi->data->data[trie_offset];
786e8c11
YO
2480 U32 *q;
2481 const U32 ucharcount = trie->uniquecharcount;
1e2e3d02 2482 const U32 numstates = trie->statecount;
786e8c11
YO
2483 const U32 ubound = trie->lasttrans + ucharcount;
2484 U32 q_read = 0;
2485 U32 q_write = 0;
2486 U32 charid;
2487 U32 base = trie->states[ 1 ].trans.base;
2488 U32 *fail;
2489 reg_ac_data *aho;
2490 const U32 data_slot = add_data( pRExC_state, 1, "T" );
2491 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
2492
2493 PERL_ARGS_ASSERT_MAKE_TRIE_FAILTABLE;
786e8c11
YO
2494#ifndef DEBUGGING
2495 PERL_UNUSED_ARG(depth);
2496#endif
2497
2498
2499 ARG_SET( stclass, data_slot );
c944940b 2500 aho = (reg_ac_data *) PerlMemShared_calloc( 1, sizeof(reg_ac_data) );
f8fc2ecf 2501 RExC_rxi->data->data[ data_slot ] = (void*)aho;
3251b653 2502 aho->trie=trie_offset;
446bd890
NC
2503 aho->states=(reg_trie_state *)PerlMemShared_malloc( numstates * sizeof(reg_trie_state) );
2504 Copy( trie->states, aho->states, numstates, reg_trie_state );
786e8c11 2505 Newxz( q, numstates, U32);
c944940b 2506 aho->fail = (U32 *) PerlMemShared_calloc( numstates, sizeof(U32) );
786e8c11
YO
2507 aho->refcount = 1;
2508 fail = aho->fail;
2509 /* initialize fail[0..1] to be 1 so that we always have
2510 a valid final fail state */
2511 fail[ 0 ] = fail[ 1 ] = 1;
2512
2513 for ( charid = 0; charid < ucharcount ; charid++ ) {
2514 const U32 newstate = TRIE_TRANS_STATE( 1, base, ucharcount, charid, 0 );
2515 if ( newstate ) {
2516 q[ q_write ] = newstate;
2517 /* set to point at the root */
2518 fail[ q[ q_write++ ] ]=1;
2519 }
2520 }
2521 while ( q_read < q_write) {
2522 const U32 cur = q[ q_read++ % numstates ];
2523 base = trie->states[ cur ].trans.base;
2524
2525 for ( charid = 0 ; charid < ucharcount ; charid++ ) {
2526 const U32 ch_state = TRIE_TRANS_STATE( cur, base, ucharcount, charid, 1 );
2527 if (ch_state) {
2528 U32 fail_state = cur;
2529 U32 fail_base;
2530 do {
2531 fail_state = fail[ fail_state ];
2532 fail_base = aho->states[ fail_state ].trans.base;
2533 } while ( !TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 ) );
2534
2535 fail_state = TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 );
2536 fail[ ch_state ] = fail_state;
2537 if ( !aho->states[ ch_state ].wordnum && aho->states[ fail_state ].wordnum )
2538 {
2539 aho->states[ ch_state ].wordnum = aho->states[ fail_state ].wordnum;
2540 }
2541 q[ q_write++ % numstates] = ch_state;
2542 }
2543 }
2544 }
2545 /* restore fail[0..1] to 0 so that we "fall out" of the AC loop
2546 when we fail in state 1, this allows us to use the
2547 charclass scan to find a valid start char. This is based on the principle
2548 that theres a good chance the string being searched contains lots of stuff
2549 that cant be a start char.
2550 */
2551 fail[ 0 ] = fail[ 1 ] = 0;
2552 DEBUG_TRIE_COMPILE_r({
6d99fb9b
JH
2553 PerlIO_printf(Perl_debug_log,
2554 "%*sStclass Failtable (%"UVuf" states): 0",
2555 (int)(depth * 2), "", (UV)numstates
1e2e3d02 2556 );
786e8c11
YO
2557 for( q_read=1; q_read<numstates; q_read++ ) {
2558 PerlIO_printf(Perl_debug_log, ", %"UVuf, (UV)fail[q_read]);
2559 }
2560 PerlIO_printf(Perl_debug_log, "\n");
2561 });
2562 Safefree(q);
2563 /*RExC_seen |= REG_SEEN_TRIEDFA;*/
a3621e74
YO
2564}
2565
786e8c11 2566
a3621e74 2567/*
5d1c421c
JH
2568 * There are strange code-generation bugs caused on sparc64 by gcc-2.95.2.
2569 * These need to be revisited when a newer toolchain becomes available.
2570 */
2571#if defined(__sparc64__) && defined(__GNUC__)
2572# if __GNUC__ < 2 || (__GNUC__ == 2 && __GNUC_MINOR__ < 96)
2573# undef SPARC64_GCC_WORKAROUND
2574# define SPARC64_GCC_WORKAROUND 1
2575# endif
2576#endif
2577
07be1b83 2578#define DEBUG_PEEP(str,scan,depth) \
b515a41d 2579 DEBUG_OPTIMISE_r({if (scan){ \
07be1b83
YO
2580 SV * const mysv=sv_newmortal(); \
2581 regnode *Next = regnext(scan); \
2582 regprop(RExC_rx, mysv, scan); \
7f69552c 2583 PerlIO_printf(Perl_debug_log, "%*s" str ">%3d: %s (%d)\n", \
07be1b83
YO
2584 (int)depth*2, "", REG_NODE_NUM(scan), SvPV_nolen_const(mysv),\
2585 Next ? (REG_NODE_NUM(Next)) : 0 ); \
b515a41d 2586 }});
07be1b83 2587
1de06328 2588
bb914485
KW
2589/* The below joins as many adjacent EXACTish nodes as possible into a single
2590 * one, and looks for problematic sequences of characters whose folds vs.
2591 * non-folds have sufficiently different lengths, that the optimizer would be
2592 * fooled into rejecting legitimate matches of them, and the trie construction
87b8b349 2593 * code needs to handle specially. The joining is only done if:
bb914485
KW
2594 * 1) there is room in the current conglomerated node to entirely contain the
2595 * next one.
2596 * 2) they are the exact same node type
2597 *
87b8b349 2598 * The adjacent nodes actually may be separated by NOTHING-kind nodes, and
bb914485
KW
2599 * these get optimized out
2600 *
9d071ca8 2601 * If there are problematic code sequences, *min_subtract is set to the delta
2d608413
KW
2602 * number of characters that the minimum size of the node can be less than its
2603 * actual size. And, the node type of the result is changed to reflect that it
2604 * contains these sequences.
bb914485 2605 *
a0c4c608
KW
2606 * And *has_exactf_sharp_s is set to indicate whether or not the node is EXACTF
2607 * and contains LATIN SMALL LETTER SHARP S
f758bddf 2608 *
bb914485
KW
2609 * This is as good a place as any to discuss the design of handling these
2610 * problematic sequences. It's been wrong in Perl for a very long time. There
87b8b349
KW
2611 * are three code points currently in Unicode whose folded lengths differ so
2612 * much from the un-folded lengths that it causes problems for the optimizer
2613 * and trie construction. Why only these are problematic, and not others where
2614 * lengths also differ is something I (khw) do not understand. New versions of
2615 * Unicode might add more such code points. Hopefully the logic in
2616 * fold_grind.t that figures out what to test (in part by verifying that each
2617 * size-combination gets tested) will catch any that do come along, so they can
2618 * be added to the special handling below. The chances of new ones are
2619 * actually rather small, as most, if not all, of the world's scripts that have
2620 * casefolding have already been encoded by Unicode. Also, a number of
2621 * Unicode's decisions were made to allow compatibility with pre-existing
2622 * standards, and almost all of those have already been dealt with. These
2623 * would otherwise be the most likely candidates for generating further tricky
2624 * sequences. In other words, Unicode by itself is unlikely to add new ones
2625 * unless it is for compatibility with pre-existing standards, and there aren't
2626 * many of those left.
bb914485
KW
2627 *
2628 * The previous designs for dealing with these involved assigning a special
2629 * node for them. This approach doesn't work, as evidenced by this example:
a0c4c608 2630 * "\xDFs" =~ /s\xDF/ui # Used to fail before these patches
87b8b349 2631 * Both these fold to "sss", but if the pattern is parsed to create a node
bb914485
KW
2632 * that would match just the \xDF, it won't be able to handle the case where a
2633 * successful match would have to cross the node's boundary. The new approach
2634 * that hopefully generally solves the problem generates an EXACTFU_SS node
2635 * that is "sss".
2636 *
2637 * There are a number of components to the approach (a lot of work for just
2638 * three code points!):
2639 * 1) This routine examines each EXACTFish node that could contain the
9d071ca8
KW
2640 * problematic sequences. It returns in *min_subtract how much to
2641 * subtract from the the actual length of the string to get a real minimum
2642 * for one that could match it. This number is usually 0 except for the
2643 * problematic sequences. This delta is used by the caller to adjust the
2644 * min length of the match, and the delta between min and max, so that the
2645 * optimizer doesn't reject these possibilities based on size constraints.
87b8b349
KW
2646 * 2) These sequences require special handling by the trie code, so this code
2647 * changes the joined node type to special ops: EXACTFU_TRICKYFOLD and
2648 * EXACTFU_SS.
bb914485
KW
2649 * 3) This is sufficient for the two Greek sequences (described below), but
2650 * the one involving the Sharp s (\xDF) needs more. The node type
2651 * EXACTFU_SS is used for an EXACTFU node that contains at least one "ss"
2652 * sequence in it. For non-UTF-8 patterns and strings, this is the only
2653 * case where there is a possible fold length change. That means that a
2654 * regular EXACTFU node without UTF-8 involvement doesn't have to concern
2655 * itself with length changes, and so can be processed faster. regexec.c
2656 * takes advantage of this. Generally, an EXACTFish node that is in UTF-8
2657 * is pre-folded by regcomp.c. This saves effort in regex matching.
87b8b349
KW
2658 * However, the pre-folding isn't done for non-UTF8 patterns because the
2659 * fold of the MICRO SIGN requires UTF-8. Also what EXACTF and EXACTFL
2660 * nodes fold to isn't known until runtime. The fold possibilities for
2661 * the non-UTF8 patterns are quite simple, except for the sharp s. All
2662 * the ones that don't involve a UTF-8 target string are members of a
2663 * fold-pair, and arrays are set up for all of them so that the other
2664 * member of the pair can be found quickly. Code elsewhere in this file
2665 * makes sure that in EXACTFU nodes, the sharp s gets folded to 'ss', even
2666 * if the pattern isn't UTF-8. This avoids the issues described in the
2667 * next item.
bb914485
KW
2668 * 4) A problem remains for the sharp s in EXACTF nodes. Whether it matches
2669 * 'ss' or not is not knowable at compile time. It will match iff the
2670 * target string is in UTF-8, unlike the EXACTFU nodes, where it always
2671 * matches; and the EXACTFL and EXACTFA nodes where it never does. Thus
6012a526 2672 * it can't be folded to "ss" at compile time, unlike EXACTFU does (as
bb914485
KW
2673 * described in item 3). An assumption that the optimizer part of
2674 * regexec.c (probably unwittingly) makes is that a character in the
2675 * pattern corresponds to at most a single character in the target string.
2676 * (And I do mean character, and not byte here, unlike other parts of the
2677 * documentation that have never been updated to account for multibyte
2678 * Unicode.) This assumption is wrong only in this case, as all other
2679 * cases are either 1-1 folds when no UTF-8 is involved; or is true by
2680 * virtue of having this file pre-fold UTF-8 patterns. I'm
2681 * reluctant to try to change this assumption, so instead the code punts.
9d071ca8
KW
2682 * This routine examines EXACTF nodes for the sharp s, and returns a
2683 * boolean indicating whether or not the node is an EXACTF node that
2684 * contains a sharp s. When it is true, the caller sets a flag that later
2685 * causes the optimizer in this file to not set values for the floating
2686 * and fixed string lengths, and thus avoids the optimizer code in
2687 * regexec.c that makes the invalid assumption. Thus, there is no
2688 * optimization based on string lengths for EXACTF nodes that contain the
2689 * sharp s. This only happens for /id rules (which means the pattern
2690 * isn't in UTF-8).
bb914485 2691 */
1de06328 2692
9d071ca8 2693#define JOIN_EXACT(scan,min_subtract,has_exactf_sharp_s, flags) \
07be1b83 2694 if (PL_regkind[OP(scan)] == EXACT) \
9d071ca8 2695 join_exact(pRExC_state,(scan),(min_subtract),has_exactf_sharp_s, (flags),NULL,depth+1)
07be1b83 2696
be8e71aa 2697STATIC U32
9d071ca8 2698S_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
2699 /* Merge several consecutive EXACTish nodes into one. */
2700 regnode *n = regnext(scan);
2701 U32 stringok = 1;
2702 regnode *next = scan + NODE_SZ_STR(scan);
2703 U32 merged = 0;
2704 U32 stopnow = 0;
2705#ifdef DEBUGGING
2706 regnode *stop = scan;
72f13be8 2707 GET_RE_DEBUG_FLAGS_DECL;
f9049ba1 2708#else
d47053eb
RGS
2709 PERL_UNUSED_ARG(depth);
2710#endif
7918f24d
NC
2711
2712 PERL_ARGS_ASSERT_JOIN_EXACT;
d47053eb 2713#ifndef EXPERIMENTAL_INPLACESCAN
f9049ba1
SP
2714 PERL_UNUSED_ARG(flags);
2715 PERL_UNUSED_ARG(val);
07be1b83 2716#endif
07be1b83 2717 DEBUG_PEEP("join",scan,depth);
bb914485 2718
3f410cf6
KW
2719 /* Look through the subsequent nodes in the chain. Skip NOTHING, merge
2720 * EXACT ones that are mergeable to the current one. */
2721 while (n
2722 && (PL_regkind[OP(n)] == NOTHING
2723 || (stringok && OP(n) == OP(scan)))
07be1b83 2724 && NEXT_OFF(n)
3f410cf6
KW
2725 && NEXT_OFF(scan) + NEXT_OFF(n) < I16_MAX)
2726 {
07be1b83
YO
2727
2728 if (OP(n) == TAIL || n > next)
2729 stringok = 0;
2730 if (PL_regkind[OP(n)] == NOTHING) {
07be1b83
YO
2731 DEBUG_PEEP("skip:",n,depth);
2732 NEXT_OFF(scan) += NEXT_OFF(n);
2733 next = n + NODE_STEP_REGNODE;
2734#ifdef DEBUGGING
2735 if (stringok)
2736 stop = n;
2737#endif
2738 n = regnext(n);
2739 }
2740 else if (stringok) {
786e8c11 2741 const unsigned int oldl = STR_LEN(scan);
07be1b83 2742 regnode * const nnext = regnext(n);
b2230d39 2743
87b8b349
KW
2744 /* XXX I (khw) kind of doubt that this works on platforms where
2745 * U8_MAX is above 255 because of lots of other assumptions */
b2230d39
KW
2746 if (oldl + STR_LEN(n) > U8_MAX)
2747 break;
07be1b83
YO
2748
2749 DEBUG_PEEP("merg",n,depth);
07be1b83 2750 merged++;
b2230d39 2751
07be1b83
YO
2752 NEXT_OFF(scan) += NEXT_OFF(n);
2753 STR_LEN(scan) += STR_LEN(n);
2754 next = n + NODE_SZ_STR(n);
2755 /* Now we can overwrite *n : */
2756 Move(STRING(n), STRING(scan) + oldl, STR_LEN(n), char);
2757#ifdef DEBUGGING
2758 stop = next - 1;
2759#endif
2760 n = nnext;
2761 if (stopnow) break;
2762 }
2763
d47053eb
RGS
2764#ifdef EXPERIMENTAL_INPLACESCAN
2765 if (flags && !NEXT_OFF(n)) {
2766 DEBUG_PEEP("atch", val, depth);
2767 if (reg_off_by_arg[OP(n)]) {
2768 ARG_SET(n, val - n);
2769 }
2770 else {
2771 NEXT_OFF(n) = val - n;
2772 }
2773 stopnow = 1;
2774 }
07be1b83
YO
2775#endif
2776 }
2c2b7f86 2777
9d071ca8 2778 *min_subtract = 0;
f758bddf 2779 *has_exactf_sharp_s = FALSE;
f646642f 2780
3f410cf6
KW
2781 /* Here, all the adjacent mergeable EXACTish nodes have been merged. We
2782 * can now analyze for sequences of problematic code points. (Prior to
2783 * this final joining, sequences could have been split over boundaries, and
a0c4c608
KW
2784 * hence missed). The sequences only happen in folding, hence for any
2785 * non-EXACT EXACTish node */
86d6fcad 2786 if (OP(scan) != EXACT) {
f758bddf
KW
2787 U8 *s;
2788 U8 * s0 = (U8*) STRING(scan);
2789 U8 * const s_end = s0 + STR_LEN(scan);
2790
2791 /* The below is perhaps overboard, but this allows us to save a test
2792 * each time through the loop at the expense of a mask. This is
2793 * because on both EBCDIC and ASCII machines, 'S' and 's' differ by a
2794 * single bit. On ASCII they are 32 apart; on EBCDIC, they are 64.
2795 * This uses an exclusive 'or' to find that bit and then inverts it to
2796 * form a mask, with just a single 0, in the bit position where 'S' and
2797 * 's' differ. */
dbeb8947 2798 const U8 S_or_s_mask = (U8) ~ ('S' ^ 's');
f758bddf
KW
2799 const U8 s_masked = 's' & S_or_s_mask;
2800
2801 /* One pass is made over the node's string looking for all the
2802 * possibilities. to avoid some tests in the loop, there are two main
2803 * cases, for UTF-8 patterns (which can't have EXACTF nodes) and
2804 * non-UTF-8 */
2805 if (UTF) {
86d6fcad 2806
f758bddf
KW
2807 /* There are two problematic Greek code points in Unicode
2808 * casefolding
86d6fcad
KW
2809 *
2810 * U+0390 - GREEK SMALL LETTER IOTA WITH DIALYTIKA AND TONOS
2811 * U+03B0 - GREEK SMALL LETTER UPSILON WITH DIALYTIKA AND TONOS
2812 *
2813 * which casefold to
2814 *
2815 * Unicode UTF-8
2816 *
2817 * U+03B9 U+0308 U+0301 0xCE 0xB9 0xCC 0x88 0xCC 0x81
2818 * U+03C5 U+0308 U+0301 0xCF 0x85 0xCC 0x88 0xCC 0x81
2819 *
2820 * This means that in case-insensitive matching (or "loose
2d608413
KW
2821 * matching", as Unicode calls it), an EXACTF of length 3 chars can
2822 * match a target string of length 1 char. This would rather mess
2823 * up the minimum length computation.
86d6fcad 2824 *
f758bddf 2825 * If these sequences are found, the minimum length is decreased by
2d608413 2826 * two.
86d6fcad 2827 *
f758bddf
KW
2828 * Similarly, 'ss' may match the single char and byte LATIN SMALL
2829 * LETTER SHARP S. We decrease the min length by 1 for each
2830 * occurrence of 'ss' found */
3f410cf6 2831
61dad979
KW
2832#define U390_FIRST_BYTE GREEK_SMALL_LETTER_IOTA_UTF8_FIRST_BYTE
2833#define U3B0_FIRST_BYTE GREEK_SMALL_LETTER_UPSILON_UTF8_FIRST_BYTE
2834 const U8 U390_tail[] = GREEK_SMALL_LETTER_IOTA_UTF8_TAIL
2835 COMBINING_DIAERESIS_UTF8
2836 COMBINING_ACUTE_ACCENT_UTF8;
2837 const U8 U3B0_tail[] = GREEK_SMALL_LETTER_UPSILON_UTF8_TAIL
2838 COMBINING_DIAERESIS_UTF8
2839 COMBINING_ACUTE_ACCENT_UTF8;
2840 const U8 len = sizeof(U390_tail); /* (-1 for NUL; +1 for 1st byte;
f758bddf
KW
2841 yields a net of 0 */
2842 /* Examine the string for one of the problematic sequences */
2843 for (s = s0;
2844 s < s_end - 1; /* Can stop 1 before the end, as minimum length
2845 * sequence we are looking for is 2 */
2846 s += UTF8SKIP(s))
86d6fcad 2847 {
bb914485 2848
f758bddf
KW
2849 /* Look for the first byte in each problematic sequence */
2850 switch (*s) {
2851 /* We don't have to worry about other things that fold to
2852 * 's' (such as the long s, U+017F), as all above-latin1
2853 * code points have been pre-folded */
2854 case 's':
2855 case 'S':
2856
a0c4c608
KW
2857 /* Current character is an 's' or 'S'. If next one is
2858 * as well, we have the dreaded sequence */
f758bddf
KW
2859 if (((*(s+1) & S_or_s_mask) == s_masked)
2860 /* These two node types don't have special handling
2861 * for 'ss' */
2862 && OP(scan) != EXACTFL && OP(scan) != EXACTFA)
2863 {
9d071ca8 2864 *min_subtract += 1;
f758bddf
KW
2865 OP(scan) = EXACTFU_SS;
2866 s++; /* No need to look at this character again */
2867 }
2868 break;
2869
61dad979 2870 case U390_FIRST_BYTE:
f758bddf
KW
2871 if (s_end - s >= len
2872
2873 /* The 1's are because are skipping comparing the
2874 * first byte */
2875 && memEQ(s + 1, U390_tail, len - 1))
2876 {
2877 goto greek_sequence;
2878 }
2879 break;
2880
61dad979 2881 case U3B0_FIRST_BYTE:
f758bddf
KW
2882 if (! (s_end - s >= len
2883 && memEQ(s + 1, U3B0_tail, len - 1)))
2884 {
2885 break;
2886 }
2887 greek_sequence:
2d608413 2888 *min_subtract += 2;
f758bddf 2889
87b8b349 2890 /* This requires special handling by trie's, so change
f758bddf
KW
2891 * the node type to indicate this. If EXACTFA and
2892 * EXACTFL were ever to be handled by trie's, this
2893 * would have to be changed. If this node has already
2894 * been changed to EXACTFU_SS in this loop, leave it as
2895 * is. (I (khw) think it doesn't matter in regexec.c
2896 * for UTF patterns, but no need to change it */
2897 if (OP(scan) == EXACTFU) {
fab2782b 2898 OP(scan) = EXACTFU_TRICKYFOLD;
f758bddf
KW
2899 }
2900 s += 6; /* We already know what this sequence is. Skip
2901 the rest of it */
2902 break;
bb914485
KW
2903 }
2904 }
2905 }
f758bddf 2906 else if (OP(scan) != EXACTFL && OP(scan) != EXACTFA) {
bb914485 2907
f758bddf
KW
2908 /* Here, the pattern is not UTF-8. We need to look only for the
2909 * 'ss' sequence, and in the EXACTF case, the sharp s, which can be
2910 * in the final position. Otherwise we can stop looking 1 byte
2911 * earlier because have to find both the first and second 's' */
2912 const U8* upper = (OP(scan) == EXACTF) ? s_end : s_end -1;
2913
2914 for (s = s0; s < upper; s++) {
2915 switch (*s) {
2916 case 'S':
2917 case 's':
2918 if (s_end - s > 1
2919 && ((*(s+1) & S_or_s_mask) == s_masked))
2920 {
9d071ca8 2921 *min_subtract += 1;
f758bddf
KW
2922
2923 /* EXACTF nodes need to know that the minimum
2924 * length changed so that a sharp s in the string
2925 * can match this ss in the pattern, but they
87b8b349
KW
2926 * remain EXACTF nodes, as they won't match this
2927 * unless the target string is is UTF-8, which we
2928 * don't know until runtime */
f758bddf
KW
2929 if (OP(scan) != EXACTF) {
2930 OP(scan) = EXACTFU_SS;
2931 }
2932 s++;
2933 }
2934 break;
2935 case LATIN_SMALL_LETTER_SHARP_S:
2936 if (OP(scan) == EXACTF) {
2937 *has_exactf_sharp_s = TRUE;
2938 }
2939 break;
86d6fcad
KW
2940 }
2941 }
2942 }
07be1b83 2943 }
3f410cf6 2944
07be1b83 2945#ifdef DEBUGGING
bb789b09
DM
2946 /* Allow dumping but overwriting the collection of skipped
2947 * ops and/or strings with fake optimized ops */
07be1b83
YO
2948 n = scan + NODE_SZ_STR(scan);
2949 while (n <= stop) {
bb789b09
DM
2950 OP(n) = OPTIMIZED;
2951 FLAGS(n) = 0;
2952 NEXT_OFF(n) = 0;
07be1b83
YO
2953 n++;
2954 }
2955#endif
2956 DEBUG_OPTIMISE_r(if (merged){DEBUG_PEEP("finl",scan,depth)});
2957 return stopnow;
2958}
2959
486ec47a 2960/* REx optimizer. Converts nodes into quicker variants "in place".
653099ff
GS
2961 Finds fixed substrings. */
2962
a0288114 2963/* Stops at toplevel WHILEM as well as at "last". At end *scanp is set
c277df42
IZ
2964 to the position after last scanned or to NULL. */
2965
40d049e4
YO
2966#define INIT_AND_WITHP \
2967 assert(!and_withp); \
2968 Newx(and_withp,1,struct regnode_charclass_class); \
2969 SAVEFREEPV(and_withp)
07be1b83 2970
b515a41d 2971/* this is a chain of data about sub patterns we are processing that
486ec47a 2972 need to be handled separately/specially in study_chunk. Its so
b515a41d
YO
2973 we can simulate recursion without losing state. */
2974struct scan_frame;
2975typedef struct scan_frame {
2976 regnode *last; /* last node to process in this frame */
2977 regnode *next; /* next node to process when last is reached */
2978 struct scan_frame *prev; /*previous frame*/
2979 I32 stop; /* what stopparen do we use */
2980} scan_frame;
2981
304ee84b
YO
2982
2983#define SCAN_COMMIT(s, data, m) scan_commit(s, data, m, is_inf)
2984
e1d1eefb
YO
2985#define CASE_SYNST_FNC(nAmE) \
2986case nAmE: \
2987 if (flags & SCF_DO_STCLASS_AND) { \
2988 for (value = 0; value < 256; value++) \
2989 if (!is_ ## nAmE ## _cp(value)) \
2990 ANYOF_BITMAP_CLEAR(data->start_class, value); \
2991 } \
2992 else { \
2993 for (value = 0; value < 256; value++) \
2994 if (is_ ## nAmE ## _cp(value)) \
2995 ANYOF_BITMAP_SET(data->start_class, value); \
2996 } \
2997 break; \
2998case N ## nAmE: \
2999 if (flags & SCF_DO_STCLASS_AND) { \
3000 for (value = 0; value < 256; value++) \
3001 if (is_ ## nAmE ## _cp(value)) \
3002 ANYOF_BITMAP_CLEAR(data->start_class, value); \
3003 } \
3004 else { \
3005 for (value = 0; value < 256; value++) \
3006 if (!is_ ## nAmE ## _cp(value)) \
3007 ANYOF_BITMAP_SET(data->start_class, value); \
3008 } \
3009 break
3010
3011
3012
76e3520e 3013STATIC I32
40d049e4 3014S_study_chunk(pTHX_ RExC_state_t *pRExC_state, regnode **scanp,
1de06328 3015 I32 *minlenp, I32 *deltap,
40d049e4
YO
3016 regnode *last,
3017 scan_data_t *data,
3018 I32 stopparen,
3019 U8* recursed,
3020 struct regnode_charclass_class *and_withp,
3021 U32 flags, U32 depth)
c277df42
IZ
3022 /* scanp: Start here (read-write). */
3023 /* deltap: Write maxlen-minlen here. */
3024 /* last: Stop before this one. */
40d049e4
YO
3025 /* data: string data about the pattern */
3026 /* stopparen: treat close N as END */
3027 /* recursed: which subroutines have we recursed into */
3028 /* and_withp: Valid if flags & SCF_DO_STCLASS_OR */
c277df42 3029{
97aff369 3030 dVAR;
2d608413
KW
3031 I32 min = 0; /* There must be at least this number of characters to match */
3032 I32 pars = 0, code;
c277df42
IZ
3033 regnode *scan = *scanp, *next;
3034 I32 delta = 0;
3035 int is_inf = (flags & SCF_DO_SUBSTR) && (data->flags & SF_IS_INF);
aca2d497 3036 int is_inf_internal = 0; /* The studied chunk is infinite */
c277df42
IZ
3037 I32 is_par = OP(scan) == OPEN ? ARG(scan) : 0;
3038 scan_data_t data_fake;
a3621e74 3039 SV *re_trie_maxbuff = NULL;
786e8c11 3040 regnode *first_non_open = scan;
e2e6a0f1 3041 I32 stopmin = I32_MAX;
8aa23a47 3042 scan_frame *frame = NULL;
a3621e74 3043 GET_RE_DEBUG_FLAGS_DECL;
8aa23a47 3044
7918f24d
NC
3045 PERL_ARGS_ASSERT_STUDY_CHUNK;
3046
13a24bad 3047#ifdef DEBUGGING
40d049e4 3048 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
13a24bad 3049#endif
40d049e4 3050
786e8c11 3051 if ( depth == 0 ) {
40d049e4 3052 while (first_non_open && OP(first_non_open) == OPEN)
786e8c11
YO
3053 first_non_open=regnext(first_non_open);
3054 }
3055
b81d288d 3056
8aa23a47
YO
3057 fake_study_recurse:
3058 while ( scan && OP(scan) != END && scan < last ){
2d608413
KW
3059 UV min_subtract = 0; /* How mmany chars to subtract from the minimum
3060 node length to get a real minimum (because
3061 the folded version may be shorter) */
f758bddf 3062 bool has_exactf_sharp_s = FALSE;
8aa23a47 3063 /* Peephole optimizer: */
304ee84b 3064 DEBUG_STUDYDATA("Peep:", data,depth);
8aa23a47 3065 DEBUG_PEEP("Peep",scan,depth);
a0c4c608
KW
3066
3067 /* Its not clear to khw or hv why this is done here, and not in the
3068 * clauses that deal with EXACT nodes. khw's guess is that it's
3069 * because of a previous design */
9d071ca8 3070 JOIN_EXACT(scan,&min_subtract, &has_exactf_sharp_s, 0);
8aa23a47
YO
3071
3072 /* Follow the next-chain of the current node and optimize
3073 away all the NOTHINGs from it. */
3074 if (OP(scan) != CURLYX) {
3075 const int max = (reg_off_by_arg[OP(scan)]
3076 ? I32_MAX
3077 /* I32 may be smaller than U16 on CRAYs! */
3078 : (I32_MAX < U16_MAX ? I32_MAX : U16_MAX));
3079 int off = (reg_off_by_arg[OP(scan)] ? ARG(scan) : NEXT_OFF(scan));
3080 int noff;
3081 regnode *n = scan;
686b73d4 3082
8aa23a47
YO
3083 /* Skip NOTHING and LONGJMP. */
3084 while ((n = regnext(n))
3085 && ((PL_regkind[OP(n)] == NOTHING && (noff = NEXT_OFF(n)))
3086 || ((OP(n) == LONGJMP) && (noff = ARG(n))))
3087 && off + noff < max)
3088 off += noff;
3089 if (reg_off_by_arg[OP(scan)])
3090 ARG(scan) = off;
3091 else
3092 NEXT_OFF(scan) = off;
3093 }
a3621e74 3094
c277df42 3095
8aa23a47
YO
3096
3097 /* The principal pseudo-switch. Cannot be a switch, since we
3098 look into several different things. */
3099 if (OP(scan) == BRANCH || OP(scan) == BRANCHJ
3100 || OP(scan) == IFTHEN) {
3101 next = regnext(scan);
3102 code = OP(scan);
3103 /* demq: the op(next)==code check is to see if we have "branch-branch" AFAICT */
686b73d4 3104
8aa23a47
YO
3105 if (OP(next) == code || code == IFTHEN) {
3106 /* NOTE - There is similar code to this block below for handling
3107 TRIE nodes on a re-study. If you change stuff here check there
3108 too. */
3109 I32 max1 = 0, min1 = I32_MAX, num = 0;
3110 struct regnode_charclass_class accum;
3111 regnode * const startbranch=scan;
686b73d4 3112
8aa23a47 3113 if (flags & SCF_DO_SUBSTR)
304ee84b 3114 SCAN_COMMIT(pRExC_state, data, minlenp); /* Cannot merge strings after this. */
8aa23a47 3115 if (flags & SCF_DO_STCLASS)
e755fd73 3116 cl_init_zero(pRExC_state, &accum);
8aa23a47
YO
3117
3118 while (OP(scan) == code) {
3119 I32 deltanext, minnext, f = 0, fake;
3120 struct regnode_charclass_class this_class;
3121
3122 num++;
3123 data_fake.flags = 0;
3124 if (data) {
3125 data_fake.whilem_c = data->whilem_c;
3126 data_fake.last_closep = data->last_closep;
3127 }
3128 else
3129 data_fake.last_closep = &fake;
58e23c8d
YO
3130
3131 data_fake.pos_delta = delta;
8aa23a47
YO
3132 next = regnext(scan);
3133 scan = NEXTOPER(scan);
3134 if (code != BRANCH)
c277df42 3135 scan = NEXTOPER(scan);
8aa23a47 3136 if (flags & SCF_DO_STCLASS) {
e755fd73 3137 cl_init(pRExC_state, &this_class);
8aa23a47
YO
3138 data_fake.start_class = &this_class;
3139 f = SCF_DO_STCLASS_AND;
58e23c8d 3140 }
8aa23a47
YO
3141 if (flags & SCF_WHILEM_VISITED_POS)
3142 f |= SCF_WHILEM_VISITED_POS;
3143
3144 /* we suppose the run is continuous, last=next...*/
3145 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
3146 next, &data_fake,
3147 stopparen, recursed, NULL, f,depth+1);
3148 if (min1 > minnext)
3149 min1 = minnext;
3150 if (max1 < minnext + deltanext)
3151 max1 = minnext + deltanext;
3152 if (deltanext == I32_MAX)
3153 is_inf = is_inf_internal = 1;
3154 scan = next;
3155 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
3156 pars++;
3157 if (data_fake.flags & SCF_SEEN_ACCEPT) {
3158 if ( stopmin > minnext)
3159 stopmin = min + min1;
3160 flags &= ~SCF_DO_SUBSTR;
3161 if (data)
3162 data->flags |= SCF_SEEN_ACCEPT;
3163 }
3164 if (data) {
3165 if (data_fake.flags & SF_HAS_EVAL)
3166 data->flags |= SF_HAS_EVAL;
3167 data->whilem_c = data_fake.whilem_c;
3dab1dad 3168 }
8aa23a47 3169 if (flags & SCF_DO_STCLASS)
3fffb88a 3170 cl_or(pRExC_state, &accum, &this_class);
8aa23a47
YO
3171 }
3172 if (code == IFTHEN && num < 2) /* Empty ELSE branch */
3173 min1 = 0;
3174 if (flags & SCF_DO_SUBSTR) {
3175 data->pos_min += min1;
3176 data->pos_delta += max1 - min1;
3177 if (max1 != min1 || is_inf)
3178 data->longest = &(data->longest_float);
3179 }
3180 min += min1;
3181 delta += max1 - min1;
3182 if (flags & SCF_DO_STCLASS_OR) {
3fffb88a 3183 cl_or(pRExC_state, data->start_class, &accum);
8aa23a47
YO
3184 if (min1) {
3185 cl_and(data->start_class, and_withp);
3186 flags &= ~SCF_DO_STCLASS;
653099ff 3187 }
8aa23a47
YO
3188 }
3189 else if (flags & SCF_DO_STCLASS_AND) {
3190 if (min1) {
3191 cl_and(data->start_class, &accum);
3192 flags &= ~SCF_DO_STCLASS;
de0c8cb8 3193 }
8aa23a47
YO
3194 else {
3195 /* Switch to OR mode: cache the old value of
3196 * data->start_class */
3197 INIT_AND_WITHP;
3198 StructCopy(data->start_class, and_withp,
3199 struct regnode_charclass_class);
3200 flags &= ~SCF_DO_STCLASS_AND;
3201 StructCopy(&accum, data->start_class,
3202 struct regnode_charclass_class);
3203 flags |= SCF_DO_STCLASS_OR;
3204 data->start_class->flags |= ANYOF_EOS;
de0c8cb8 3205 }
8aa23a47 3206 }
a3621e74 3207
8aa23a47
YO
3208 if (PERL_ENABLE_TRIE_OPTIMISATION && OP( startbranch ) == BRANCH ) {
3209 /* demq.
a3621e74 3210
8aa23a47
YO
3211 Assuming this was/is a branch we are dealing with: 'scan' now
3212 points at the item that follows the branch sequence, whatever
3213 it is. We now start at the beginning of the sequence and look
3214 for subsequences of
a3621e74 3215
8aa23a47
YO
3216 BRANCH->EXACT=>x1
3217 BRANCH->EXACT=>x2
3218 tail
a3621e74 3219
8aa23a47 3220 which would be constructed from a pattern like /A|LIST|OF|WORDS/
a3621e74 3221
486ec47a 3222 If we can find such a subsequence we need to turn the first
8aa23a47
YO
3223 element into a trie and then add the subsequent branch exact
3224 strings to the trie.
a3621e74 3225
8aa23a47 3226 We have two cases
a3621e74 3227
3b753521 3228 1. patterns where the whole set of branches can be converted.
a3621e74 3229
8aa23a47 3230 2. patterns where only a subset can be converted.
a3621e74 3231
8aa23a47
YO
3232 In case 1 we can replace the whole set with a single regop
3233 for the trie. In case 2 we need to keep the start and end
3b753521 3234 branches so
a3621e74 3235
8aa23a47
YO
3236 'BRANCH EXACT; BRANCH EXACT; BRANCH X'
3237 becomes BRANCH TRIE; BRANCH X;
786e8c11 3238
8aa23a47
YO
3239 There is an additional case, that being where there is a
3240 common prefix, which gets split out into an EXACT like node
3241 preceding the TRIE node.
a3621e74 3242
8aa23a47
YO
3243 If x(1..n)==tail then we can do a simple trie, if not we make
3244 a "jump" trie, such that when we match the appropriate word
486ec47a 3245 we "jump" to the appropriate tail node. Essentially we turn
8aa23a47 3246 a nested if into a case structure of sorts.
b515a41d 3247
8aa23a47 3248 */
686b73d4 3249
8aa23a47
YO
3250 int made=0;
3251 if (!re_trie_maxbuff) {
3252 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
3253 if (!SvIOK(re_trie_maxbuff))
3254 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
3255 }
3256 if ( SvIV(re_trie_maxbuff)>=0 ) {
3257 regnode *cur;
3258 regnode *first = (regnode *)NULL;
3259 regnode *last = (regnode *)NULL;
3260 regnode *tail = scan;
fab2782b 3261 U8 trietype = 0;
8aa23a47 3262 U32 count=0;
a3621e74
YO
3263
3264#ifdef DEBUGGING
8aa23a47 3265 SV * const mysv = sv_newmortal(); /* for dumping */
a3621e74 3266#endif
8aa23a47
YO
3267 /* var tail is used because there may be a TAIL
3268 regop in the way. Ie, the exacts will point to the
3269 thing following the TAIL, but the last branch will
3270 point at the TAIL. So we advance tail. If we
3271 have nested (?:) we may have to move through several
3272 tails.
3273 */
3274
3275 while ( OP( tail ) == TAIL ) {
3276 /* this is the TAIL generated by (?:) */
3277 tail = regnext( tail );
3278 }
a3621e74 3279
8aa23a47 3280
df826430 3281 DEBUG_TRIE_COMPILE_r({
8aa23a47
YO
3282 regprop(RExC_rx, mysv, tail );
3283 PerlIO_printf( Perl_debug_log, "%*s%s%s\n",
3284 (int)depth * 2 + 2, "",
3285 "Looking for TRIE'able sequences. Tail node is: ",
3286 SvPV_nolen_const( mysv )
3287 );
3288 });
3289
3290 /*
3291
fab2782b
YO
3292 Step through the branches
3293 cur represents each branch,
3294 noper is the first thing to be matched as part of that branch
3295 noper_next is the regnext() of that node.
3296
3297 We normally handle a case like this /FOO[xyz]|BAR[pqr]/
3298 via a "jump trie" but we also support building with NOJUMPTRIE,
3299 which restricts the trie logic to structures like /FOO|BAR/.
3300
3301 If noper is a trieable nodetype then the branch is a possible optimization
3302 target. If we are building under NOJUMPTRIE then we require that noper_next
3303 is the same as scan (our current position in the regex program).
3304
3305 Once we have two or more consecutive such branches we can create a
3306 trie of the EXACT's contents and stitch it in place into the program.
3307
3308 If the sequence represents all of the branches in the alternation we
3309 replace the entire thing with a single TRIE node.
3310
3311 Otherwise when it is a subsequence we need to stitch it in place and
3312 replace only the relevant branches. This means the first branch has
3313 to remain as it is used by the alternation logic, and its next pointer,
3314 and needs to be repointed at the item on the branch chain following
3315 the last branch we have optimized away.
3316
3317 This could be either a BRANCH, in which case the subsequence is internal,
3318 or it could be the item following the branch sequence in which case the
3319 subsequence is at the end (which does not necessarily mean the first node
3320 is the start of the alternation).
3321
3322 TRIE_TYPE(X) is a define which maps the optype to a trietype.
3323
3324 optype | trietype
3325 ----------------+-----------
3326 NOTHING | NOTHING
3327 EXACT | EXACT
3328 EXACTFU | EXACTFU
3329 EXACTFU_SS | EXACTFU
3330 EXACTFU_TRICKYFOLD | EXACTFU
3331 EXACTFA | 0
3332
8aa23a47
YO
3333
3334 */
fab2782b
YO
3335#define TRIE_TYPE(X) ( ( NOTHING == (X) ) ? NOTHING : \
3336 ( EXACT == (X) ) ? EXACT : \
3337 ( EXACTFU == (X) || EXACTFU_SS == (X) || EXACTFU_TRICKYFOLD == (X) ) ? EXACTFU : \
3338 0 )
8aa23a47
YO
3339
3340 /* dont use tail as the end marker for this traverse */
3341 for ( cur = startbranch ; cur != scan ; cur = regnext( cur ) ) {
3342 regnode * const noper = NEXTOPER( cur );
fab2782b
YO
3343 U8 noper_type = OP( noper );
3344 U8 noper_trietype = TRIE_TYPE( noper_type );
b515a41d 3345#if defined(DEBUGGING) || defined(NOJUMPTRIE)
8aa23a47 3346 regnode * const noper_next = regnext( noper );
df826430
YO
3347 U8 noper_next_type = (noper_next && noper_next != tail) ? OP(noper_next) : 0;
3348 U8 noper_next_trietype = (noper_next && noper_next != tail) ? TRIE_TYPE( noper_next_type ) :0;
b515a41d
YO
3349#endif
3350
df826430 3351 DEBUG_TRIE_COMPILE_r({
8aa23a47
YO
3352 regprop(RExC_rx, mysv, cur);
3353 PerlIO_printf( Perl_debug_log, "%*s- %s (%d)",
3354 (int)depth * 2 + 2,"", SvPV_nolen_const( mysv ), REG_NODE_NUM(cur) );
3355
3356 regprop(RExC_rx, mysv, noper);
3357 PerlIO_printf( Perl_debug_log, " -> %s",
3358 SvPV_nolen_const(mysv));
3359
3360 if ( noper_next ) {
3361 regprop(RExC_rx, mysv, noper_next );
3362 PerlIO_printf( Perl_debug_log,"\t=> %s\t",
3363 SvPV_nolen_const(mysv));
3364 }
df826430
YO
3365 PerlIO_printf( Perl_debug_log, "(First==%d,Last==%d,Cur==%d,tt==%s,nt==%s,nnt==%s)\n",
3366 REG_NODE_NUM(first), REG_NODE_NUM(last), REG_NODE_NUM(cur),
3367 PL_reg_name[trietype], PL_reg_name[noper_trietype], PL_reg_name[noper_next_trietype]
3368 );
8aa23a47 3369 });
fab2782b
YO
3370
3371 /* Is noper a trieable nodetype that can be merged with the
3372 * current trie (if there is one)? */
3373 if ( noper_trietype
3374 &&
3375 (
df826430
YO
3376 ( noper_trietype == NOTHING)
3377 || ( trietype == NOTHING )
a40630bf 3378 || ( trietype == noper_trietype )
fab2782b 3379 )
786e8c11 3380#ifdef NOJUMPTRIE
8aa23a47 3381 && noper_next == tail
786e8c11 3382#endif
8aa23a47
YO
3383 && count < U16_MAX)
3384 {
fab2782b
YO
3385 /* Handle mergable triable node
3386 * Either we are the first node in a new trieable sequence,
3387 * in which case we do some bookkeeping, otherwise we update
3388 * the end pointer. */
fab2782b 3389 if ( !first ) {
3b6759a6 3390 first = cur;
df826430
YO
3391 if ( noper_trietype == NOTHING ) {
3392#if !defined(DEBUGGING) && !defined(NOJUMPTRIE)
3393 regnode * const noper_next = regnext( noper );
3b6759a6 3394 U8 noper_next_type = (noper_next && noper_next!=tail) ? OP(noper_next) : 0;
df826430
YO
3395 U8 noper_next_trietype = noper_next_type ? TRIE_TYPE( noper_next_type ) :0;
3396#endif
3397
190c1910 3398 if ( noper_next_trietype ) {
df826430 3399 trietype = noper_next_trietype;
190c1910
YO
3400 } else if (noper_next_type) {
3401 /* a NOTHING regop is 1 regop wide. We need at least two
3402 * for a trie so we can't merge this in */
3403 first = NULL;
3404 }
3405 } else {
3406 trietype = noper_trietype;
3b6759a6 3407 }
8aa23a47 3408 } else {
fab2782b
YO
3409 if ( trietype == NOTHING )
3410 trietype = noper_trietype;
8aa23a47
YO
3411 last = cur;
3412 }
df826430
YO
3413 if (first)
3414 count++;
fab2782b
YO
3415 } /* end handle mergable triable node */
3416 else {
3417 /* handle unmergable node -
3418 * noper may either be a triable node which can not be tried
3419 * together with the current trie, or a non triable node */
729aaeb5
YO
3420 if ( last ) {
3421 /* If last is set and trietype is not NOTHING then we have found
3422 * at least two triable branch sequences in a row of a similar
3423 * trietype so we can turn them into a trie. If/when we
3424 * allow NOTHING to start a trie sequence this condition will be
3425 * required, and it isn't expensive so we leave it in for now. */
e6351b37 3426 if ( trietype && trietype != NOTHING )
729aaeb5
YO
3427 make_trie( pRExC_state,
3428 startbranch, first, cur, tail, count,
3429 trietype, depth+1 );
fab2782b 3430 last = NULL; /* note: we clear/update first, trietype etc below, so we dont do it here */
8aa23a47 3431 }
fab2782b 3432 if ( noper_trietype
786e8c11 3433#ifdef NOJUMPTRIE
8aa23a47 3434 && noper_next == tail
786e8c11 3435#endif
8aa23a47 3436 ){
fab2782b 3437 /* noper is triable, so we can start a new trie sequence */
8aa23a47
YO
3438 count = 1;
3439 first = cur;
fab2782b
YO
3440 trietype = noper_trietype;
3441 } else if (first) {
3442 /* if we already saw a first but the current node is not triable then we have
3443 * to reset the first information. */
8aa23a47
YO
3444 count = 0;
3445 first = NULL;
fab2782b 3446 trietype = 0;
8aa23a47 3447 }
fab2782b
YO
3448 } /* end handle unmergable node */
3449 } /* loop over branches */
df826430 3450 DEBUG_TRIE_COMPILE_r({
8aa23a47
YO
3451 regprop(RExC_rx, mysv, cur);
3452 PerlIO_printf( Perl_debug_log,
3453 "%*s- %s (%d) <SCAN FINISHED>\n", (int)depth * 2 + 2,
3454 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
3455
3456 });
e6351b37 3457 if ( last && trietype ) {
3b6759a6
YO
3458 if ( trietype != NOTHING ) {
3459 /* the last branch of the sequence was part of a trie,
3460 * so we have to construct it here outside of the loop
3461 */
3462 made= make_trie( pRExC_state, startbranch, first, scan, tail, count, trietype, depth+1 );
686b73d4 3463#ifdef TRIE_STUDY_OPT
3b6759a6
YO
3464 if ( ((made == MADE_EXACT_TRIE &&
3465 startbranch == first)
3466 || ( first_non_open == first )) &&
3467 depth==0 ) {
3468 flags |= SCF_TRIE_RESTUDY;
3469 if ( startbranch == first
3470 && scan == tail )
3471 {
3472 RExC_seen &=~REG_TOP_LEVEL_BRANCHES;
3473 }
8aa23a47 3474 }
3dab1dad 3475#endif
3b6759a6
YO
3476 } else {
3477 /* at this point we know whatever we have is a NOTHING sequence/branch
3478 * AND if 'startbranch' is 'first' then we can turn the whole thing into a NOTHING
3479 */
3480 if ( startbranch == first ) {
3481 regnode *opt;
3482 /* the entire thing is a NOTHING sequence, something like this:
3483 * (?:|) So we can turn it into a plain NOTHING op. */
3484 DEBUG_TRIE_COMPILE_r({
3485 regprop(RExC_rx, mysv, cur);
3486 PerlIO_printf( Perl_debug_log,
3487 "%*s- %s (%d) <NOTHING BRANCH SEQUENCE>\n", (int)depth * 2 + 2,
3488 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
3489
3490 });
3491 OP(startbranch)= NOTHING;
3492 NEXT_OFF(startbranch)= tail - startbranch;
3493 for ( opt= startbranch + 1; opt < tail ; opt++ )
3494 OP(opt)= OPTIMIZED;
3495 }
3496 }
fab2782b
YO
3497 } /* end if ( last) */
3498 } /* TRIE_MAXBUF is non zero */
8aa23a47
YO
3499
3500 } /* do trie */
3501
653099ff 3502 }
8aa23a47
YO
3503 else if ( code == BRANCHJ ) { /* single branch is optimized. */
3504 scan = NEXTOPER(NEXTOPER(scan));
3505 } else /* single branch is optimized. */
3506 scan = NEXTOPER(scan);
3507 continue;
3508 } else if (OP(scan) == SUSPEND || OP(scan) == GOSUB || OP(scan) == GOSTART) {
3509 scan_frame *newframe = NULL;
3510 I32 paren;
3511 regnode *start;
3512 regnode *end;
3513
3514 if (OP(scan) != SUSPEND) {
3515 /* set the pointer */
3516 if (OP(scan) == GOSUB) {
3517 paren = ARG(scan);
3518 RExC_recurse[ARG2L(scan)] = scan;
3519 start = RExC_open_parens[paren-1];
3520 end = RExC_close_parens[paren-1];
3521 } else {
3522 paren = 0;
f8fc2ecf 3523 start = RExC_rxi->program + 1;
8aa23a47
YO
3524 end = RExC_opend;
3525 }
3526 if (!recursed) {
3527 Newxz(recursed, (((RExC_npar)>>3) +1), U8);
3528 SAVEFREEPV(recursed);
3529 }
3530 if (!PAREN_TEST(recursed,paren+1)) {
3531 PAREN_SET(recursed,paren+1);
3532 Newx(newframe,1,scan_frame);
3533 } else {
3534 if (flags & SCF_DO_SUBSTR) {
304ee84b 3535 SCAN_COMMIT(pRExC_state,data,minlenp);
8aa23a47
YO
3536 data->longest = &(data->longest_float);
3537 }
3538 is_inf = is_inf_internal = 1;
3539 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
3fffb88a 3540 cl_anything(pRExC_state, data->start_class);
8aa23a47
YO
3541 flags &= ~SCF_DO_STCLASS;
3542 }
3543 } else {
3544 Newx(newframe,1,scan_frame);
3545 paren = stopparen;
3546 start = scan+2;
3547 end = regnext(scan);
3548 }
3549 if (newframe) {
3550 assert(start);
3551 assert(end);
3552 SAVEFREEPV(newframe);
3553 newframe->next = regnext(scan);
3554 newframe->last = last;
3555 newframe->stop = stopparen;
3556 newframe->prev = frame;
3557
3558 frame = newframe;
3559 scan = start;
3560 stopparen = paren;
3561 last = end;
3562
3563 continue;
3564 }
3565 }
3566 else if (OP(scan) == EXACT) {
3567 I32 l = STR_LEN(scan);
3568 UV uc;
3569 if (UTF) {
3570 const U8 * const s = (U8*)STRING(scan);
4b88fb76 3571 uc = utf8_to_uvchr_buf(s, s + l, NULL);
8aa23a47 3572 l = utf8_length(s, s + l);
8aa23a47
YO
3573 } else {
3574 uc = *((U8*)STRING(scan));
3575 }
3576 min += l;
3577 if (flags & SCF_DO_SUBSTR) { /* Update longest substr. */
3578 /* The code below prefers earlier match for fixed
3579 offset, later match for variable offset. */
3580 if (data->last_end == -1) { /* Update the start info. */
3581