This is a live mirror of the Perl 5 development currently hosted at https://github.com/perl/perl5
regcomp.c; ([?...]) are single char width
[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
c277df42 77#define REG_COMP_C
54df2634
NC
78#ifdef PERL_IN_XSUB_RE
79# include "re_comp.h"
afda64e8 80EXTERN_C const struct regexp_engine my_reg_engine;
54df2634
NC
81#else
82# include "regcomp.h"
83#endif
a687059c 84
b992490d 85#include "invlist_inline.h"
1b0f46bf 86#include "unicode_constants.h"
04e98a4d 87
538e84ed
KW
88#define HAS_NONLATIN1_FOLD_CLOSURE(i) \
89 _HAS_NONLATIN1_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)
f12c0118
KW
90#define HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(i) \
91 _HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)
26faadbd 92#define IS_NON_FINAL_FOLD(c) _IS_NON_FINAL_FOLD_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
2c61f163 93#define IS_IN_SOME_FOLD_L1(c) _IS_IN_SOME_FOLD_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
94dc5c2d 94
a687059c
LW
95#ifndef STATIC
96#define STATIC static
97#endif
98
3f910e62
YO
99/* this is a chain of data about sub patterns we are processing that
100 need to be handled separately/specially in study_chunk. Its so
101 we can simulate recursion without losing state. */
102struct scan_frame;
103typedef struct scan_frame {
104 regnode *last_regnode; /* last node to process in this frame */
105 regnode *next_regnode; /* next node to process when last is reached */
106 U32 prev_recursed_depth;
107 I32 stopparen; /* what stopparen do we use */
3f910e62
YO
108
109 struct scan_frame *this_prev_frame; /* this previous frame */
110 struct scan_frame *prev_frame; /* previous frame */
111 struct scan_frame *next_frame; /* next frame */
112} scan_frame;
113
f4ae5a27
KW
114/* Certain characters are output as a sequence with the first being a
115 * backslash. */
4aada8b9 116#define isBACKSLASHED_PUNCT(c) memCHRs("-[]\\^", c)
f4ae5a27
KW
117
118
09b2b2e6 119struct RExC_state_t {
514a91f1
DM
120 U32 flags; /* RXf_* are we folding, multilining? */
121 U32 pm_flags; /* PMf_* stuff from the calling PMOP */
830247a4 122 char *precomp; /* uncompiled string. */
711b303b 123 char *precomp_end; /* pointer to end of uncompiled string. */
288b8c02 124 REGEXP *rx_sv; /* The SV that is the regexp. */
f8fc2ecf 125 regexp *rx; /* perl core regexp structure */
538e84ed
KW
126 regexp_internal *rxi; /* internal data for regexp object
127 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. */
51684e3c
KW
131 char *copy_start; /* start of copy of input within
132 constructed parse string */
cc16d262
KW
133 char *save_copy_start; /* Provides one level of saving
134 and restoring 'copy_start' */
51684e3c
KW
135 char *copy_start_in_input; /* Position in input string
136 corresponding to copy_start */
ea3daa5d 137 SSize_t whilem_seen; /* number of WHILEM in this expr */
fac92740 138 regnode *emit_start; /* Start of emitted-code area */
f55b7b26 139 regnode_offset emit; /* Code-emit pointer */
830247a4
IZ
140 I32 naughty; /* How bad is this pattern? */
141 I32 sawback; /* Did we see \1, ...? */
142 U32 seen;
88f063b4
KW
143 SSize_t size; /* Number of regnode equivalents in
144 pattern */
7c932d07
KW
145
146 /* position beyond 'precomp' of the warning message furthest away from
147 * 'precomp'. During the parse, no warnings are raised for any problems
148 * earlier in the parse than this position. This works if warnings are
149 * raised the first time a given spot is parsed, and if only one
150 * independent warning is raised for any given spot */
151 Size_t latest_warn_offset;
152
153 I32 npar; /* Capture buffer count so far in the
154 parse, (OPEN) plus one. ("par" 0 is
155 the whole pattern)*/
156 I32 total_par; /* During initial parse, is either 0,
157 or -1; the latter indicating a
158 reparse is needed. After that pass,
159 it is what 'npar' became after the
160 pass. Hence, it being > 0 indicates
161 we are in a reparse situation */
538e84ed
KW
162 I32 nestroot; /* root parens we are in - used by
163 accept */
830247a4 164 I32 seen_zerolen;
f55b7b26
KW
165 regnode_offset *open_parens; /* offsets to open parens */
166 regnode_offset *close_parens; /* offsets to close parens */
0275405c 167 I32 parens_buf_size; /* #slots malloced open/close_parens */
d5a00e4a 168 regnode *end_op; /* END node in program */
02daf0ab
YO
169 I32 utf8; /* whether the pattern is utf8 or not */
170 I32 orig_utf8; /* whether the pattern was originally in utf8 */
171 /* XXX use this for future optimisation of case
172 * where pattern must be upgraded to utf8. */
e40e74fe
KW
173 I32 uni_semantics; /* If a d charset modifier should use unicode
174 rules, even if the pattern is not in
175 utf8 */
81714fb9 176 HV *paren_names; /* Paren names */
538e84ed 177
40d049e4 178 regnode **recurse; /* Recurse regops */
88f063b4 179 I32 recurse_count; /* Number of recurse regops we have generated */
4286711a 180 U8 *study_chunk_recursed; /* bitmap of which subs we have moved
538e84ed 181 through */
09a65838 182 U32 study_chunk_recursed_bytes; /* bytes in bitmap */
b57e4118 183 I32 in_lookbehind;
105c827d 184 I32 in_lookahead;
4624b182 185 I32 contains_locale;
bb3f3ed2 186 I32 override_recoding;
4caef9c3 187 I32 recode_x_to_native;
9d53c457 188 I32 in_multi_char_class;
1acab4c5 189 struct reg_code_blocks *code_blocks;/* positions of literal (?{})
68e2671b 190 within pattern */
b1603ef8 191 int code_index; /* next code_blocks[] slot */
ee273784 192 SSize_t maxlen; /* mininum possible number of chars in string to match */
3f910e62
YO
193 scan_frame *frame_head;
194 scan_frame *frame_last;
195 U32 frame_count;
7eec73eb 196 AV *warn_text;
945961fd 197 HV *unlexed_names;
dc6d7f5c 198#ifdef ADD_TO_REGEXEC
830247a4
IZ
199 char *starttry; /* -Dr: where regtry was called. */
200#define RExC_starttry (pRExC_state->starttry)
201#endif
d24ca0c5 202 SV *runtime_code_qr; /* qr with the runtime code blocks */
3dab1dad 203#ifdef DEBUGGING
be8e71aa 204 const char *lastparse;
3dab1dad 205 I32 lastnum;
1f1031fe 206 AV *paren_name_list; /* idx -> name */
d9a72fcc 207 U32 study_chunk_recursed_count;
c9f0d54c
YO
208 SV *mysv1;
209 SV *mysv2;
88f063b4 210
3dab1dad
YO
211#define RExC_lastparse (pRExC_state->lastparse)
212#define RExC_lastnum (pRExC_state->lastnum)
1f1031fe 213#define RExC_paren_name_list (pRExC_state->paren_name_list)
d9a72fcc 214#define RExC_study_chunk_recursed_count (pRExC_state->study_chunk_recursed_count)
c9f0d54c
YO
215#define RExC_mysv (pRExC_state->mysv1)
216#define RExC_mysv1 (pRExC_state->mysv1)
217#define RExC_mysv2 (pRExC_state->mysv2)
218
3dab1dad 219#endif
bd0b62a6 220 bool seen_d_op;
911bd04e 221 bool strict;
da7cf1cc 222 bool study_started;
034602eb 223 bool in_script_run;
7c932d07 224 bool use_BRANCHJ;
09b2b2e6 225};
830247a4 226
e2509266 227#define RExC_flags (pRExC_state->flags)
514a91f1 228#define RExC_pm_flags (pRExC_state->pm_flags)
830247a4 229#define RExC_precomp (pRExC_state->precomp)
51684e3c
KW
230#define RExC_copy_start_in_input (pRExC_state->copy_start_in_input)
231#define RExC_copy_start_in_constructed (pRExC_state->copy_start)
cc16d262 232#define RExC_save_copy_start_in_constructed (pRExC_state->save_copy_start)
711b303b 233#define RExC_precomp_end (pRExC_state->precomp_end)
288b8c02 234#define RExC_rx_sv (pRExC_state->rx_sv)
830247a4 235#define RExC_rx (pRExC_state->rx)
f8fc2ecf 236#define RExC_rxi (pRExC_state->rxi)
fac92740 237#define RExC_start (pRExC_state->start)
830247a4
IZ
238#define RExC_end (pRExC_state->end)
239#define RExC_parse (pRExC_state->parse)
7c932d07 240#define RExC_latest_warn_offset (pRExC_state->latest_warn_offset )
830247a4 241#define RExC_whilem_seen (pRExC_state->whilem_seen)
bd0b62a6
KW
242#define RExC_seen_d_op (pRExC_state->seen_d_op) /* Seen something that differs
243 under /d from /u ? */
512c0f5a 244
7122b237 245#ifdef RE_TRACK_PATTERN_OFFSETS
44f36cd3 246# define RExC_offsets (RExC_rxi->u.offsets) /* I am not like the
538e84ed 247 others */
7122b237 248#endif
830247a4 249#define RExC_emit (pRExC_state->emit)
fac92740 250#define RExC_emit_start (pRExC_state->emit_start)
830247a4
IZ
251#define RExC_sawback (pRExC_state->sawback)
252#define RExC_seen (pRExC_state->seen)
253#define RExC_size (pRExC_state->size)
ee273784 254#define RExC_maxlen (pRExC_state->maxlen)
830247a4 255#define RExC_npar (pRExC_state->npar)
1533d1d2 256#define RExC_total_parens (pRExC_state->total_par)
0275405c 257#define RExC_parens_buf_size (pRExC_state->parens_buf_size)
e2e6a0f1 258#define RExC_nestroot (pRExC_state->nestroot)
830247a4 259#define RExC_seen_zerolen (pRExC_state->seen_zerolen)
1aa99e6b 260#define RExC_utf8 (pRExC_state->utf8)
e40e74fe 261#define RExC_uni_semantics (pRExC_state->uni_semantics)
02daf0ab 262#define RExC_orig_utf8 (pRExC_state->orig_utf8)
40d049e4
YO
263#define RExC_open_parens (pRExC_state->open_parens)
264#define RExC_close_parens (pRExC_state->close_parens)
5bd2d46e 265#define RExC_end_op (pRExC_state->end_op)
81714fb9 266#define RExC_paren_names (pRExC_state->paren_names)
40d049e4
YO
267#define RExC_recurse (pRExC_state->recurse)
268#define RExC_recurse_count (pRExC_state->recurse_count)
09a65838 269#define RExC_study_chunk_recursed (pRExC_state->study_chunk_recursed)
538e84ed
KW
270#define RExC_study_chunk_recursed_bytes \
271 (pRExC_state->study_chunk_recursed_bytes)
b57e4118 272#define RExC_in_lookbehind (pRExC_state->in_lookbehind)
105c827d 273#define RExC_in_lookahead (pRExC_state->in_lookahead)
4624b182 274#define RExC_contains_locale (pRExC_state->contains_locale)
4caef9c3
KW
275#define RExC_recode_x_to_native (pRExC_state->recode_x_to_native)
276
b6d67071 277#ifdef EBCDIC
4caef9c3
KW
278# define SET_recode_x_to_native(x) \
279 STMT_START { RExC_recode_x_to_native = (x); } STMT_END
280#else
281# define SET_recode_x_to_native(x) NOOP
b6d67071 282#endif
4caef9c3 283
9d53c457 284#define RExC_in_multi_char_class (pRExC_state->in_multi_char_class)
3f910e62
YO
285#define RExC_frame_head (pRExC_state->frame_head)
286#define RExC_frame_last (pRExC_state->frame_last)
287#define RExC_frame_count (pRExC_state->frame_count)
67cdf558 288#define RExC_strict (pRExC_state->strict)
da7cf1cc 289#define RExC_study_started (pRExC_state->study_started)
7eec73eb 290#define RExC_warn_text (pRExC_state->warn_text)
034602eb 291#define RExC_in_script_run (pRExC_state->in_script_run)
7c932d07 292#define RExC_use_BRANCHJ (pRExC_state->use_BRANCHJ)
945961fd 293#define RExC_unlexed_names (pRExC_state->unlexed_names)
830247a4 294
99807a43
HS
295/* Heuristic check on the complexity of the pattern: if TOO_NAUGHTY, we set
296 * a flag to disable back-off on the fixed/floating substrings - if it's
297 * a high complexity pattern we assume the benefit of avoiding a full match
298 * is worth the cost of checking for the substrings even if they rarely help.
299 */
300#define RExC_naughty (pRExC_state->naughty)
301#define TOO_NAUGHTY (10)
302#define MARK_NAUGHTY(add) \
303 if (RExC_naughty < TOO_NAUGHTY) \
304 RExC_naughty += (add)
305#define MARK_NAUGHTY_EXP(exp, add) \
306 if (RExC_naughty < TOO_NAUGHTY) \
307 RExC_naughty += RExC_naughty / (exp) + (add)
cde0cee5 308
a687059c
LW
309#define ISMULT1(c) ((c) == '*' || (c) == '+' || (c) == '?')
310#define ISMULT2(s) ((*s) == '*' || (*s) == '+' || (*s) == '?' || \
412f55bb 311 ((*s) == '{' && regcurly(s)))
a687059c
LW
312
313/*
314 * Flags to be passed up and down.
315 */
a687059c 316#define WORST 0 /* Worst case. */
c4ead770
KW
317#define HASWIDTH 0x01 /* Known to not match null strings, could match
318 non-null ones. */
fda99bee 319
e64f369d 320/* Simple enough to be STAR/PLUS operand; in an EXACTish node must be a single
2fd92675
KW
321 * character. (There needs to be a case: in the switch statement in regexec.c
322 * for any node marked SIMPLE.) Note that this is not the same thing as
323 * REGNODE_SIMPLE */
fda99bee 324#define SIMPLE 0x02
e64f369d 325#define SPSTART 0x04 /* Starts with * or + */
8d9c2815
NC
326#define POSTPONED 0x08 /* (?1),(?&name), (??{...}) or similar */
327#define TRYAGAIN 0x10 /* Weeded out a declaration. */
c36361b2
KW
328#define RESTART_PARSE 0x20 /* Need to redo the parse */
329#define NEED_UTF8 0x40 /* In conjunction with RESTART_PARSE, need to
330 calcuate sizes as UTF-8 */
a687059c 331
3dab1dad
YO
332#define REG_NODE_NUM(x) ((x) ? (int)((x)-RExC_emit_start) : -1)
333
07be1b83
YO
334/* whether trie related optimizations are enabled */
335#if PERL_ENABLE_EXTENDED_TRIE_OPTIMISATION
336#define TRIE_STUDY_OPT
786e8c11 337#define FULL_TRIE_STUDY
07be1b83
YO
338#define TRIE_STCLASS
339#endif
1de06328
YO
340
341
40d049e4
YO
342
343#define PBYTE(u8str,paren) ((U8*)(u8str))[(paren) >> 3]
344#define PBITVAL(paren) (1 << ((paren) & 7))
ac3ef366
HS
345#define PAREN_OFFSET(depth) \
346 (RExC_study_chunk_recursed + (depth) * RExC_study_chunk_recursed_bytes)
347#define PAREN_TEST(depth, paren) \
348 (PBYTE(PAREN_OFFSET(depth), paren) & PBITVAL(paren))
349#define PAREN_SET(depth, paren) \
350 (PBYTE(PAREN_OFFSET(depth), paren) |= PBITVAL(paren))
351#define PAREN_UNSET(depth, paren) \
352 (PBYTE(PAREN_OFFSET(depth), paren) &= ~PBITVAL(paren))
40d049e4 353
82a6ada4 354#define REQUIRE_UTF8(flagp) STMT_START { \
8d9c2815 355 if (!UTF) { \
c36361b2 356 *flagp = RESTART_PARSE|NEED_UTF8; \
f55b7b26 357 return 0; \
8d9c2815 358 } \
82a6ada4 359 } STMT_END
40d049e4 360
d63d75f5 361/* Change from /d into /u rules, and restart the parse. RExC_uni_semantics is
f6e8f31e
KW
362 * a flag that indicates we need to override /d with /u as a result of
363 * something in the pattern. It should only be used in regards to calling
b2090782 364 * set_regex_charset() or get_regex_charset() */
512c0f5a
KW
365#define REQUIRE_UNI_RULES(flagp, restart_retval) \
366 STMT_START { \
367 if (DEPENDS_SEMANTICS) { \
512c0f5a
KW
368 set_regex_charset(&RExC_flags, REGEX_UNICODE_CHARSET); \
369 RExC_uni_semantics = 1; \
b76e1c04 370 if (RExC_seen_d_op && LIKELY(! IN_PARENS_PASS)) { \
bd0b62a6
KW
371 /* No need to restart the parse if we haven't seen \
372 * anything that differs between /u and /d, and no need \
373 * to restart immediately if we're going to reparse \
374 * anyway to count parens */ \
c36361b2 375 *flagp |= RESTART_PARSE; \
512c0f5a 376 return restart_retval; \
d1be211c 377 } \
512c0f5a
KW
378 } \
379 } STMT_END
380
7c932d07
KW
381#define REQUIRE_BRANCHJ(flagp, restart_retval) \
382 STMT_START { \
383 RExC_use_BRANCHJ = 1; \
439a3bfe
KW
384 *flagp |= RESTART_PARSE; \
385 return restart_retval; \
7c932d07
KW
386 } STMT_END
387
b76e1c04
KW
388/* Until we have completed the parse, we leave RExC_total_parens at 0 or
389 * less. After that, it must always be positive, because the whole re is
390 * considered to be surrounded by virtual parens. Setting it to negative
391 * indicates there is some construct that needs to know the actual number of
392 * parens to be properly handled. And that means an extra pass will be
393 * required after we've counted them all */
394#define ALL_PARENS_COUNTED (RExC_total_parens > 0)
7c932d07 395#define REQUIRE_PARENS_PASS \
b76e1c04
KW
396 STMT_START { /* No-op if have completed a pass */ \
397 if (! ALL_PARENS_COUNTED) RExC_total_parens = -1; \
7c932d07 398 } STMT_END
b76e1c04
KW
399#define IN_PARENS_PASS (RExC_total_parens < 0)
400
7c932d07 401
f358228e
KW
402/* This is used to return failure (zero) early from the calling function if
403 * various flags in 'flags' are set. Two flags always cause a return:
404 * 'RESTART_PARSE' and 'NEED_UTF8'. 'extra' can be used to specify any
405 * additional flags that should cause a return; 0 if none. If the return will
406 * be done, '*flagp' is first set to be all of the flags that caused the
407 * return. */
408#define RETURN_FAIL_ON_RESTART_OR_FLAGS(flags,flagp,extra) \
fd7fc6ae 409 STMT_START { \
c36361b2
KW
410 if ((flags) & (RESTART_PARSE|NEED_UTF8|(extra))) { \
411 *(flagp) = (flags) & (RESTART_PARSE|NEED_UTF8|(extra)); \
f358228e 412 return 0; \
fd7fc6ae
YO
413 } \
414 } STMT_END
415
c36361b2 416#define MUST_RESTART(flags) ((flags) & (RESTART_PARSE))
fd7fc6ae 417
5d555bb8 418#define RETURN_FAIL_ON_RESTART(flags,flagp) \
f358228e 419 RETURN_FAIL_ON_RESTART_OR_FLAGS( flags, flagp, 0)
5d555bb8 420#define RETURN_FAIL_ON_RESTART_FLAGP(flagp) \
f358228e 421 if (MUST_RESTART(*(flagp))) return 0
fd7fc6ae 422
f19b1a63
KW
423/* This converts the named class defined in regcomp.h to its equivalent class
424 * number defined in handy.h. */
425#define namedclass_to_classnum(class) ((int) ((class) / 2))
426#define classnum_to_namedclass(classnum) ((classnum) * 2)
427
de92f5e6
KW
428#define _invlist_union_complement_2nd(a, b, output) \
429 _invlist_union_maybe_complement_2nd(a, b, TRUE, output)
430#define _invlist_intersection_complement_2nd(a, b, output) \
431 _invlist_intersection_maybe_complement_2nd(a, b, TRUE, output)
432
4a808c87
KW
433/* We add a marker if we are deferring expansion of a property that is both
434 * 1) potentiallly user-defined; and
435 * 2) could also be an official Unicode property.
436 *
437 * Without this marker, any deferred expansion can only be for a user-defined
438 * one. This marker shouldn't conflict with any that could be in a legal name,
439 * and is appended to its name to indicate this. There is a string and
440 * character form */
043cdb9a
KW
441#define DEFERRED_COULD_BE_OFFICIAL_MARKERs "~"
442#define DEFERRED_COULD_BE_OFFICIAL_MARKERc '~'
1c2f3d7a 443
1de06328
YO
444/* About scan_data_t.
445
446 During optimisation we recurse through the regexp program performing
447 various inplace (keyhole style) optimisations. In addition study_chunk
448 and scan_commit populate this data structure with information about
538e84ed 449 what strings MUST appear in the pattern. We look for the longest
3b753521 450 string that must appear at a fixed location, and we look for the
1de06328
YO
451 longest string that may appear at a floating location. So for instance
452 in the pattern:
538e84ed 453
1de06328 454 /FOO[xX]A.*B[xX]BAR/
538e84ed 455
1de06328
YO
456 Both 'FOO' and 'A' are fixed strings. Both 'B' and 'BAR' are floating
457 strings (because they follow a .* construct). study_chunk will identify
458 both FOO and BAR as being the longest fixed and floating strings respectively.
538e84ed 459
1de06328 460 The strings can be composites, for instance
538e84ed 461
1de06328 462 /(f)(o)(o)/
538e84ed 463
1de06328 464 will result in a composite fixed substring 'foo'.
538e84ed 465
1de06328 466 For each string some basic information is maintained:
538e84ed 467
6ea77169 468 - min_offset
1de06328
YO
469 This is the position the string must appear at, or not before.
470 It also implicitly (when combined with minlenp) tells us how many
3b753521
FN
471 characters must match before the string we are searching for.
472 Likewise when combined with minlenp and the length of the string it
538e84ed 473 tells us how many characters must appear after the string we have
1de06328 474 found.
538e84ed 475
1de06328
YO
476 - max_offset
477 Only used for floating strings. This is the rightmost point that
ea3daa5d 478 the string can appear at. If set to SSize_t_MAX it indicates that the
1de06328 479 string can occur infinitely far to the right.
6e12f832 480 For fixed strings, it is equal to min_offset.
538e84ed 481
1de06328 482 - minlenp
2d608413
KW
483 A pointer to the minimum number of characters of the pattern that the
484 string was found inside. This is important as in the case of positive
538e84ed 485 lookahead or positive lookbehind we can have multiple patterns
1de06328 486 involved. Consider
538e84ed 487
1de06328 488 /(?=FOO).*F/
538e84ed 489
1de06328
YO
490 The minimum length of the pattern overall is 3, the minimum length
491 of the lookahead part is 3, but the minimum length of the part that
538e84ed 492 will actually match is 1. So 'FOO's minimum length is 3, but the
1de06328 493 minimum length for the F is 1. This is important as the minimum length
538e84ed 494 is used to determine offsets in front of and behind the string being
1de06328 495 looked for. Since strings can be composites this is the length of the
486ec47a 496 pattern at the time it was committed with a scan_commit. Note that
1de06328 497 the length is calculated by study_chunk, so that the minimum lengths
538e84ed 498 are not known until the full pattern has been compiled, thus the
1de06328 499 pointer to the value.
538e84ed 500
1de06328 501 - lookbehind
538e84ed 502
1de06328 503 In the case of lookbehind the string being searched for can be
538e84ed 504 offset past the start point of the final matching string.
1de06328
YO
505 If this value was just blithely removed from the min_offset it would
506 invalidate some of the calculations for how many chars must match
507 before or after (as they are derived from min_offset and minlen and
538e84ed 508 the length of the string being searched for).
1de06328
YO
509 When the final pattern is compiled and the data is moved from the
510 scan_data_t structure into the regexp structure the information
538e84ed
KW
511 about lookbehind is factored in, with the information that would
512 have been lost precalculated in the end_shift field for the
1de06328
YO
513 associated string.
514
515 The fields pos_min and pos_delta are used to store the minimum offset
538e84ed 516 and the delta to the maximum offset at the current point in the pattern.
1de06328
YO
517
518*/
2c2d71f5 519
cb6aef8b
DM
520struct scan_data_substrs {
521 SV *str; /* longest substring found in pattern */
522 SSize_t min_offset; /* earliest point in string it can appear */
523 SSize_t max_offset; /* latest point in string it can appear */
524 SSize_t *minlenp; /* pointer to the minlen relevant to the string */
525 SSize_t lookbehind; /* is the pos of the string modified by LB */
11683ecb 526 I32 flags; /* per substring SF_* and SCF_* flags */
cb6aef8b
DM
527};
528
2c2d71f5 529typedef struct scan_data_t {
1de06328
YO
530 /*I32 len_min; unused */
531 /*I32 len_delta; unused */
49f55535 532 SSize_t pos_min;
ea3daa5d 533 SSize_t pos_delta;
2c2d71f5 534 SV *last_found;
ea3daa5d 535 SSize_t last_end; /* min value, <0 unless valid. */
49f55535 536 SSize_t last_start_min;
ea3daa5d 537 SSize_t last_start_max;
2099df82
DM
538 U8 cur_is_floating; /* whether the last_* values should be set as
539 * the next fixed (0) or floating (1)
540 * substring */
6ea77169 541
2099df82
DM
542 /* [0] is longest fixed substring so far, [1] is longest float so far */
543 struct scan_data_substrs substrs[2];
6ea77169 544
11683ecb 545 I32 flags; /* common SF_* and SCF_* flags */
2c2d71f5 546 I32 whilem_c;
49f55535 547 SSize_t *last_closep;
b8f7bb16 548 regnode_ssc *start_class;
2c2d71f5
JH
549} scan_data_t;
550
a687059c 551/*
e50aee73 552 * Forward declarations for pregcomp()'s friends.
a687059c 553 */
a0d0e21e 554
6ea77169 555static const scan_data_t zero_scan_data = {
a2ca2017 556 0, 0, NULL, 0, 0, 0, 0,
6ea77169 557 {
11683ecb
DM
558 { NULL, 0, 0, 0, 0, 0 },
559 { NULL, 0, 0, 0, 0, 0 },
6ea77169
DM
560 },
561 0, 0, NULL, NULL
562};
c277df42 563
11683ecb
DM
564/* study flags */
565
07be1b83
YO
566#define SF_BEFORE_SEOL 0x0001
567#define SF_BEFORE_MEOL 0x0002
11683ecb 568#define SF_BEFORE_EOL (SF_BEFORE_SEOL|SF_BEFORE_MEOL)
c277df42 569
07be1b83
YO
570#define SF_IS_INF 0x0040
571#define SF_HAS_PAR 0x0080
572#define SF_IN_PAR 0x0100
573#define SF_HAS_EVAL 0x0200
bc604ad8
KW
574
575
576/* SCF_DO_SUBSTR is the flag that tells the regexp analyzer to track the
577 * longest substring in the pattern. When it is not set the optimiser keeps
578 * track of position, but does not keep track of the actual strings seen,
579 *
580 * So for instance /foo/ will be parsed with SCF_DO_SUBSTR being true, but
581 * /foo/i will not.
582 *
583 * Similarly, /foo.*(blah|erm|huh).*fnorble/ will have "foo" and "fnorble"
584 * parsed with SCF_DO_SUBSTR on, but while processing the (...) it will be
585 * turned off because of the alternation (BRANCH). */
07be1b83 586#define SCF_DO_SUBSTR 0x0400
bc604ad8 587
653099ff
GS
588#define SCF_DO_STCLASS_AND 0x0800
589#define SCF_DO_STCLASS_OR 0x1000
590#define SCF_DO_STCLASS (SCF_DO_STCLASS_AND|SCF_DO_STCLASS_OR)
e1901655 591#define SCF_WHILEM_VISITED_POS 0x2000
c277df42 592
786e8c11 593#define SCF_TRIE_RESTUDY 0x4000 /* Do restudy? */
538e84ed 594#define SCF_SEEN_ACCEPT 0x8000
688e0391 595#define SCF_TRIE_DOING_RESTUDY 0x10000
4286711a
YO
596#define SCF_IN_DEFINE 0x20000
597
598
599
07be1b83 600
43fead97 601#define UTF cBOOL(RExC_utf8)
00b27cfc
KW
602
603/* The enums for all these are ordered so things work out correctly */
a62b1201 604#define LOC (get_regex_charset(RExC_flags) == REGEX_LOCALE_CHARSET)
538e84ed
KW
605#define DEPENDS_SEMANTICS (get_regex_charset(RExC_flags) \
606 == REGEX_DEPENDS_CHARSET)
00b27cfc 607#define UNI_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_UNICODE_CHARSET)
538e84ed
KW
608#define AT_LEAST_UNI_SEMANTICS (get_regex_charset(RExC_flags) \
609 >= REGEX_UNICODE_CHARSET)
610#define ASCII_RESTRICTED (get_regex_charset(RExC_flags) \
611 == REGEX_ASCII_RESTRICTED_CHARSET)
612#define AT_LEAST_ASCII_RESTRICTED (get_regex_charset(RExC_flags) \
613 >= REGEX_ASCII_RESTRICTED_CHARSET)
614#define ASCII_FOLD_RESTRICTED (get_regex_charset(RExC_flags) \
615 == REGEX_ASCII_MORE_RESTRICTED_CHARSET)
a62b1201 616
43fead97 617#define FOLD cBOOL(RExC_flags & RXf_PMf_FOLD)
a0ed51b3 618
f2c2a6ab
KW
619/* For programs that want to be strictly Unicode compatible by dying if any
620 * attempt is made to match a non-Unicode code point against a Unicode
621 * property. */
622#define ALWAYS_WARN_SUPER ckDEAD(packWARN(WARN_NON_UNICODE))
623
93733859 624#define OOB_NAMEDCLASS -1
b8c5462f 625
8e661ac5
KW
626/* There is no code point that is out-of-bounds, so this is problematic. But
627 * its only current use is to initialize a variable that is always set before
628 * looked at. */
629#define OOB_UNICODE 0xDEADBEEF
630
a0ed51b3 631#define CHR_SVLEN(sv) (UTF ? sv_len_utf8(sv) : SvCUR(sv))
a0ed51b3 632
8615cb43 633
b45f050a
JF
634/* length of regex to show in messages that don't mark a position within */
635#define RegexLengthToShowInErrorMessages 127
636
637/*
638 * If MARKER[12] are adjusted, be sure to adjust the constants at the top
639 * of t/op/regmesg.t, the tests in t/op/re_tests, and those in
640 * op/pragma/warn/regcomp.
641 */
7253e4e3
RK
642#define MARKER1 "<-- HERE" /* marker as it appears in the description */
643#define MARKER2 " <-- HERE " /* marker as it appears within the regex */
b81d288d 644
538e84ed 645#define REPORT_LOCATION " in regex; marked by " MARKER1 \
147e3846 646 " in m/%" UTF8f MARKER2 "%" UTF8f "/"
b45f050a 647
285b5ca0
KW
648/* The code in this file in places uses one level of recursion with parsing
649 * rebased to an alternate string constructed by us in memory. This can take
650 * the form of something that is completely different from the input, or
651 * something that uses the input as part of the alternate. In the first case,
652 * there should be no possibility of an error, as we are in complete control of
51684e3c
KW
653 * the alternate string. But in the second case we don't completely control
654 * the input portion, so there may be errors in that. Here's an example:
285b5ca0
KW
655 * /[abc\x{DF}def]/ui
656 * is handled specially because \x{df} folds to a sequence of more than one
51684e3c 657 * character: 'ss'. What is done is to create and parse an alternate string,
285b5ca0
KW
658 * which looks like this:
659 * /(?:\x{DF}|[abc\x{DF}def])/ui
660 * where it uses the input unchanged in the middle of something it constructs,
661 * which is a branch for the DF outside the character class, and clustering
662 * parens around the whole thing. (It knows enough to skip the DF inside the
663 * class while in this substitute parse.) 'abc' and 'def' may have errors that
664 * need to be reported. The general situation looks like this:
665 *
51684e3c 666 * |<------- identical ------>|
285b5ca0 667 * sI tI xI eI
51684e3c 668 * Input: ---------------------------------------------------------------
285b5ca0
KW
669 * Constructed: ---------------------------------------------------
670 * sC tC xC eC EC
51684e3c 671 * |<------- identical ------>|
285b5ca0 672 *
51684e3c
KW
673 * sI..eI is the portion of the input pattern we are concerned with here.
674 * sC..EC is the constructed substitute parse string.
675 * sC..tC is constructed by us
676 * tC..eC is an exact duplicate of the portion of the input pattern tI..eI.
677 * In the diagram, these are vertically aligned.
678 * eC..EC is also constructed by us.
679 * xC is the position in the substitute parse string where we found a
680 * problem.
681 * xI is the position in the original pattern corresponding to xC.
285b5ca0 682 *
51684e3c
KW
683 * We want to display a message showing the real input string. Thus we need to
684 * translate from xC to xI. We know that xC >= tC, since the portion of the
685 * string sC..tC has been constructed by us, and so shouldn't have errors. We
686 * get:
687 * xI = tI + (xC - tC)
285b5ca0 688 *
51684e3c
KW
689 * When the substitute parse is constructed, the code needs to set:
690 * RExC_start (sC)
691 * RExC_end (eC)
692 * RExC_copy_start_in_input (tI)
693 * RExC_copy_start_in_constructed (tC)
694 * and restore them when done.
285b5ca0 695 *
51684e3c
KW
696 * During normal processing of the input pattern, both
697 * 'RExC_copy_start_in_input' and 'RExC_copy_start_in_constructed' are set to
698 * sI, so that xC equals xI.
285b5ca0
KW
699 */
700
51684e3c
KW
701#define sI RExC_precomp
702#define eI RExC_precomp_end
703#define sC RExC_start
704#define eC RExC_end
705#define tI RExC_copy_start_in_input
706#define tC RExC_copy_start_in_constructed
707#define xI(xC) (tI + (xC - tC))
708#define xI_offset(xC) (xI(xC) - sI)
285b5ca0
KW
709
710#define REPORT_LOCATION_ARGS(xC) \
711 UTF8fARG(UTF, \
51684e3c 712 (xI(xC) > eI) /* Don't run off end */ \
232b691f 713 ? eI - sI /* Length before the <--HERE */ \
bc929b61
KW
714 : ((xI_offset(xC) >= 0) \
715 ? xI_offset(xC) \
716 : (Perl_croak(aTHX_ "panic: %s: %d: negative offset: %" \
717 IVdf " trying to output message for " \
718 " pattern %.*s", \
51684e3c 719 __FILE__, __LINE__, (IV) xI_offset(xC), \
bc929b61 720 ((int) (eC - sC)), sC), 0)), \
51684e3c 721 sI), /* The input pattern printed up to the <--HERE */ \
285b5ca0 722 UTF8fARG(UTF, \
51684e3c
KW
723 (xI(xC) > eI) ? 0 : eI - xI(xC), /* Length after <--HERE */ \
724 (xI(xC) > eI) ? eI : xI(xC)) /* pattern after <--HERE */
c1d900c3 725
8a6d8ec6
HS
726/* Used to point after bad bytes for an error message, but avoid skipping
727 * past a nul byte. */
e3304382 728#define SKIP_IF_CHAR(s, e) (!*(s) ? 0 : UTF ? UTF8_SAFE_SKIP(s, e) : 1)
8a6d8ec6 729
ed68fe11
KW
730/* Set up to clean up after our imminent demise */
731#define PREPARE_TO_DIE \
732 STMT_START { \
733 if (RExC_rx_sv) \
734 SAVEFREESV(RExC_rx_sv); \
7c932d07
KW
735 if (RExC_open_parens) \
736 SAVEFREEPV(RExC_open_parens); \
737 if (RExC_close_parens) \
738 SAVEFREEPV(RExC_close_parens); \
ed68fe11
KW
739 } STMT_END
740
b45f050a
JF
741/*
742 * Calls SAVEDESTRUCTOR_X if needed, then calls Perl_croak with the given
743 * arg. Show regex, up to a maximum length. If it's too long, chop and add
744 * "...".
745 */
58e23c8d 746#define _FAIL(code) STMT_START { \
bfed75c6 747 const char *ellipses = ""; \
88f063b4 748 IV len = RExC_precomp_end - RExC_precomp; \
ccb2c380 749 \
ed68fe11 750 PREPARE_TO_DIE; \
ccb2c380
MP
751 if (len > RegexLengthToShowInErrorMessages) { \
752 /* chop 10 shorter than the max, to ensure meaning of "..." */ \
753 len = RegexLengthToShowInErrorMessages - 10; \
754 ellipses = "..."; \
755 } \
58e23c8d 756 code; \
ccb2c380 757} STMT_END
8615cb43 758
58e23c8d 759#define FAIL(msg) _FAIL( \
147e3846 760 Perl_croak(aTHX_ "%s in regex m/%" UTF8f "%s/", \
c1d900c3 761 msg, UTF8fARG(UTF, len, RExC_precomp), ellipses))
58e23c8d
YO
762
763#define FAIL2(msg,arg) _FAIL( \
147e3846 764 Perl_croak(aTHX_ msg " in regex m/%" UTF8f "%s/", \
c1d900c3 765 arg, UTF8fARG(UTF, len, RExC_precomp), ellipses))
58e23c8d 766
4b7718d6
KW
767#define FAIL3(msg,arg1,arg2) _FAIL( \
768 Perl_croak(aTHX_ msg " in regex m/%" UTF8f "%s/", \
769 arg1, arg2, UTF8fARG(UTF, len, RExC_precomp), ellipses))
770
b45f050a 771/*
b45f050a
JF
772 * Simple_vFAIL -- like FAIL, but marks the current location in the scan
773 */
ccb2c380 774#define Simple_vFAIL(m) STMT_START { \
ccb2c380 775 Perl_croak(aTHX_ "%s" REPORT_LOCATION, \
d528642a 776 m, REPORT_LOCATION_ARGS(RExC_parse)); \
ccb2c380 777} STMT_END
b45f050a
JF
778
779/*
780 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL()
781 */
ccb2c380 782#define vFAIL(m) STMT_START { \
ed68fe11 783 PREPARE_TO_DIE; \
ccb2c380
MP
784 Simple_vFAIL(m); \
785} STMT_END
b45f050a
JF
786
787/*
788 * Like Simple_vFAIL(), but accepts two arguments.
789 */
ccb2c380 790#define Simple_vFAIL2(m,a1) STMT_START { \
d528642a
KW
791 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, \
792 REPORT_LOCATION_ARGS(RExC_parse)); \
ccb2c380 793} STMT_END
b45f050a
JF
794
795/*
796 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL2().
797 */
ccb2c380 798#define vFAIL2(m,a1) STMT_START { \
ed68fe11 799 PREPARE_TO_DIE; \
ccb2c380
MP
800 Simple_vFAIL2(m, a1); \
801} STMT_END
b45f050a
JF
802
803
804/*
805 * Like Simple_vFAIL(), but accepts three arguments.
806 */
ccb2c380 807#define Simple_vFAIL3(m, a1, a2) STMT_START { \
c1d900c3 808 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, \
d528642a 809 REPORT_LOCATION_ARGS(RExC_parse)); \
ccb2c380 810} STMT_END
b45f050a
JF
811
812/*
813 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL3().
814 */
ccb2c380 815#define vFAIL3(m,a1,a2) STMT_START { \
ed68fe11 816 PREPARE_TO_DIE; \
ccb2c380
MP
817 Simple_vFAIL3(m, a1, a2); \
818} STMT_END
b45f050a
JF
819
820/*
821 * Like Simple_vFAIL(), but accepts four arguments.
822 */
ccb2c380 823#define Simple_vFAIL4(m, a1, a2, a3) STMT_START { \
d528642a
KW
824 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, a3, \
825 REPORT_LOCATION_ARGS(RExC_parse)); \
ccb2c380 826} STMT_END
b45f050a 827
95db3ffa 828#define vFAIL4(m,a1,a2,a3) STMT_START { \
ed68fe11 829 PREPARE_TO_DIE; \
95db3ffa
KW
830 Simple_vFAIL4(m, a1, a2, a3); \
831} STMT_END
832
946095af 833/* A specialized version of vFAIL2 that works with UTF8f */
d528642a 834#define vFAIL2utf8f(m, a1) STMT_START { \
ed68fe11 835 PREPARE_TO_DIE; \
d528642a
KW
836 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, \
837 REPORT_LOCATION_ARGS(RExC_parse)); \
946095af
BF
838} STMT_END
839
3ba22297 840#define vFAIL3utf8f(m, a1, a2) STMT_START { \
ed68fe11 841 PREPARE_TO_DIE; \
3ba22297 842 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, \
d528642a 843 REPORT_LOCATION_ARGS(RExC_parse)); \
3ba22297
KW
844} STMT_END
845
d62e749e 846/* Setting this to NULL is a signal to not output warnings */
cc16d262
KW
847#define TURN_OFF_WARNINGS_IN_SUBSTITUTE_PARSE \
848 STMT_START { \
849 RExC_save_copy_start_in_constructed = RExC_copy_start_in_constructed;\
850 RExC_copy_start_in_constructed = NULL; \
851 } STMT_END
852#define RESTORE_WARNINGS \
853 RExC_copy_start_in_constructed = RExC_save_copy_start_in_constructed
d62e749e 854
7c932d07
KW
855/* Since a warning can be generated multiple times as the input is reparsed, we
856 * output it the first time we come to that point in the parse, but suppress it
857 * otherwise. 'RExC_copy_start_in_constructed' being NULL is a flag to not
858 * generate any warnings */
55bcf857 859#define TO_OUTPUT_WARNINGS(loc) \
7c932d07
KW
860 ( RExC_copy_start_in_constructed \
861 && ((xI(loc)) - RExC_precomp) > (Ptrdiff_t) RExC_latest_warn_offset)
55bcf857 862
7c932d07
KW
863/* After we've emitted a warning, we save the position in the input so we don't
864 * output it again */
865#define UPDATE_WARNINGS_LOC(loc) \
866 STMT_START { \
867 if (TO_OUTPUT_WARNINGS(loc)) { \
5d894ca5
KW
868 RExC_latest_warn_offset = MAX(sI, MIN(eI, xI(loc))) \
869 - RExC_precomp; \
7c932d07
KW
870 } \
871 } STMT_END
77ccba3c 872
086f21bf 873/* 'warns' is the output of the packWARNx macro used in 'code' */
f8220fe2
KW
874#define _WARN_HELPER(loc, warns, code) \
875 STMT_START { \
d62e749e
KW
876 if (! RExC_copy_start_in_constructed) { \
877 Perl_croak( aTHX_ "panic! %s: %d: Tried to warn when none" \
878 " expected at '%s'", \
879 __FILE__, __LINE__, loc); \
880 } \
55bcf857 881 if (TO_OUTPUT_WARNINGS(loc)) { \
086f21bf
KW
882 if (ckDEAD(warns)) \
883 PREPARE_TO_DIE; \
338190a1 884 code; \
77ccba3c 885 UPDATE_WARNINGS_LOC(loc); \
338190a1 886 } \
f8220fe2 887 } STMT_END
946095af 888
5e0a247b 889/* m is not necessarily a "literal string", in this macro */
58ad8160
KW
890#define warn_non_literal_string(loc, packed_warn, m) \
891 _WARN_HELPER(loc, packed_warn, \
892 Perl_warner(aTHX_ packed_warn, \
d528642a 893 "%s" REPORT_LOCATION, \
f8220fe2 894 m, REPORT_LOCATION_ARGS(loc)))
58ad8160
KW
895#define reg_warn_non_literal_string(loc, m) \
896 warn_non_literal_string(loc, packWARN(WARN_REGEXP), m)
5e0a247b 897
eb761011
KW
898#define ckWARN2_non_literal_string(loc, packwarn, m, a1) \
899 STMT_START { \
900 char * format; \
901 Size_t format_size = strlen(m) + strlen(REPORT_LOCATION)+ 1;\
902 Newx(format, format_size, char); \
903 my_strlcpy(format, m, format_size); \
904 my_strlcat(format, REPORT_LOCATION, format_size); \
905 SAVEFREEPV(format); \
906 _WARN_HELPER(loc, packwarn, \
907 Perl_ck_warner(aTHX_ packwarn, \
908 format, \
909 a1, REPORT_LOCATION_ARGS(loc))); \
910 } STMT_END
911
f8220fe2
KW
912#define ckWARNreg(loc,m) \
913 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
914 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
d528642a 915 m REPORT_LOCATION, \
f8220fe2 916 REPORT_LOCATION_ARGS(loc)))
ccb2c380 917
f8220fe2
KW
918#define vWARN(loc, m) \
919 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
920 Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
d528642a 921 m REPORT_LOCATION, \
f8220fe2 922 REPORT_LOCATION_ARGS(loc))) \
b927b7e9 923
f8220fe2
KW
924#define vWARN_dep(loc, m) \
925 _WARN_HELPER(loc, packWARN(WARN_DEPRECATED), \
926 Perl_warner(aTHX_ packWARN(WARN_DEPRECATED), \
d528642a 927 m REPORT_LOCATION, \
f8220fe2 928 REPORT_LOCATION_ARGS(loc)))
0d6106aa 929
f8220fe2
KW
930#define ckWARNdep(loc,m) \
931 _WARN_HELPER(loc, packWARN(WARN_DEPRECATED), \
932 Perl_ck_warner_d(aTHX_ packWARN(WARN_DEPRECATED), \
d528642a 933 m REPORT_LOCATION, \
f8220fe2 934 REPORT_LOCATION_ARGS(loc)))
147508a2 935
f8220fe2
KW
936#define ckWARNregdep(loc,m) \
937 _WARN_HELPER(loc, packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
938 Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, \
d528642a
KW
939 WARN_REGEXP), \
940 m REPORT_LOCATION, \
f8220fe2 941 REPORT_LOCATION_ARGS(loc)))
ccb2c380 942
f8220fe2
KW
943#define ckWARN2reg_d(loc,m, a1) \
944 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
945 Perl_ck_warner_d(aTHX_ packWARN(WARN_REGEXP), \
d528642a 946 m REPORT_LOCATION, \
f8220fe2 947 a1, REPORT_LOCATION_ARGS(loc)))
2335b3d3 948
f8220fe2
KW
949#define ckWARN2reg(loc, m, a1) \
950 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
951 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
d528642a 952 m REPORT_LOCATION, \
f8220fe2 953 a1, REPORT_LOCATION_ARGS(loc)))
ccb2c380 954
f8220fe2
KW
955#define vWARN3(loc, m, a1, a2) \
956 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
957 Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
d528642a 958 m REPORT_LOCATION, \
f8220fe2 959 a1, a2, REPORT_LOCATION_ARGS(loc)))
ccb2c380 960
f8220fe2
KW
961#define ckWARN3reg(loc, m, a1, a2) \
962 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
963 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
d528642a
KW
964 m REPORT_LOCATION, \
965 a1, a2, \
f8220fe2 966 REPORT_LOCATION_ARGS(loc)))
668c081a 967
f8220fe2
KW
968#define vWARN4(loc, m, a1, a2, a3) \
969 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
970 Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
d528642a
KW
971 m REPORT_LOCATION, \
972 a1, a2, a3, \
f8220fe2 973 REPORT_LOCATION_ARGS(loc)))
ccb2c380 974
f8220fe2
KW
975#define ckWARN4reg(loc, m, a1, a2, a3) \
976 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
977 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
d528642a
KW
978 m REPORT_LOCATION, \
979 a1, a2, a3, \
f8220fe2 980 REPORT_LOCATION_ARGS(loc)))
668c081a 981
f8220fe2
KW
982#define vWARN5(loc, m, a1, a2, a3, a4) \
983 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
984 Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
d528642a
KW
985 m REPORT_LOCATION, \
986 a1, a2, a3, a4, \
f8220fe2 987 REPORT_LOCATION_ARGS(loc)))
9d1d55b5 988
aea5f181
KW
989#define ckWARNexperimental(loc, class, m) \
990 _WARN_HELPER(loc, packWARN(class), \
991 Perl_ck_warner_d(aTHX_ packWARN(class), \
992 m REPORT_LOCATION, \
993 REPORT_LOCATION_ARGS(loc)))
994
1a322bb4
KW
995/* Convert between a pointer to a node and its offset from the beginning of the
996 * program */
997#define REGNODE_p(offset) (RExC_emit_start + (offset))
998#define REGNODE_OFFSET(node) ((node) - RExC_emit_start)
999
538e84ed 1000/* Macros for recording node offsets. 20001227 mjd@plover.com
fac92740
MJD
1001 * Nodes are numbered 1, 2, 3, 4. Node #n's position is recorded in
1002 * element 2*n-1 of the array. Element #2n holds the byte length node #n.
1003 * Element 0 holds the number n.
07be1b83 1004 * Position is 1 indexed.
fac92740 1005 */
7122b237 1006#ifndef RE_TRACK_PATTERN_OFFSETS
d7c442ce 1007#define Set_Node_Offset_To_R(offset,byte)
7122b237
YO
1008#define Set_Node_Offset(node,byte)
1009#define Set_Cur_Node_Offset
1010#define Set_Node_Length_To_R(node,len)
1011#define Set_Node_Length(node,len)
6a86c6ad 1012#define Set_Node_Cur_Length(node,start)
538e84ed
KW
1013#define Node_Offset(n)
1014#define Node_Length(n)
7122b237
YO
1015#define Set_Node_Offset_Length(node,offset,len)
1016#define ProgLen(ri) ri->u.proglen
1017#define SetProgLen(ri,x) ri->u.proglen = x
4cbeff97 1018#define Track_Code(code)
7122b237
YO
1019#else
1020#define ProgLen(ri) ri->u.offsets[0]
1021#define SetProgLen(ri,x) ri->u.offsets[0] = x
d7c442ce 1022#define Set_Node_Offset_To_R(offset,byte) STMT_START { \
ccb2c380 1023 MJD_OFFSET_DEBUG(("** (%d) offset of node %d is %d.\n", \
d7c442ce
KW
1024 __LINE__, (int)(offset), (int)(byte))); \
1025 if((offset) < 0) { \
538e84ed 1026 Perl_croak(aTHX_ "value of node is %d in Offset macro", \
d7c442ce 1027 (int)(offset)); \
ccb2c380 1028 } else { \
d7c442ce 1029 RExC_offsets[2*(offset)-1] = (byte); \
ccb2c380 1030 } \
ccb2c380
MP
1031} STMT_END
1032
88f063b4 1033#define Set_Node_Offset(node,byte) \
1a322bb4 1034 Set_Node_Offset_To_R(REGNODE_OFFSET(node), (byte)-RExC_start)
ccb2c380
MP
1035#define Set_Cur_Node_Offset Set_Node_Offset(RExC_emit, RExC_parse)
1036
1037#define Set_Node_Length_To_R(node,len) STMT_START { \
ccb2c380 1038 MJD_OFFSET_DEBUG(("** (%d) size of node %d is %d.\n", \
551405c4 1039 __LINE__, (int)(node), (int)(len))); \
ccb2c380 1040 if((node) < 0) { \
538e84ed
KW
1041 Perl_croak(aTHX_ "value of node is %d in Length macro", \
1042 (int)(node)); \
ccb2c380
MP
1043 } else { \
1044 RExC_offsets[2*(node)] = (len); \
1045 } \
ccb2c380
MP
1046} STMT_END
1047
1048#define Set_Node_Length(node,len) \
1a322bb4 1049 Set_Node_Length_To_R(REGNODE_OFFSET(node), len)
6a86c6ad
NC
1050#define Set_Node_Cur_Length(node, start) \
1051 Set_Node_Length(node, RExC_parse - start)
fac92740
MJD
1052
1053/* Get offsets and lengths */
1a322bb4
KW
1054#define Node_Offset(n) (RExC_offsets[2*(REGNODE_OFFSET(n))-1])
1055#define Node_Length(n) (RExC_offsets[2*(REGNODE_OFFSET(n))])
fac92740 1056
07be1b83 1057#define Set_Node_Offset_Length(node,offset,len) STMT_START { \
1a322bb4
KW
1058 Set_Node_Offset_To_R(REGNODE_OFFSET(node), (offset)); \
1059 Set_Node_Length_To_R(REGNODE_OFFSET(node), (len)); \
07be1b83 1060} STMT_END
4cbeff97
TC
1061
1062#define Track_Code(code) STMT_START { code } STMT_END
7122b237 1063#endif
07be1b83
YO
1064
1065#if PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS
1066#define EXPERIMENTAL_INPLACESCAN
f427392e 1067#endif /*PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS*/
07be1b83 1068
cb41e5d6
YO
1069#ifdef DEBUGGING
1070int
6ad9a8ab 1071Perl_re_printf(pTHX_ const char *fmt, ...)
cb41e5d6
YO
1072{
1073 va_list ap;
1074 int result;
1075 PerlIO *f= Perl_debug_log;
1076 PERL_ARGS_ASSERT_RE_PRINTF;
1077 va_start(ap, fmt);
1078 result = PerlIO_vprintf(f, fmt, ap);
1079 va_end(ap);
1080 return result;
1081}
1082
1083int
7b031478 1084Perl_re_indentf(pTHX_ const char *fmt, U32 depth, ...)
cb41e5d6
YO
1085{
1086 va_list ap;
1087 int result;
1088 PerlIO *f= Perl_debug_log;
1089 PERL_ARGS_ASSERT_RE_INDENTF;
1090 va_start(ap, depth);
daeb874b 1091 PerlIO_printf(f, "%*s", ( (int)depth % 20 ) * 2, "");
cb41e5d6
YO
1092 result = PerlIO_vprintf(f, fmt, ap);
1093 va_end(ap);
1094 return result;
1095}
1096#endif /* DEBUGGING */
1097
7b031478 1098#define DEBUG_RExC_seen() \
538e84ed 1099 DEBUG_OPTIMISE_MORE_r({ \
88f063b4 1100 Perl_re_printf( aTHX_ "RExC_seen: "); \
538e84ed 1101 \
e384d5c1 1102 if (RExC_seen & REG_ZERO_LEN_SEEN) \
88f063b4 1103 Perl_re_printf( aTHX_ "REG_ZERO_LEN_SEEN "); \
538e84ed 1104 \
e384d5c1 1105 if (RExC_seen & REG_LOOKBEHIND_SEEN) \
88f063b4 1106 Perl_re_printf( aTHX_ "REG_LOOKBEHIND_SEEN "); \
538e84ed 1107 \
e384d5c1 1108 if (RExC_seen & REG_GPOS_SEEN) \
88f063b4 1109 Perl_re_printf( aTHX_ "REG_GPOS_SEEN "); \
538e84ed 1110 \
e384d5c1 1111 if (RExC_seen & REG_RECURSE_SEEN) \
88f063b4 1112 Perl_re_printf( aTHX_ "REG_RECURSE_SEEN "); \
538e84ed 1113 \
7b031478 1114 if (RExC_seen & REG_TOP_LEVEL_BRANCHES_SEEN) \
88f063b4 1115 Perl_re_printf( aTHX_ "REG_TOP_LEVEL_BRANCHES_SEEN "); \
538e84ed 1116 \
e384d5c1 1117 if (RExC_seen & REG_VERBARG_SEEN) \
88f063b4 1118 Perl_re_printf( aTHX_ "REG_VERBARG_SEEN "); \
538e84ed 1119 \
e384d5c1 1120 if (RExC_seen & REG_CUTGROUP_SEEN) \
88f063b4 1121 Perl_re_printf( aTHX_ "REG_CUTGROUP_SEEN "); \
538e84ed 1122 \
e384d5c1 1123 if (RExC_seen & REG_RUN_ON_COMMENT_SEEN) \
88f063b4 1124 Perl_re_printf( aTHX_ "REG_RUN_ON_COMMENT_SEEN "); \
538e84ed 1125 \
e384d5c1 1126 if (RExC_seen & REG_UNFOLDED_MULTI_SEEN) \
88f063b4 1127 Perl_re_printf( aTHX_ "REG_UNFOLDED_MULTI_SEEN "); \
538e84ed 1128 \
7b031478 1129 if (RExC_seen & REG_UNBOUNDED_QUANTIFIER_SEEN) \
88f063b4 1130 Perl_re_printf( aTHX_ "REG_UNBOUNDED_QUANTIFIER_SEEN "); \
ee273784 1131 \
88f063b4 1132 Perl_re_printf( aTHX_ "\n"); \
9e9ecfdf
YO
1133 });
1134
fdfb4f21 1135#define DEBUG_SHOW_STUDY_FLAG(flags,flag) \
6ad9a8ab 1136 if ((flags) & flag) Perl_re_printf( aTHX_ "%s ", #flag)
fdfb4f21 1137
fdfb4f21 1138
f5a36d78
DM
1139#ifdef DEBUGGING
1140static void
1141S_debug_show_study_flags(pTHX_ U32 flags, const char *open_str,
1142 const char *close_str)
1143{
1144 if (!flags)
1145 return;
1146
1147 Perl_re_printf( aTHX_ "%s", open_str);
11683ecb
DM
1148 DEBUG_SHOW_STUDY_FLAG(flags, SF_BEFORE_SEOL);
1149 DEBUG_SHOW_STUDY_FLAG(flags, SF_BEFORE_MEOL);
f5a36d78
DM
1150 DEBUG_SHOW_STUDY_FLAG(flags, SF_IS_INF);
1151 DEBUG_SHOW_STUDY_FLAG(flags, SF_HAS_PAR);
1152 DEBUG_SHOW_STUDY_FLAG(flags, SF_IN_PAR);
1153 DEBUG_SHOW_STUDY_FLAG(flags, SF_HAS_EVAL);
1154 DEBUG_SHOW_STUDY_FLAG(flags, SCF_DO_SUBSTR);
1155 DEBUG_SHOW_STUDY_FLAG(flags, SCF_DO_STCLASS_AND);
1156 DEBUG_SHOW_STUDY_FLAG(flags, SCF_DO_STCLASS_OR);
1157 DEBUG_SHOW_STUDY_FLAG(flags, SCF_DO_STCLASS);
1158 DEBUG_SHOW_STUDY_FLAG(flags, SCF_WHILEM_VISITED_POS);
1159 DEBUG_SHOW_STUDY_FLAG(flags, SCF_TRIE_RESTUDY);
1160 DEBUG_SHOW_STUDY_FLAG(flags, SCF_SEEN_ACCEPT);
1161 DEBUG_SHOW_STUDY_FLAG(flags, SCF_TRIE_DOING_RESTUDY);
1162 DEBUG_SHOW_STUDY_FLAG(flags, SCF_IN_DEFINE);
1163 Perl_re_printf( aTHX_ "%s", close_str);
1164}
1165
1166
1167static void
1168S_debug_studydata(pTHX_ const char *where, scan_data_t *data,
1169 U32 depth, int is_inf)
1170{
1171 GET_RE_DEBUG_FLAGS_DECL;
1172
1173 DEBUG_OPTIMISE_MORE_r({
1174 if (!data)
1175 return;
1176 Perl_re_indentf(aTHX_ "%s: Pos:%" IVdf "/%" IVdf " Flags: 0x%" UVXf,
1177 depth,
1178 where,
1179 (IV)data->pos_min,
1180 (IV)data->pos_delta,
1181 (UV)data->flags
1182 );
1183
1184 S_debug_show_study_flags(aTHX_ data->flags," [","]");
fdfb4f21 1185
f5a36d78
DM
1186 Perl_re_printf( aTHX_
1187 " Whilem_c: %" IVdf " Lcp: %" IVdf " %s",
1188 (IV)data->whilem_c,
1189 (IV)(data->last_closep ? *((data)->last_closep) : -1),
1190 is_inf ? "INF " : ""
1191 );
1192
11683ecb 1193 if (data->last_found) {
37b6262f 1194 int i;
11683ecb
DM
1195 Perl_re_printf(aTHX_
1196 "Last:'%s' %" IVdf ":%" IVdf "/%" IVdf,
1197 SvPVX_const(data->last_found),
1198 (IV)data->last_end,
1199 (IV)data->last_start_min,
1200 (IV)data->last_start_max
1201 );
1202
2099df82 1203 for (i = 0; i < 2; i++) {
37b6262f
DM
1204 Perl_re_printf(aTHX_
1205 " %s%s: '%s' @ %" IVdf "/%" IVdf,
2099df82
DM
1206 data->cur_is_floating == i ? "*" : "",
1207 i ? "Float" : "Fixed",
37b6262f
DM
1208 SvPVX_const(data->substrs[i].str),
1209 (IV)data->substrs[i].min_offset,
1210 (IV)data->substrs[i].max_offset
1211 );
1212 S_debug_show_study_flags(aTHX_ data->substrs[i].flags," [","]");
1213 }
11683ecb 1214 }
f5a36d78
DM
1215
1216 Perl_re_printf( aTHX_ "\n");
1217 });
1218}
1219
1220
1221static void
1222S_debug_peep(pTHX_ const char *str, const RExC_state_t *pRExC_state,
1223 regnode *scan, U32 depth, U32 flags)
1224{
1225 GET_RE_DEBUG_FLAGS_DECL;
1226
1227 DEBUG_OPTIMISE_r({
1228 regnode *Next;
1229
1230 if (!scan)
1231 return;
1232 Next = regnext(scan);
1233 regprop(RExC_rx, RExC_mysv, scan, NULL, pRExC_state);
1234 Perl_re_indentf( aTHX_ "%s>%3d: %s (%d)",
1235 depth,
1236 str,
1237 REG_NODE_NUM(scan), SvPV_nolen_const(RExC_mysv),
1238 Next ? (REG_NODE_NUM(Next)) : 0 );
1239 S_debug_show_study_flags(aTHX_ flags," [ ","]");
1240 Perl_re_printf( aTHX_ "\n");
1241 });
1242}
1243
1244
1245# define DEBUG_STUDYDATA(where, data, depth, is_inf) \
1246 S_debug_studydata(aTHX_ where, data, depth, is_inf)
1247
1248# define DEBUG_PEEP(str, scan, depth, flags) \
1249 S_debug_peep(aTHX_ str, pRExC_state, scan, depth, flags)
1250
1251#else
1252# define DEBUG_STUDYDATA(where, data, depth, is_inf) NOOP
1253# define DEBUG_PEEP(str, scan, depth, flags) NOOP
1254#endif
1de06328 1255
cb41e5d6 1256
c6871b76
KW
1257/* =========================================================
1258 * BEGIN edit_distance stuff.
1259 *
1260 * This calculates how many single character changes of any type are needed to
1261 * transform a string into another one. It is taken from version 3.1 of
1262 *
1263 * https://metacpan.org/pod/Text::Levenshtein::Damerau::XS
1264 */
1265
1266/* Our unsorted dictionary linked list. */
1267/* Note we use UVs, not chars. */
1268
1269struct dictionary{
1270 UV key;
1271 UV value;
1272 struct dictionary* next;
1273};
1274typedef struct dictionary item;
1275
1276
1277PERL_STATIC_INLINE item*
88f063b4 1278push(UV key, item* curr)
c6871b76
KW
1279{
1280 item* head;
d09f14bf 1281 Newx(head, 1, item);
c6871b76
KW
1282 head->key = key;
1283 head->value = 0;
1284 head->next = curr;
1285 return head;
1286}
1287
1288
1289PERL_STATIC_INLINE item*
1290find(item* head, UV key)
1291{
1292 item* iterator = head;
1293 while (iterator){
1294 if (iterator->key == key){
1295 return iterator;
1296 }
1297 iterator = iterator->next;
1298 }
1299
1300 return NULL;
1301}
1302
1303PERL_STATIC_INLINE item*
88f063b4 1304uniquePush(item* head, UV key)
c6871b76
KW
1305{
1306 item* iterator = head;
1307
1308 while (iterator){
1309 if (iterator->key == key) {
1310 return head;
1311 }
1312 iterator = iterator->next;
1313 }
1314
88f063b4 1315 return push(key, head);
c6871b76
KW
1316}
1317
1318PERL_STATIC_INLINE void
1319dict_free(item* head)
1320{
1321 item* iterator = head;
1322
1323 while (iterator) {
1324 item* temp = iterator;
1325 iterator = iterator->next;
1326 Safefree(temp);
1327 }
1328
1329 head = NULL;
1330}
1331
1332/* End of Dictionary Stuff */
1333
1334/* All calculations/work are done here */
1335STATIC int
1336S_edit_distance(const UV* src,
1337 const UV* tgt,
1338 const STRLEN x, /* length of src[] */
1339 const STRLEN y, /* length of tgt[] */
1340 const SSize_t maxDistance
1341)
1342{
1343 item *head = NULL;
88f063b4 1344 UV swapCount, swapScore, targetCharCount, i, j;
c6871b76
KW
1345 UV *scores;
1346 UV score_ceil = x + y;
1347
1348 PERL_ARGS_ASSERT_EDIT_DISTANCE;
1349
1350 /* intialize matrix start values */
d09f14bf 1351 Newx(scores, ( (x + 2) * (y + 2)), UV);
c6871b76
KW
1352 scores[0] = score_ceil;
1353 scores[1 * (y + 2) + 0] = score_ceil;
1354 scores[0 * (y + 2) + 1] = score_ceil;
1355 scores[1 * (y + 2) + 1] = 0;
88f063b4 1356 head = uniquePush(uniquePush(head, src[0]), tgt[0]);
c6871b76
KW
1357
1358 /* work loops */
1359 /* i = src index */
1360 /* j = tgt index */
1361 for (i=1;i<=x;i++) {
1362 if (i < x)
88f063b4 1363 head = uniquePush(head, src[i]);
c6871b76
KW
1364 scores[(i+1) * (y + 2) + 1] = i;
1365 scores[(i+1) * (y + 2) + 0] = score_ceil;
1366 swapCount = 0;
1367
1368 for (j=1;j<=y;j++) {
1369 if (i == 1) {
1370 if(j < y)
88f063b4 1371 head = uniquePush(head, tgt[j]);
c6871b76
KW
1372 scores[1 * (y + 2) + (j + 1)] = j;
1373 scores[0 * (y + 2) + (j + 1)] = score_ceil;
1374 }
1375
88f063b4 1376 targetCharCount = find(head, tgt[j-1])->value;
c6871b76
KW
1377 swapScore = scores[targetCharCount * (y + 2) + swapCount] + i - targetCharCount - 1 + j - swapCount;
1378
1379 if (src[i-1] != tgt[j-1]){
1380 scores[(i+1) * (y + 2) + (j + 1)] = MIN(swapScore,(MIN(scores[i * (y + 2) + j], MIN(scores[(i+1) * (y + 2) + j], scores[i * (y + 2) + (j + 1)])) + 1));
1381 }
1382 else {
1383 swapCount = j;
1384 scores[(i+1) * (y + 2) + (j + 1)] = MIN(scores[i * (y + 2) + j], swapScore);
1385 }
1386 }
1387
88f063b4 1388 find(head, src[i-1])->value = i;
c6871b76
KW
1389 }
1390
1391 {
1392 IV score = scores[(x+1) * (y + 2) + (y + 1)];
1393 dict_free(head);
1394 Safefree(scores);
1395 return (maxDistance != 0 && maxDistance < score)?(-1):score;
1396 }
1397}
1398
1399/* END of edit_distance() stuff
1400 * ========================================================= */
1401
653099ff 1402/* Mark that we cannot extend a found fixed substring at this point.
37b6262f 1403 Update the longest found anchored substring or the longest found
653099ff
GS
1404 floating substrings if needed. */
1405
4327152a 1406STATIC void
ea3daa5d
FC
1407S_scan_commit(pTHX_ const RExC_state_t *pRExC_state, scan_data_t *data,
1408 SSize_t *minlenp, int is_inf)
c277df42 1409{
e1ec3a88 1410 const STRLEN l = CHR_SVLEN(data->last_found);
2099df82 1411 SV * const longest_sv = data->substrs[data->cur_is_floating].str;
a2ca2017 1412 const STRLEN old_l = CHR_SVLEN(longest_sv);
1de06328 1413 GET_RE_DEBUG_FLAGS_DECL;
b81d288d 1414
7918f24d
NC
1415 PERL_ARGS_ASSERT_SCAN_COMMIT;
1416
c277df42 1417 if ((l >= old_l) && ((l > old_l) || (data->flags & SF_BEFORE_EOL))) {
2099df82 1418 const U8 i = data->cur_is_floating;
a2ca2017 1419 SvSetMagicSV(longest_sv, data->last_found);
37b6262f
DM
1420 data->substrs[i].min_offset = l ? data->last_start_min : data->pos_min;
1421
2099df82
DM
1422 if (!i) /* fixed */
1423 data->substrs[0].max_offset = data->substrs[0].min_offset;
1424 else { /* float */
1425 data->substrs[1].max_offset = (l
646e8787
DM
1426 ? data->last_start_max
1427 : (data->pos_delta > SSize_t_MAX - data->pos_min
ea3daa5d
FC
1428 ? SSize_t_MAX
1429 : data->pos_min + data->pos_delta));
1430 if (is_inf
2099df82
DM
1431 || (STRLEN)data->substrs[1].max_offset > (STRLEN)SSize_t_MAX)
1432 data->substrs[1].max_offset = SSize_t_MAX;
37b6262f 1433 }
11683ecb 1434
37b6262f
DM
1435 if (data->flags & SF_BEFORE_EOL)
1436 data->substrs[i].flags |= (data->flags & SF_BEFORE_EOL);
1437 else
1438 data->substrs[i].flags &= ~SF_BEFORE_EOL;
1439 data->substrs[i].minlenp = minlenp;
1440 data->substrs[i].lookbehind = 0;
c277df42 1441 }
37b6262f 1442
c277df42 1443 SvCUR_set(data->last_found, 0);
0eda9292 1444 {
a28509cc 1445 SV * const sv = data->last_found;
097eb12c
AL
1446 if (SvUTF8(sv) && SvMAGICAL(sv)) {
1447 MAGIC * const mg = mg_find(sv, PERL_MAGIC_utf8);
1448 if (mg)
1449 mg->mg_len = 0;
1450 }
0eda9292 1451 }
c277df42
IZ
1452 data->last_end = -1;
1453 data->flags &= ~SF_BEFORE_EOL;
f5a36d78 1454 DEBUG_STUDYDATA("commit", data, 0, is_inf);
c277df42
IZ
1455}
1456
cdd87c1d
KW
1457/* An SSC is just a regnode_charclass_posix with an extra field: the inversion
1458 * list that describes which code points it matches */
1459
653099ff 1460STATIC void
3420edd7 1461S_ssc_anything(pTHX_ regnode_ssc *ssc)
653099ff 1462{
cdd87c1d
KW
1463 /* Set the SSC 'ssc' to match an empty string or any code point */
1464
557bd3fb 1465 PERL_ARGS_ASSERT_SSC_ANYTHING;
7918f24d 1466
71068078 1467 assert(is_ANYOF_SYNTHETIC(ssc));
3fffb88a 1468
0854ea0b
KW
1469 /* mortalize so won't leak */
1470 ssc->invlist = sv_2mortal(_add_range_to_invlist(NULL, 0, UV_MAX));
93e92956 1471 ANYOF_FLAGS(ssc) |= SSC_MATCHES_EMPTY_STRING; /* Plus matches empty */
653099ff
GS
1472}
1473
653099ff 1474STATIC int
dc3bf405 1475S_ssc_is_anything(const regnode_ssc *ssc)
653099ff 1476{
c144baaa
KW
1477 /* Returns TRUE if the SSC 'ssc' can match the empty string and any code
1478 * point; FALSE otherwise. Thus, this is used to see if using 'ssc' buys
1479 * us anything: if the function returns TRUE, 'ssc' hasn't been restricted
1480 * in any way, so there's no point in using it */
cdd87c1d
KW
1481
1482 UV start, end;
1483 bool ret;
653099ff 1484
557bd3fb 1485 PERL_ARGS_ASSERT_SSC_IS_ANYTHING;
7918f24d 1486
71068078 1487 assert(is_ANYOF_SYNTHETIC(ssc));
cdd87c1d 1488
93e92956 1489 if (! (ANYOF_FLAGS(ssc) & SSC_MATCHES_EMPTY_STRING)) {
cdd87c1d
KW
1490 return FALSE;
1491 }
1492
1493 /* See if the list consists solely of the range 0 - Infinity */
1494 invlist_iterinit(ssc->invlist);
1495 ret = invlist_iternext(ssc->invlist, &start, &end)
1496 && start == 0
1497 && end == UV_MAX;
1498
1499 invlist_iterfinish(ssc->invlist);
1500
1501 if (ret) {
1502 return TRUE;
1503 }
1504
1505 /* If e.g., both \w and \W are set, matches everything */
e0e1be5f 1506 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
cdd87c1d
KW
1507 int i;
1508 for (i = 0; i < ANYOF_POSIXL_MAX; i += 2) {
1509 if (ANYOF_POSIXL_TEST(ssc, i) && ANYOF_POSIXL_TEST(ssc, i+1)) {
1510 return TRUE;
1511 }
1512 }
1513 }
1514
1515 return FALSE;
653099ff
GS
1516}
1517
653099ff 1518STATIC void
cdd87c1d 1519S_ssc_init(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc)
653099ff 1520{
cdd87c1d
KW
1521 /* Initializes the SSC 'ssc'. This includes setting it to match an empty
1522 * string, any code point, or any posix class under locale */
1523
557bd3fb 1524 PERL_ARGS_ASSERT_SSC_INIT;
7918f24d 1525
557bd3fb 1526 Zero(ssc, 1, regnode_ssc);
71068078 1527 set_ANYOF_SYNTHETIC(ssc);
93e92956 1528 ARG_SET(ssc, ANYOF_ONLY_HAS_BITMAP);
3420edd7 1529 ssc_anything(ssc);
cdd87c1d 1530
2f306ab9
KW
1531 /* If any portion of the regex is to operate under locale rules that aren't
1532 * fully known at compile time, initialization includes it. The reason
1533 * this isn't done for all regexes is that the optimizer was written under
1534 * the assumption that locale was all-or-nothing. Given the complexity and
1535 * lack of documentation in the optimizer, and that there are inadequate
1536 * test cases for locale, many parts of it may not work properly, it is
1537 * safest to avoid locale unless necessary. */
cdd87c1d
KW
1538 if (RExC_contains_locale) {
1539 ANYOF_POSIXL_SETALL(ssc);
cdd87c1d
KW
1540 }
1541 else {
1542 ANYOF_POSIXL_ZERO(ssc);
1543 }
653099ff
GS
1544}
1545
b423522f 1546STATIC int
dc3bf405
BF
1547S_ssc_is_cp_posixl_init(const RExC_state_t *pRExC_state,
1548 const regnode_ssc *ssc)
b423522f
KW
1549{
1550 /* Returns TRUE if the SSC 'ssc' is in its initial state with regard only
1551 * to the list of code points matched, and locale posix classes; hence does
1552 * not check its flags) */
1553
1554 UV start, end;
1555 bool ret;
1556
1557 PERL_ARGS_ASSERT_SSC_IS_CP_POSIXL_INIT;
1558
71068078 1559 assert(is_ANYOF_SYNTHETIC(ssc));
b423522f
KW
1560
1561 invlist_iterinit(ssc->invlist);
1562 ret = invlist_iternext(ssc->invlist, &start, &end)
1563 && start == 0
1564 && end == UV_MAX;
1565
1566 invlist_iterfinish(ssc->invlist);
1567
1568 if (! ret) {
1569 return FALSE;
1570 }
1571
e0e1be5f 1572 if (RExC_contains_locale && ! ANYOF_POSIXL_SSC_TEST_ALL_SET(ssc)) {
31f05a37 1573 return FALSE;
b423522f
KW
1574 }
1575
1576 return TRUE;
1577}
1578
4ebed06a
KW
1579#define INVLIST_INDEX 0
1580#define ONLY_LOCALE_MATCHES_INDEX 1
1581#define DEFERRED_USER_DEFINED_INDEX 2
1582
b423522f
KW
1583STATIC SV*
1584S_get_ANYOF_cp_list_for_ssc(pTHX_ const RExC_state_t *pRExC_state,
5c0f85ef 1585 const regnode_charclass* const node)
b423522f
KW
1586{
1587 /* Returns a mortal inversion list defining which code points are matched
1588 * by 'node', which is of type ANYOF. Handles complementing the result if
1589 * appropriate. If some code points aren't knowable at this time, the
31f05a37
KW
1590 * returned list must, and will, contain every code point that is a
1591 * possibility. */
b423522f 1592
1565c085 1593 dVAR;
e2506fa7 1594 SV* invlist = NULL;
1ee208c4 1595 SV* only_utf8_locale_invlist = NULL;
b423522f
KW
1596 unsigned int i;
1597 const U32 n = ARG(node);
31f05a37 1598 bool new_node_has_latin1 = FALSE;
2d5613be 1599 const U8 flags = (inRANGE(OP(node), ANYOFH, ANYOFRb))
29a889ef
KW
1600 ? 0
1601 : ANYOF_FLAGS(node);
b423522f
KW
1602
1603 PERL_ARGS_ASSERT_GET_ANYOF_CP_LIST_FOR_SSC;
1604
1605 /* Look at the data structure created by S_set_ANYOF_arg() */
93e92956 1606 if (n != ANYOF_ONLY_HAS_BITMAP) {
b423522f
KW
1607 SV * const rv = MUTABLE_SV(RExC_rxi->data->data[n]);
1608 AV * const av = MUTABLE_AV(SvRV(rv));
1609 SV **const ary = AvARRAY(av);
1610 assert(RExC_rxi->data->what[n] == 's');
1611
4ebed06a 1612 if (av_tindex_skip_len_mg(av) >= DEFERRED_USER_DEFINED_INDEX) {
b423522f 1613
6ed02bb6
KW
1614 /* Here there are things that won't be known until runtime -- we
1615 * have to assume it could be anything */
e2506fa7 1616 invlist = sv_2mortal(_new_invlist(1));
b423522f
KW
1617 return _add_range_to_invlist(invlist, 0, UV_MAX);
1618 }
4ebed06a 1619 else if (ary[INVLIST_INDEX]) {
b423522f 1620
6ed02bb6 1621 /* Use the node's inversion list */
4ebed06a 1622 invlist = sv_2mortal(invlist_clone(ary[INVLIST_INDEX], NULL));
1ee208c4
KW
1623 }
1624
1625 /* Get the code points valid only under UTF-8 locales */
766d6d33 1626 if ( (flags & ANYOFL_FOLD)
4ebed06a 1627 && av_tindex_skip_len_mg(av) >= ONLY_LOCALE_MATCHES_INDEX)
1ee208c4 1628 {
4ebed06a 1629 only_utf8_locale_invlist = ary[ONLY_LOCALE_MATCHES_INDEX];
b423522f
KW
1630 }
1631 }
1632
e2506fa7
KW
1633 if (! invlist) {
1634 invlist = sv_2mortal(_new_invlist(0));
1635 }
1636
dcb20b36
KW
1637 /* An ANYOF node contains a bitmap for the first NUM_ANYOF_CODE_POINTS
1638 * code points, and an inversion list for the others, but if there are code
1639 * points that should match only conditionally on the target string being
1640 * UTF-8, those are placed in the inversion list, and not the bitmap.
1641 * Since there are circumstances under which they could match, they are
1642 * included in the SSC. But if the ANYOF node is to be inverted, we have
1643 * to exclude them here, so that when we invert below, the end result
1644 * actually does include them. (Think about "\xe0" =~ /[^\xc0]/di;). We
1645 * have to do this here before we add the unconditionally matched code
1646 * points */
766d6d33 1647 if (flags & ANYOF_INVERT) {
b423522f
KW
1648 _invlist_intersection_complement_2nd(invlist,
1649 PL_UpperLatin1,
1650 &invlist);
1651 }
1652
1653 /* Add in the points from the bit map */
2d5613be 1654 if (! inRANGE(OP(node), ANYOFH, ANYOFRb)) {
90973738
KW
1655 for (i = 0; i < NUM_ANYOF_CODE_POINTS; i++) {
1656 if (ANYOF_BITMAP_TEST(node, i)) {
1657 unsigned int start = i++;
1658
1659 for (; i < NUM_ANYOF_CODE_POINTS
1660 && ANYOF_BITMAP_TEST(node, i); ++i)
1661 {
1662 /* empty */
1663 }
1664 invlist = _add_range_to_invlist(invlist, start, i-1);
1665 new_node_has_latin1 = TRUE;
6f8848d5 1666 }
b423522f
KW
1667 }
1668 }
1669
1670 /* If this can match all upper Latin1 code points, have to add them
ac33c516
KW
1671 * as well. But don't add them if inverting, as when that gets done below,
1672 * it would exclude all these characters, including the ones it shouldn't
1673 * that were added just above */
766d6d33
KW
1674 if (! (flags & ANYOF_INVERT) && OP(node) == ANYOFD
1675 && (flags & ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER))
f240c685 1676 {
b423522f
KW
1677 _invlist_union(invlist, PL_UpperLatin1, &invlist);
1678 }
1679
1680 /* Similarly for these */
766d6d33 1681 if (flags & ANYOF_MATCHES_ALL_ABOVE_BITMAP) {
e0a1ff7a 1682 _invlist_union_complement_2nd(invlist, PL_InBitmap, &invlist);
b423522f
KW
1683 }
1684
766d6d33 1685 if (flags & ANYOF_INVERT) {
b423522f
KW
1686 _invlist_invert(invlist);
1687 }
766d6d33 1688 else if (flags & ANYOFL_FOLD) {
35b8412f 1689 if (new_node_has_latin1) {
31f05a37 1690
26be5fe6
KW
1691 /* Under /li, any 0-255 could fold to any other 0-255, depending on
1692 * the locale. We can skip this if there are no 0-255 at all. */
1693 _invlist_union(invlist, PL_Latin1, &invlist);
35b8412f
KW
1694
1695 invlist = add_cp_to_invlist(invlist, LATIN_SMALL_LETTER_DOTLESS_I);
1696 invlist = add_cp_to_invlist(invlist, LATIN_CAPITAL_LETTER_I_WITH_DOT_ABOVE);
1697 }
1698 else {
1699 if (_invlist_contains_cp(invlist, LATIN_SMALL_LETTER_DOTLESS_I)) {
1700 invlist = add_cp_to_invlist(invlist, 'I');
1701 }
1702 if (_invlist_contains_cp(invlist,
1703 LATIN_CAPITAL_LETTER_I_WITH_DOT_ABOVE))
1704 {
1705 invlist = add_cp_to_invlist(invlist, 'i');
1706 }
1707 }
31f05a37
KW
1708 }
1709
1ee208c4
KW
1710 /* Similarly add the UTF-8 locale possible matches. These have to be
1711 * deferred until after the non-UTF-8 locale ones are taken care of just
1712 * above, or it leads to wrong results under ANYOF_INVERT */
1713 if (only_utf8_locale_invlist) {
31f05a37 1714 _invlist_union_maybe_complement_2nd(invlist,
1ee208c4 1715 only_utf8_locale_invlist,
766d6d33 1716 flags & ANYOF_INVERT,
31f05a37
KW
1717 &invlist);
1718 }
b423522f
KW
1719
1720 return invlist;
1721}
1722
1051e1c4 1723/* These two functions currently do the exact same thing */
557bd3fb 1724#define ssc_init_zero ssc_init
653099ff 1725
cdd87c1d
KW
1726#define ssc_add_cp(ssc, cp) ssc_add_range((ssc), (cp), (cp))
1727#define ssc_match_all_cp(ssc) ssc_add_range(ssc, 0, UV_MAX)
1728
557bd3fb 1729/* 'AND' a given class with another one. Can create false positives. 'ssc'
93e92956
KW
1730 * should not be inverted. 'and_with->flags & ANYOF_MATCHES_POSIXL' should be
1731 * 0 if 'and_with' is a regnode_charclass instead of a regnode_ssc. */
cdd87c1d 1732
653099ff 1733STATIC void
b423522f 1734S_ssc_and(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc,
7dcac5f6 1735 const regnode_charclass *and_with)
653099ff 1736{
cdd87c1d
KW
1737 /* Accumulate into SSC 'ssc' its 'AND' with 'and_with', which is either
1738 * another SSC or a regular ANYOF class. Can create false positives. */
40d049e4 1739
a0dd4231 1740 SV* anded_cp_list;
2d5613be 1741 U8 and_with_flags = inRANGE(OP(and_with), ANYOFH, ANYOFRb)
29a889ef
KW
1742 ? 0
1743 : ANYOF_FLAGS(and_with);
a0dd4231 1744 U8 anded_flags;
1e6ade67 1745
cdd87c1d 1746 PERL_ARGS_ASSERT_SSC_AND;
653099ff 1747
71068078 1748 assert(is_ANYOF_SYNTHETIC(ssc));
a0dd4231
KW
1749
1750 /* 'and_with' is used as-is if it too is an SSC; otherwise have to extract
1751 * the code point inversion list and just the relevant flags */
71068078 1752 if (is_ANYOF_SYNTHETIC(and_with)) {
7dcac5f6 1753 anded_cp_list = ((regnode_ssc *)and_with)->invlist;
765e6ecf 1754 anded_flags = and_with_flags;
e9b08962
KW
1755
1756 /* XXX This is a kludge around what appears to be deficiencies in the
1757 * optimizer. If we make S_ssc_anything() add in the WARN_SUPER flag,
1758 * there are paths through the optimizer where it doesn't get weeded
1759 * out when it should. And if we don't make some extra provision for
1760 * it like the code just below, it doesn't get added when it should.
1761 * This solution is to add it only when AND'ing, which is here, and
1762 * only when what is being AND'ed is the pristine, original node
1763 * matching anything. Thus it is like adding it to ssc_anything() but
1764 * only when the result is to be AND'ed. Probably the same solution
1765 * could be adopted for the same problem we have with /l matching,
1766 * which is solved differently in S_ssc_init(), and that would lead to
1767 * fewer false positives than that solution has. But if this solution
1768 * creates bugs, the consequences are only that a warning isn't raised
1769 * that should be; while the consequences for having /l bugs is
1770 * incorrect matches */
7dcac5f6 1771 if (ssc_is_anything((regnode_ssc *)and_with)) {
f240c685 1772 anded_flags |= ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER;
e9b08962 1773 }
a0dd4231
KW
1774 }
1775 else {
5c0f85ef 1776 anded_cp_list = get_ANYOF_cp_list_for_ssc(pRExC_state, and_with);
f240c685 1777 if (OP(and_with) == ANYOFD) {
765e6ecf 1778 anded_flags = and_with_flags & ANYOF_COMMON_FLAGS;
f240c685
KW
1779 }
1780 else {
765e6ecf 1781 anded_flags = and_with_flags
f240c685 1782 &( ANYOF_COMMON_FLAGS
108316fb
KW
1783 |ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER
1784 |ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP);
765e6ecf 1785 if (ANYOFL_UTF8_LOCALE_REQD(and_with_flags)) {
d1c40ef5
KW
1786 anded_flags &=
1787 ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
1788 }
f240c685 1789 }
a0dd4231
KW
1790 }
1791
1792 ANYOF_FLAGS(ssc) &= anded_flags;
cdd87c1d
KW
1793
1794 /* Below, C1 is the list of code points in 'ssc'; P1, its posix classes.
1795 * C2 is the list of code points in 'and-with'; P2, its posix classes.
1796 * 'and_with' may be inverted. When not inverted, we have the situation of
1797 * computing:
1798 * (C1 | P1) & (C2 | P2)
1799 * = (C1 & (C2 | P2)) | (P1 & (C2 | P2))
1800 * = ((C1 & C2) | (C1 & P2)) | ((P1 & C2) | (P1 & P2))
1801 * <= ((C1 & C2) | P2)) | ( P1 | (P1 & P2))
1802 * <= ((C1 & C2) | P1 | P2)
1803 * Alternatively, the last few steps could be:
1804 * = ((C1 & C2) | (C1 & P2)) | ((P1 & C2) | (P1 & P2))
1805 * <= ((C1 & C2) | C1 ) | ( C2 | (P1 & P2))
1806 * <= (C1 | C2 | (P1 & P2))
1807 * We favor the second approach if either P1 or P2 is non-empty. This is
1808 * because these components are a barrier to doing optimizations, as what
1809 * they match cannot be known until the moment of matching as they are
1810 * dependent on the current locale, 'AND"ing them likely will reduce or
1811 * eliminate them.
1812 * But we can do better if we know that C1,P1 are in their initial state (a
1813 * frequent occurrence), each matching everything:
1814 * (<everything>) & (C2 | P2) = C2 | P2
1815 * Similarly, if C2,P2 are in their initial state (again a frequent
1816 * occurrence), the result is a no-op
1817 * (C1 | P1) & (<everything>) = C1 | P1
1818 *
1819 * Inverted, we have
1820 * (C1 | P1) & ~(C2 | P2) = (C1 | P1) & (~C2 & ~P2)
1821 * = (C1 & (~C2 & ~P2)) | (P1 & (~C2 & ~P2))
1822 * <= (C1 & ~C2) | (P1 & ~P2)
1823 * */
1aa99e6b 1824
765e6ecf 1825 if ((and_with_flags & ANYOF_INVERT)
71068078 1826 && ! is_ANYOF_SYNTHETIC(and_with))
a0dd4231 1827 {
cdd87c1d 1828 unsigned int i;
8951c461 1829
cdd87c1d
KW
1830 ssc_intersection(ssc,
1831 anded_cp_list,
1832 FALSE /* Has already been inverted */
1833 );
c6b76537 1834
cdd87c1d
KW
1835 /* If either P1 or P2 is empty, the intersection will be also; can skip
1836 * the loop */
765e6ecf 1837 if (! (and_with_flags & ANYOF_MATCHES_POSIXL)) {
cdd87c1d
KW
1838 ANYOF_POSIXL_ZERO(ssc);
1839 }
e0e1be5f 1840 else if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
cdd87c1d
KW
1841
1842 /* Note that the Posix class component P from 'and_with' actually
1843 * looks like:
1844 * P = Pa | Pb | ... | Pn
1845 * where each component is one posix class, such as in [\w\s].
1846 * Thus
1847 * ~P = ~(Pa | Pb | ... | Pn)
1848 * = ~Pa & ~Pb & ... & ~Pn
1849 * <= ~Pa | ~Pb | ... | ~Pn
1850 * The last is something we can easily calculate, but unfortunately
1851 * is likely to have many false positives. We could do better
1852 * in some (but certainly not all) instances if two classes in
1853 * P have known relationships. For example
1854 * :lower: <= :alpha: <= :alnum: <= \w <= :graph: <= :print:
1855 * So
1856 * :lower: & :print: = :lower:
1857 * And similarly for classes that must be disjoint. For example,
1858 * since \s and \w can have no elements in common based on rules in
1859 * the POSIX standard,
1860 * \w & ^\S = nothing
1861 * Unfortunately, some vendor locales do not meet the Posix
1862 * standard, in particular almost everything by Microsoft.
1863 * The loop below just changes e.g., \w into \W and vice versa */
1864
1ee208c4 1865 regnode_charclass_posixl temp;
cdd87c1d
KW
1866 int add = 1; /* To calculate the index of the complement */
1867
b1234259 1868 Zero(&temp, 1, regnode_charclass_posixl);
cdd87c1d
KW
1869 ANYOF_POSIXL_ZERO(&temp);
1870 for (i = 0; i < ANYOF_MAX; i++) {
1871 assert(i % 2 != 0
7dcac5f6
KW
1872 || ! ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i)
1873 || ! ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i + 1));
cdd87c1d 1874
7dcac5f6 1875 if (ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i)) {
cdd87c1d
KW
1876 ANYOF_POSIXL_SET(&temp, i + add);
1877 }
1878 add = 0 - add; /* 1 goes to -1; -1 goes to 1 */
1879 }
1880 ANYOF_POSIXL_AND(&temp, ssc);
c6b76537 1881
cdd87c1d
KW
1882 } /* else ssc already has no posixes */
1883 } /* else: Not inverted. This routine is a no-op if 'and_with' is an SSC
1884 in its initial state */
71068078 1885 else if (! is_ANYOF_SYNTHETIC(and_with)
7dcac5f6 1886 || ! ssc_is_cp_posixl_init(pRExC_state, (regnode_ssc *)and_with))
cdd87c1d
KW
1887 {
1888 /* But if 'ssc' is in its initial state, the result is just 'and_with';
1889 * copy it over 'ssc' */
1890 if (ssc_is_cp_posixl_init(pRExC_state, ssc)) {
71068078 1891 if (is_ANYOF_SYNTHETIC(and_with)) {
cdd87c1d
KW
1892 StructCopy(and_with, ssc, regnode_ssc);
1893 }
1894 else {
1895 ssc->invlist = anded_cp_list;
1896 ANYOF_POSIXL_ZERO(ssc);
765e6ecf 1897 if (and_with_flags & ANYOF_MATCHES_POSIXL) {
7dcac5f6 1898 ANYOF_POSIXL_OR((regnode_charclass_posixl*) and_with, ssc);
cdd87c1d
KW
1899 }
1900 }
1901 }
e0e1be5f 1902 else if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)
765e6ecf 1903 || (and_with_flags & ANYOF_MATCHES_POSIXL))
cdd87c1d
KW
1904 {
1905 /* One or the other of P1, P2 is non-empty. */
765e6ecf 1906 if (and_with_flags & ANYOF_MATCHES_POSIXL) {
1ea8b7fe
KW
1907 ANYOF_POSIXL_AND((regnode_charclass_posixl*) and_with, ssc);
1908 }
cdd87c1d
KW
1909 ssc_union(ssc, anded_cp_list, FALSE);
1910 }
1911 else { /* P1 = P2 = empty */
1912 ssc_intersection(ssc, anded_cp_list, FALSE);
1913 }
137165a6 1914 }
653099ff
GS
1915}
1916
653099ff 1917STATIC void
cdd87c1d 1918S_ssc_or(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc,
7dcac5f6 1919 const regnode_charclass *or_with)
653099ff 1920{
cdd87c1d
KW
1921 /* Accumulate into SSC 'ssc' its 'OR' with 'or_with', which is either
1922 * another SSC or a regular ANYOF class. Can create false positives if
1923 * 'or_with' is to be inverted. */
7918f24d 1924
a0dd4231
KW
1925 SV* ored_cp_list;
1926 U8 ored_flags;
2d5613be 1927 U8 or_with_flags = inRANGE(OP(or_with), ANYOFH, ANYOFRb)
29a889ef
KW
1928 ? 0
1929 : ANYOF_FLAGS(or_with);
c6b76537 1930
cdd87c1d 1931 PERL_ARGS_ASSERT_SSC_OR;
c6b76537 1932
71068078 1933 assert(is_ANYOF_SYNTHETIC(ssc));
a0dd4231
KW
1934
1935 /* 'or_with' is used as-is if it too is an SSC; otherwise have to extract
1936 * the code point inversion list and just the relevant flags */
71068078 1937 if (is_ANYOF_SYNTHETIC(or_with)) {
7dcac5f6 1938 ored_cp_list = ((regnode_ssc*) or_with)->invlist;
765e6ecf 1939 ored_flags = or_with_flags;
a0dd4231
KW
1940 }
1941 else {
5c0f85ef 1942 ored_cp_list = get_ANYOF_cp_list_for_ssc(pRExC_state, or_with);
765e6ecf 1943 ored_flags = or_with_flags & ANYOF_COMMON_FLAGS;
f240c685
KW
1944 if (OP(or_with) != ANYOFD) {
1945 ored_flags
765e6ecf 1946 |= or_with_flags
108316fb
KW
1947 & ( ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER
1948 |ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP);
765e6ecf 1949 if (ANYOFL_UTF8_LOCALE_REQD(or_with_flags)) {
d1c40ef5
KW
1950 ored_flags |=
1951 ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
1952 }
f240c685 1953 }
a0dd4231
KW
1954 }
1955
1956 ANYOF_FLAGS(ssc) |= ored_flags;
cdd87c1d
KW
1957
1958 /* Below, C1 is the list of code points in 'ssc'; P1, its posix classes.
1959 * C2 is the list of code points in 'or-with'; P2, its posix classes.
1960 * 'or_with' may be inverted. When not inverted, we have the simple
1961 * situation of computing:
1962 * (C1 | P1) | (C2 | P2) = (C1 | C2) | (P1 | P2)
1963 * If P1|P2 yields a situation with both a class and its complement are
1964 * set, like having both \w and \W, this matches all code points, and we
1965 * can delete these from the P component of the ssc going forward. XXX We
1966 * might be able to delete all the P components, but I (khw) am not certain
1967 * about this, and it is better to be safe.
1968 *
1969 * Inverted, we have
1970 * (C1 | P1) | ~(C2 | P2) = (C1 | P1) | (~C2 & ~P2)
1971 * <= (C1 | P1) | ~C2
1972 * <= (C1 | ~C2) | P1
1973 * (which results in actually simpler code than the non-inverted case)
1974 * */
9826f543 1975
765e6ecf 1976 if ((or_with_flags & ANYOF_INVERT)
71068078 1977 && ! is_ANYOF_SYNTHETIC(or_with))
a0dd4231 1978 {
cdd87c1d 1979 /* We ignore P2, leaving P1 going forward */
1ea8b7fe 1980 } /* else Not inverted */
765e6ecf 1981 else if (or_with_flags & ANYOF_MATCHES_POSIXL) {
7dcac5f6 1982 ANYOF_POSIXL_OR((regnode_charclass_posixl*)or_with, ssc);
e0e1be5f 1983 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
cdd87c1d
KW
1984 unsigned int i;
1985 for (i = 0; i < ANYOF_MAX; i += 2) {
1986 if (ANYOF_POSIXL_TEST(ssc, i) && ANYOF_POSIXL_TEST(ssc, i + 1))
1987 {
1988 ssc_match_all_cp(ssc);
1989 ANYOF_POSIXL_CLEAR(ssc, i);
1990 ANYOF_POSIXL_CLEAR(ssc, i+1);
cdd87c1d
KW
1991 }
1992 }
1993 }
1aa99e6b 1994 }
cdd87c1d
KW
1995
1996 ssc_union(ssc,
1997 ored_cp_list,
1998 FALSE /* Already has been inverted */
1999 );
653099ff
GS
2000}
2001
b423522f
KW
2002PERL_STATIC_INLINE void
2003S_ssc_union(pTHX_ regnode_ssc *ssc, SV* const invlist, const bool invert2nd)
2004{
2005 PERL_ARGS_ASSERT_SSC_UNION;
2006
71068078 2007 assert(is_ANYOF_SYNTHETIC(ssc));
b423522f
KW
2008
2009 _invlist_union_maybe_complement_2nd(ssc->invlist,
2010 invlist,
2011 invert2nd,
2012 &ssc->invlist);
2013}
2014
2015PERL_STATIC_INLINE void
2016S_ssc_intersection(pTHX_ regnode_ssc *ssc,
2017 SV* const invlist,
2018 const bool invert2nd)
2019{
2020 PERL_ARGS_ASSERT_SSC_INTERSECTION;
2021
71068078 2022 assert(is_ANYOF_SYNTHETIC(ssc));
b423522f
KW
2023
2024 _invlist_intersection_maybe_complement_2nd(ssc->invlist,
2025 invlist,
2026 invert2nd,
2027 &ssc->invlist);
2028}
2029
2030PERL_STATIC_INLINE void
2031S_ssc_add_range(pTHX_ regnode_ssc *ssc, const UV start, const UV end)
2032{
2033 PERL_ARGS_ASSERT_SSC_ADD_RANGE;
2034
71068078 2035 assert(is_ANYOF_SYNTHETIC(ssc));
b423522f
KW
2036
2037 ssc->invlist = _add_range_to_invlist(ssc->invlist, start, end);
2038}
2039
2040PERL_STATIC_INLINE void
2041S_ssc_cp_and(pTHX_ regnode_ssc *ssc, const UV cp)
2042{
2043 /* AND just the single code point 'cp' into the SSC 'ssc' */
2044
2045 SV* cp_list = _new_invlist(2);
2046
2047 PERL_ARGS_ASSERT_SSC_CP_AND;
2048
71068078 2049 assert(is_ANYOF_SYNTHETIC(ssc));
b423522f
KW
2050
2051 cp_list = add_cp_to_invlist(cp_list, cp);
2052 ssc_intersection(ssc, cp_list,
2053 FALSE /* Not inverted */
2054 );
2055 SvREFCNT_dec_NN(cp_list);
2056}
2057
2058PERL_STATIC_INLINE void
dc3bf405 2059S_ssc_clear_locale(regnode_ssc *ssc)
b423522f
KW
2060{
2061 /* Set the SSC 'ssc' to not match any locale things */
b423522f
KW
2062 PERL_ARGS_ASSERT_SSC_CLEAR_LOCALE;
2063
71068078 2064 assert(is_ANYOF_SYNTHETIC(ssc));
b423522f
KW
2065
2066 ANYOF_POSIXL_ZERO(ssc);
2067 ANYOF_FLAGS(ssc) &= ~ANYOF_LOCALE_FLAGS;
2068}
2069
b35552de
KW
2070#define NON_OTHER_COUNT NON_OTHER_COUNT_FOR_USE_ONLY_BY_REGCOMP_DOT_C
2071
2072STATIC bool
2073S_is_ssc_worth_it(const RExC_state_t * pRExC_state, const regnode_ssc * ssc)
2074{
2075 /* The synthetic start class is used to hopefully quickly winnow down
2076 * places where a pattern could start a match in the target string. If it
2077 * doesn't really narrow things down that much, there isn't much point to
2078 * having the overhead of using it. This function uses some very crude
2079 * heuristics to decide if to use the ssc or not.
2080 *
2081 * It returns TRUE if 'ssc' rules out more than half what it considers to
2082 * be the "likely" possible matches, but of course it doesn't know what the
2083 * actual things being matched are going to be; these are only guesses
2084 *
2085 * For /l matches, it assumes that the only likely matches are going to be
2086 * in the 0-255 range, uniformly distributed, so half of that is 127
2087 * For /a and /d matches, it assumes that the likely matches will be just
2088 * the ASCII range, so half of that is 63
2089 * For /u and there isn't anything matching above the Latin1 range, it
2090 * assumes that that is the only range likely to be matched, and uses
2091 * half that as the cut-off: 127. If anything matches above Latin1,
2092 * it assumes that all of Unicode could match (uniformly), except for
2093 * non-Unicode code points and things in the General Category "Other"
2094 * (unassigned, private use, surrogates, controls and formats). This
2095 * is a much large number. */
2096
b35552de
KW
2097 U32 count = 0; /* Running total of number of code points matched by
2098 'ssc' */
2099 UV start, end; /* Start and end points of current range in inversion
26be5fe6 2100 XXX outdated. UTF-8 locales are common, what about invert? list */
72400949
KW
2101 const U32 max_code_points = (LOC)
2102 ? 256
f6e8f31e 2103 : (( ! UNI_SEMANTICS
1a26cbcb 2104 || invlist_highest(ssc->invlist) < 256)
72400949
KW
2105 ? 128
2106 : NON_OTHER_COUNT);
2107 const U32 max_match = max_code_points / 2;
b35552de
KW
2108
2109 PERL_ARGS_ASSERT_IS_SSC_WORTH_IT;
2110
2111 invlist_iterinit(ssc->invlist);
2112 while (invlist_iternext(ssc->invlist, &start, &end)) {
72400949
KW
2113 if (start >= max_code_points) {
2114 break;
b35552de 2115 }
72400949 2116 end = MIN(end, max_code_points - 1);
b35552de 2117 count += end - start + 1;
72400949 2118 if (count >= max_match) {
b35552de
KW
2119 invlist_iterfinish(ssc->invlist);
2120 return FALSE;
2121 }
2122 }
2123
2124 return TRUE;
2125}
2126
2127
b423522f
KW
2128STATIC void
2129S_ssc_finalize(pTHX_ RExC_state_t *pRExC_state, regnode_ssc *ssc)
2130{
2131 /* The inversion list in the SSC is marked mortal; now we need a more
2132 * permanent copy, which is stored the same way that is done in a regular
dcb20b36
KW
2133 * ANYOF node, with the first NUM_ANYOF_CODE_POINTS code points in a bit
2134 * map */
b423522f 2135
28118b9c 2136 SV* invlist = invlist_clone(ssc->invlist, NULL);
b423522f
KW
2137
2138 PERL_ARGS_ASSERT_SSC_FINALIZE;
2139
71068078 2140 assert(is_ANYOF_SYNTHETIC(ssc));
b423522f 2141
a0dd4231 2142 /* The code in this file assumes that all but these flags aren't relevant
93e92956
KW
2143 * to the SSC, except SSC_MATCHES_EMPTY_STRING, which should be cleared
2144 * by the time we reach here */
f240c685
KW
2145 assert(! (ANYOF_FLAGS(ssc)
2146 & ~( ANYOF_COMMON_FLAGS
108316fb
KW
2147 |ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER
2148 |ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP)));
a0dd4231 2149
b423522f
KW
2150 populate_ANYOF_from_invlist( (regnode *) ssc, &invlist);
2151
4c404f26 2152 set_ANYOF_arg(pRExC_state, (regnode *) ssc, invlist, NULL, NULL);
21c3fd9d 2153 SvREFCNT_dec(invlist);
b423522f 2154
85c8e306
KW
2155 /* Make sure is clone-safe */
2156 ssc->invlist = NULL;
2157
e0e1be5f 2158 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
93e92956 2159 ANYOF_FLAGS(ssc) |= ANYOF_MATCHES_POSIXL;
d156f5cb 2160 OP(ssc) = ANYOFPOSIXL;
e0e1be5f 2161 }
d156f5cb 2162 else if (RExC_contains_locale) {
b2e90ddf
KW
2163 OP(ssc) = ANYOFL;
2164 }
2165
1462525b 2166 assert(! (ANYOF_FLAGS(ssc) & ANYOF_LOCALE_FLAGS) || RExC_contains_locale);
b423522f
KW
2167}
2168
a3621e74
YO
2169#define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
2170#define TRIE_LIST_CUR(state) ( TRIE_LIST_ITEM( state, 0 ).forid )
2171#define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
538e84ed
KW
2172#define TRIE_LIST_USED(idx) ( trie->states[state].trans.list \
2173 ? (TRIE_LIST_CUR( idx ) - 1) \
2174 : 0 )
a3621e74 2175
3dab1dad
YO
2176
2177#ifdef DEBUGGING
07be1b83 2178/*
2b8b4781
NC
2179 dump_trie(trie,widecharmap,revcharmap)
2180 dump_trie_interim_list(trie,widecharmap,revcharmap,next_alloc)
2181 dump_trie_interim_table(trie,widecharmap,revcharmap,next_alloc)
3dab1dad
YO
2182
2183 These routines dump out a trie in a somewhat readable format.
07be1b83
YO
2184 The _interim_ variants are used for debugging the interim
2185 tables that are used to generate the final compressed
2186 representation which is what dump_trie expects.
2187
486ec47a 2188 Part of the reason for their existence is to provide a form
3dab1dad 2189 of documentation as to how the different representations function.
07be1b83
YO
2190
2191*/
3dab1dad
YO
2192
2193/*
3dab1dad
YO
2194 Dumps the final compressed table form of the trie to Perl_debug_log.
2195 Used for debugging make_trie().
2196*/
b9a59e08 2197
3dab1dad 2198STATIC void
2b8b4781
NC
2199S_dump_trie(pTHX_ const struct _reg_trie_data *trie, HV *widecharmap,
2200 AV *revcharmap, U32 depth)
3dab1dad
YO
2201{
2202 U32 state;
ab3bbdeb 2203 SV *sv=sv_newmortal();
55eed653 2204 int colwidth= widecharmap ? 6 : 4;
2e64971a 2205 U16 word;
3dab1dad
YO
2206 GET_RE_DEBUG_FLAGS_DECL;
2207
7918f24d 2208 PERL_ARGS_ASSERT_DUMP_TRIE;
ab3bbdeb 2209
6ad9a8ab 2210 Perl_re_indentf( aTHX_ "Char : %-6s%-6s%-4s ",
1e37780e 2211 depth+1, "Match","Base","Ofs" );
3dab1dad
YO
2212
2213 for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
2b8b4781 2214 SV ** const tmp = av_fetch( revcharmap, state, 0);
3dab1dad 2215 if ( tmp ) {
6ad9a8ab 2216 Perl_re_printf( aTHX_ "%*s",
ab3bbdeb 2217 colwidth,
538e84ed 2218 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
ab3bbdeb
YO
2219 PL_colors[0], PL_colors[1],
2220 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
538e84ed
KW
2221 PERL_PV_ESCAPE_FIRSTCHAR
2222 )
ab3bbdeb 2223 );
3dab1dad
YO
2224 }
2225 }
1e37780e
YO
2226 Perl_re_printf( aTHX_ "\n");
2227 Perl_re_indentf( aTHX_ "State|-----------------------", depth+1);
3dab1dad
YO
2228
2229 for( state = 0 ; state < trie->uniquecharcount ; state++ )
6ad9a8ab
YO
2230 Perl_re_printf( aTHX_ "%.*s", colwidth, "--------");
2231 Perl_re_printf( aTHX_ "\n");
3dab1dad 2232
1e2e3d02 2233 for( state = 1 ; state < trie->statecount ; state++ ) {
be8e71aa 2234 const U32 base = trie->states[ state ].trans.base;
3dab1dad 2235
147e3846 2236 Perl_re_indentf( aTHX_ "#%4" UVXf "|", depth+1, (UV)state);
3dab1dad
YO
2237
2238 if ( trie->states[ state ].wordnum ) {
1e37780e 2239 Perl_re_printf( aTHX_ " W%4X", trie->states[ state ].wordnum );
3dab1dad 2240 } else {
6ad9a8ab 2241 Perl_re_printf( aTHX_ "%6s", "" );
3dab1dad
YO
2242 }
2243
147e3846 2244 Perl_re_printf( aTHX_ " @%4" UVXf " ", (UV)base );
3dab1dad
YO
2245
2246 if ( base ) {
2247 U32 ofs = 0;
2248
2249 while( ( base + ofs < trie->uniquecharcount ) ||
2250 ( base + ofs - trie->uniquecharcount < trie->lasttrans
538e84ed
KW
2251 && trie->trans[ base + ofs - trie->uniquecharcount ].check
2252 != state))
3dab1dad
YO
2253 ofs++;
2254
147e3846 2255 Perl_re_printf( aTHX_ "+%2" UVXf "[ ", (UV)ofs);
3dab1dad
YO
2256
2257 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
538e84ed
KW
2258 if ( ( base + ofs >= trie->uniquecharcount )
2259 && ( base + ofs - trie->uniquecharcount
2260 < trie->lasttrans )
2261 && trie->trans[ base + ofs
2262 - trie->uniquecharcount ].check == state )
3dab1dad 2263 {
147e3846 2264 Perl_re_printf( aTHX_ "%*" UVXf, colwidth,
1e37780e
YO
2265 (UV)trie->trans[ base + ofs - trie->uniquecharcount ].next
2266 );
3dab1dad 2267 } else {
88f063b4 2268 Perl_re_printf( aTHX_ "%*s", colwidth," ." );
3dab1dad
YO
2269 }
2270 }
2271
6ad9a8ab 2272 Perl_re_printf( aTHX_ "]");
3dab1dad
YO
2273
2274 }
6ad9a8ab 2275 Perl_re_printf( aTHX_ "\n" );
3dab1dad 2276 }
6ad9a8ab 2277 Perl_re_indentf( aTHX_ "word_info N:(prev,len)=",
cb41e5d6 2278 depth);
2e64971a 2279 for (word=1; word <= trie->wordcount; word++) {
6ad9a8ab 2280 Perl_re_printf( aTHX_ " %d:(%d,%d)",
2e64971a
DM
2281 (int)word, (int)(trie->wordinfo[word].prev),
2282 (int)(trie->wordinfo[word].len));
2283 }
6ad9a8ab 2284 Perl_re_printf( aTHX_ "\n" );
538e84ed 2285}
3dab1dad 2286/*
3dab1dad 2287 Dumps a fully constructed but uncompressed trie in list form.
538e84ed 2288 List tries normally only are used for construction when the number of
3dab1dad
YO
2289 possible chars (trie->uniquecharcount) is very high.
2290 Used for debugging make_trie().
2291*/
2292STATIC void
55eed653 2293S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie,
2b8b4781
NC
2294 HV *widecharmap, AV *revcharmap, U32 next_alloc,
2295 U32 depth)
3dab1dad
YO
2296{
2297 U32 state;
ab3bbdeb 2298 SV *sv=sv_newmortal();
55eed653 2299 int colwidth= widecharmap ? 6 : 4;
3dab1dad 2300 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
2301
2302 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_LIST;
2303
3dab1dad 2304 /* print out the table precompression. */
6ad9a8ab 2305 Perl_re_indentf( aTHX_ "State :Word | Transition Data\n",
cb41e5d6 2306 depth+1 );
6ad9a8ab 2307 Perl_re_indentf( aTHX_ "%s",
cb41e5d6 2308 depth+1, "------:-----+-----------------\n" );
538e84ed 2309
3dab1dad
YO
2310 for( state=1 ; state < next_alloc ; state ++ ) {
2311 U16 charid;
538e84ed 2312
147e3846 2313 Perl_re_indentf( aTHX_ " %4" UVXf " :",
cb41e5d6 2314 depth+1, (UV)state );
3dab1dad 2315 if ( ! trie->states[ state ].wordnum ) {
6ad9a8ab 2316 Perl_re_printf( aTHX_ "%5s| ","");
3dab1dad 2317 } else {
6ad9a8ab 2318 Perl_re_printf( aTHX_ "W%4x| ",
3dab1dad
YO
2319 trie->states[ state ].wordnum
2320 );
2321 }
2322 for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
538e84ed 2323 SV ** const tmp = av_fetch( revcharmap,
88f063b4 2324 TRIE_LIST_ITEM(state, charid).forid, 0);
ab3bbdeb 2325 if ( tmp ) {
147e3846 2326 Perl_re_printf( aTHX_ "%*s:%3X=%4" UVXf " | ",
ab3bbdeb 2327 colwidth,
538e84ed
KW
2328 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp),
2329 colwidth,
2330 PL_colors[0], PL_colors[1],
2331 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0)
2332 | PERL_PV_ESCAPE_FIRSTCHAR
ab3bbdeb 2333 ) ,
88f063b4
KW
2334 TRIE_LIST_ITEM(state, charid).forid,
2335 (UV)TRIE_LIST_ITEM(state, charid).newstate
1e2e3d02 2336 );
538e84ed 2337 if (!(charid % 10))
6ad9a8ab 2338 Perl_re_printf( aTHX_ "\n%*s| ",
664e119d 2339 (int)((depth * 2) + 14), "");
1e2e3d02 2340 }
ab3bbdeb 2341 }
6ad9a8ab 2342 Perl_re_printf( aTHX_ "\n");
3dab1dad 2343 }
538e84ed 2344}
3dab1dad
YO
2345
2346/*
3dab1dad 2347 Dumps a fully constructed but uncompressed trie in table form.
538e84ed
KW
2348 This is the normal DFA style state transition table, with a few
2349 twists to facilitate compression later.
3dab1dad
YO
2350 Used for debugging make_trie().
2351*/
2352STATIC void
55eed653 2353S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie,
2b8b4781
NC
2354 HV *widecharmap, AV *revcharmap, U32 next_alloc,
2355 U32 depth)
3dab1dad
YO
2356{
2357 U32 state;
2358 U16 charid;
ab3bbdeb 2359 SV *sv=sv_newmortal();
55eed653 2360 int colwidth= widecharmap ? 6 : 4;
3dab1dad 2361 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
2362
2363 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_TABLE;
538e84ed 2364
3dab1dad
YO
2365 /*
2366 print out the table precompression so that we can do a visual check
2367 that they are identical.
2368 */
538e84ed 2369
6ad9a8ab 2370 Perl_re_indentf( aTHX_ "Char : ", depth+1 );
3dab1dad
YO
2371
2372 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
2b8b4781 2373 SV ** const tmp = av_fetch( revcharmap, charid, 0);
3dab1dad 2374 if ( tmp ) {
6ad9a8ab 2375 Perl_re_printf( aTHX_ "%*s",
ab3bbdeb 2376 colwidth,
538e84ed 2377 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
ab3bbdeb
YO
2378 PL_colors[0], PL_colors[1],
2379 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
538e84ed
KW
2380 PERL_PV_ESCAPE_FIRSTCHAR
2381 )
ab3bbdeb 2382 );
3dab1dad
YO
2383 }
2384 }
2385
4aaafc03
YO
2386 Perl_re_printf( aTHX_ "\n");
2387 Perl_re_indentf( aTHX_ "State+-", depth+1 );
3dab1dad
YO
2388
2389 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
6ad9a8ab 2390 Perl_re_printf( aTHX_ "%.*s", colwidth,"--------");
3dab1dad
YO
2391 }
2392
6ad9a8ab 2393 Perl_re_printf( aTHX_ "\n" );
3dab1dad
YO
2394
2395 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
2396
147e3846 2397 Perl_re_indentf( aTHX_ "%4" UVXf " : ",
cb41e5d6 2398 depth+1,
3dab1dad
YO
2399 (UV)TRIE_NODENUM( state ) );
2400
2401 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
ab3bbdeb
YO
2402 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
2403 if (v)
147e3846 2404 Perl_re_printf( aTHX_ "%*" UVXf, colwidth, v );
ab3bbdeb 2405 else
6ad9a8ab 2406 Perl_re_printf( aTHX_ "%*s", colwidth, "." );
3dab1dad
YO
2407 }
2408 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
147e3846 2409 Perl_re_printf( aTHX_ " (%4" UVXf ")\n",
538e84ed 2410 (UV)trie->trans[ state ].check );
3dab1dad 2411 } else {
147e3846 2412 Perl_re_printf( aTHX_ " (%4" UVXf ") W%4X\n",
538e84ed 2413 (UV)trie->trans[ state ].check,
3dab1dad
YO
2414 trie->states[ TRIE_NODENUM( state ) ].wordnum );
2415 }
2416 }
07be1b83 2417}
3dab1dad
YO
2418
2419#endif
2420
2e64971a 2421
786e8c11
YO
2422/* make_trie(startbranch,first,last,tail,word_count,flags,depth)
2423 startbranch: the first branch in the whole branch sequence
2424 first : start branch of sequence of branch-exact nodes.
2425 May be the same as startbranch
2426 last : Thing following the last branch.
2427 May be the same as tail.
2428 tail : item following the branch sequence
2429 count : words in the sequence
a4525e78 2430 flags : currently the OP() type we will be building one of /EXACT(|F|FA|FU|FU_SS|L|FLU8)/
786e8c11 2431 depth : indent depth
3dab1dad 2432
786e8c11 2433Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
07be1b83 2434
786e8c11
YO
2435A trie is an N'ary tree where the branches are determined by digital
2436decomposition of the key. IE, at the root node you look up the 1st character and
2437follow that branch repeat until you find the end of the branches. Nodes can be
2438marked as "accepting" meaning they represent a complete word. Eg:
07be1b83 2439
786e8c11 2440 /he|she|his|hers/
72f13be8 2441
786e8c11
YO
2442would convert into the following structure. Numbers represent states, letters
2443following numbers represent valid transitions on the letter from that state, if
2444the number is in square brackets it represents an accepting state, otherwise it
2445will be in parenthesis.
07be1b83 2446
786e8c11
YO
2447 +-h->+-e->[3]-+-r->(8)-+-s->[9]
2448 | |
2449 | (2)
2450 | |
2451 (1) +-i->(6)-+-s->[7]
2452 |
2453 +-s->(3)-+-h->(4)-+-e->[5]
07be1b83 2454
786e8c11
YO
2455 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
2456
2457This shows that when matching against the string 'hers' we will begin at state 1
2458read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
2459then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
2460is also accepting. Thus we know that we can match both 'he' and 'hers' with a
2461single traverse. We store a mapping from accepting to state to which word was
2462matched, and then when we have multiple possibilities we try to complete the
b8fda935 2463rest of the regex in the order in which they occurred in the alternation.
786e8c11
YO
2464
2465The only prior NFA like behaviour that would be changed by the TRIE support is
2466the silent ignoring of duplicate alternations which are of the form:
2467
2468 / (DUPE|DUPE) X? (?{ ... }) Y /x
2469
4b714af6 2470Thus EVAL blocks following a trie may be called a different number of times with
786e8c11 2471and without the optimisation. With the optimisations dupes will be silently
486ec47a 2472ignored. This inconsistent behaviour of EVAL type nodes is well established as
786e8c11
YO
2473the following demonstrates:
2474
2475 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
2476
2477which prints out 'word' three times, but
2478
2479 'words'=~/(word|word|word)(?{ print $1 })S/
2480
2481which doesnt print it out at all. This is due to other optimisations kicking in.
2482
2483Example of what happens on a structural level:
2484
486ec47a 2485The regexp /(ac|ad|ab)+/ will produce the following debug output:
786e8c11
YO
2486
2487 1: CURLYM[1] {1,32767}(18)
2488 5: BRANCH(8)
2489 6: EXACT <ac>(16)
2490 8: BRANCH(11)
2491 9: EXACT <ad>(16)
2492 11: BRANCH(14)
2493 12: EXACT <ab>(16)
2494 16: SUCCEED(0)
2495 17: NOTHING(18)
2496 18: END(0)
2497
2498This would be optimizable with startbranch=5, first=5, last=16, tail=16
2499and should turn into:
2500
2501 1: CURLYM[1] {1,32767}(18)
2502 5: TRIE(16)
2503 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
2504 <ac>
2505 <ad>
2506 <ab>
2507 16: SUCCEED(0)
2508 17: NOTHING(18)
2509 18: END(0)
2510
2511Cases where tail != last would be like /(?foo|bar)baz/:
2512
2513 1: BRANCH(4)
2514 2: EXACT <foo>(8)
2515 4: BRANCH(7)
2516 5: EXACT <bar>(8)
2517 7: TAIL(8)
2518 8: EXACT <baz>(10)
2519 10: END(0)
2520
2521which would be optimizable with startbranch=1, first=1, last=7, tail=8
2522and would end up looking like:
2523
2524 1: TRIE(8)
2525 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
2526 <foo>
2527 <bar>
2528 7: TAIL(8)
2529 8: EXACT <baz>(10)
2530 10: END(0)
2531
c80e42f3 2532 d = uvchr_to_utf8_flags(d, uv, 0);
786e8c11
YO
2533
2534is the recommended Unicode-aware way of saying
2535
2536 *(d++) = uv;
2537*/
2538
fab2782b 2539#define TRIE_STORE_REVCHAR(val) \
786e8c11 2540 STMT_START { \
73031816 2541 if (UTF) { \
668fcfea 2542 SV *zlopp = newSV(UTF8_MAXBYTES); \
88c9ea1e 2543 unsigned char *flrbbbbb = (unsigned char *) SvPVX(zlopp); \
1d84a256
DM
2544 unsigned char *const kapow = uvchr_to_utf8(flrbbbbb, val); \
2545 *kapow = '\0'; \
73031816
NC
2546 SvCUR_set(zlopp, kapow - flrbbbbb); \
2547 SvPOK_on(zlopp); \
2548 SvUTF8_on(zlopp); \
2549 av_push(revcharmap, zlopp); \
2550 } else { \
fab2782b 2551 char ooooff = (char)val; \
73031816
NC
2552 av_push(revcharmap, newSVpvn(&ooooff, 1)); \
2553 } \
2554 } STMT_END
786e8c11 2555
914a25d5
KW
2556/* This gets the next character from the input, folding it if not already
2557 * folded. */
2558#define TRIE_READ_CHAR STMT_START { \
2559 wordlen++; \
2560 if ( UTF ) { \
2561 /* if it is UTF then it is either already folded, or does not need \
2562 * folding */ \
1c1d615a 2563 uvc = valid_utf8_to_uvchr( (const U8*) uc, &len); \
914a25d5
KW
2564 } \
2565 else if (folder == PL_fold_latin1) { \
7d006b13
KW
2566 /* This folder implies Unicode rules, which in the range expressible \
2567 * by not UTF is the lower case, with the two exceptions, one of \
2568 * which should have been taken care of before calling this */ \
2569 assert(*uc != LATIN_SMALL_LETTER_SHARP_S); \
2570 uvc = toLOWER_L1(*uc); \
2571 if (UNLIKELY(uvc == MICRO_SIGN)) uvc = GREEK_SMALL_LETTER_MU; \
2572 len = 1; \
914a25d5
KW
2573 } else { \
2574 /* raw data, will be folded later if needed */ \
2575 uvc = (U32)*uc; \
2576 len = 1; \
2577 } \
786e8c11
YO
2578} STMT_END
2579
2580
2581
2582#define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
2583 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
00195859 2584 U32 ging = TRIE_LIST_LEN( state ) * 2; \
f9003953 2585 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
00195859 2586 TRIE_LIST_LEN( state ) = ging; \
786e8c11
YO
2587 } \
2588 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
2589 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
2590 TRIE_LIST_CUR( state )++; \
2591} STMT_END
07be1b83 2592
786e8c11 2593#define TRIE_LIST_NEW(state) STMT_START { \
d09f14bf 2594 Newx( trie->states[ state ].trans.list, \
786e8c11
YO
2595 4, reg_trie_trans_le ); \
2596 TRIE_LIST_CUR( state ) = 1; \
2597 TRIE_LIST_LEN( state ) = 4; \
2598} STMT_END
07be1b83 2599
786e8c11
YO
2600#define TRIE_HANDLE_WORD(state) STMT_START { \
2601 U16 dupe= trie->states[ state ].wordnum; \
2602 regnode * const noper_next = regnext( noper ); \
2603 \
786e8c11
YO
2604 DEBUG_r({ \
2605 /* store the word for dumping */ \
2606 SV* tmp; \
2607 if (OP(noper) != NOTHING) \
740cce10 2608 tmp = newSVpvn_utf8(STRING(noper), STR_LEN(noper), UTF); \
786e8c11 2609 else \
740cce10 2610 tmp = newSVpvn_utf8( "", 0, UTF ); \
2b8b4781 2611 av_push( trie_words, tmp ); \
786e8c11
YO
2612 }); \
2613 \
2614 curword++; \
2e64971a
DM
2615 trie->wordinfo[curword].prev = 0; \
2616 trie->wordinfo[curword].len = wordlen; \
2617 trie->wordinfo[curword].accept = state; \
786e8c11
YO
2618 \
2619 if ( noper_next < tail ) { \
2620 if (!trie->jump) \
538e84ed
KW
2621 trie->jump = (U16 *) PerlMemShared_calloc( word_count + 1, \
2622 sizeof(U16) ); \
7f69552c 2623 trie->jump[curword] = (U16)(noper_next - convert); \
786e8c11
YO
2624 if (!jumper) \
2625 jumper = noper_next; \
2626 if (!nextbranch) \
2627 nextbranch= regnext(cur); \
2628 } \
2629 \
2630 if ( dupe ) { \
2e64971a
DM
2631 /* It's a dupe. Pre-insert into the wordinfo[].prev */\
2632 /* chain, so that when the bits of chain are later */\
2633 /* linked together, the dups appear in the chain */\
2634 trie->wordinfo[curword].prev = trie->wordinfo[dupe].prev; \
2635 trie->wordinfo[dupe].prev = curword; \
786e8c11
YO
2636 } else { \
2637 /* we haven't inserted this word yet. */ \
2638 trie->states[ state ].wordnum = curword; \
2639 } \
2640} STMT_END
07be1b83 2641
3dab1dad 2642
786e8c11
YO
2643#define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
2644 ( ( base + charid >= ucharcount \
2645 && base + charid < ubound \
2646 && state == trie->trans[ base - ucharcount + charid ].check \
2647 && trie->trans[ base - ucharcount + charid ].next ) \
2648 ? trie->trans[ base - ucharcount + charid ].next \
2649 : ( state==1 ? special : 0 ) \
2650 )
3dab1dad 2651
8bcafbf4
YO
2652#define TRIE_BITMAP_SET_FOLDED(trie, uvc, folder) \
2653STMT_START { \
2654 TRIE_BITMAP_SET(trie, uvc); \
2655 /* store the folded codepoint */ \
2656 if ( folder ) \
2657 TRIE_BITMAP_SET(trie, folder[(U8) uvc ]); \
2658 \
2659 if ( !UTF ) { \
2660 /* store first byte of utf8 representation of */ \
2661 /* variant codepoints */ \
2662 if (! UVCHR_IS_INVARIANT(uvc)) { \
2663 TRIE_BITMAP_SET(trie, UTF8_TWO_BYTE_HI(uvc)); \
2664 } \
2665 } \
2666} STMT_END
786e8c11
YO
2667#define MADE_TRIE 1
2668#define MADE_JUMP_TRIE 2
2669#define MADE_EXACT_TRIE 4
3dab1dad 2670
a3621e74 2671STATIC I32
538e84ed
KW
2672S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch,
2673 regnode *first, regnode *last, regnode *tail,
2674 U32 word_count, U32 flags, U32 depth)
a3621e74
YO
2675{
2676 /* first pass, loop through and scan words */
2677 reg_trie_data *trie;
55eed653 2678 HV *widecharmap = NULL;
2b8b4781 2679 AV *revcharmap = newAV();
a3621e74 2680 regnode *cur;
a3621e74
YO
2681 STRLEN len = 0;
2682 UV uvc = 0;
2683 U16 curword = 0;
2684 U32 next_alloc = 0;
786e8c11
YO
2685 regnode *jumper = NULL;
2686 regnode *nextbranch = NULL;
7f69552c 2687 regnode *convert = NULL;
2e64971a 2688 U32 *prev_states; /* temp array mapping each state to previous one */
a3621e74 2689 /* we just use folder as a flag in utf8 */
1e696034 2690 const U8 * folder = NULL;
a3621e74 2691
3a611511
YO
2692 /* in the below add_data call we are storing either 'tu' or 'tuaa'
2693 * which stands for one trie structure, one hash, optionally followed
2694 * by two arrays */
2b8b4781 2695#ifdef DEBUGGING
3a611511 2696 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("tuaa"));
2b8b4781
NC
2697 AV *trie_words = NULL;
2698 /* along with revcharmap, this only used during construction but both are
2699 * useful during debugging so we store them in the struct when debugging.
8e11feef 2700 */
2b8b4781 2701#else
cf78de0b 2702 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("tu"));
3dab1dad 2703 STRLEN trie_charcount=0;
3dab1dad 2704#endif
2b8b4781 2705 SV *re_trie_maxbuff;
a3621e74 2706 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
2707
2708 PERL_ARGS_ASSERT_MAKE_TRIE;
72f13be8
YO
2709#ifndef DEBUGGING
2710 PERL_UNUSED_ARG(depth);
2711#endif
a3621e74 2712
1e696034 2713 switch (flags) {
3f2416ae 2714 case EXACT: case EXACT_REQ8: case EXACTL: break;
89829bb5 2715 case EXACTFAA:
0ea669f4 2716 case EXACTFUP:
a4525e78
KW
2717 case EXACTFU:
2718 case EXACTFLU8: folder = PL_fold_latin1; break;
1e696034 2719 case EXACTF: folder = PL_fold; break;
fab2782b 2720 default: Perl_croak( aTHX_ "panic! In trie construction, unknown node type %u %s", (unsigned) flags, PL_reg_name[flags] );
1e696034
KW
2721 }
2722
c944940b 2723 trie = (reg_trie_data *) PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
a3621e74 2724 trie->refcount = 1;
3dab1dad 2725 trie->startstate = 1;
786e8c11 2726 trie->wordcount = word_count;
f8fc2ecf 2727 RExC_rxi->data->data[ data_slot ] = (void*)trie;
c944940b 2728 trie->charmap = (U16 *) PerlMemShared_calloc( 256, sizeof(U16) );
3f2416ae 2729 if (flags == EXACT || flags == EXACT_REQ8 || flags == EXACTL)
c944940b 2730 trie->bitmap = (char *) PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
2e64971a
DM
2731 trie->wordinfo = (reg_trie_wordinfo *) PerlMemShared_calloc(
2732 trie->wordcount+1, sizeof(reg_trie_wordinfo));
2733
a3621e74 2734 DEBUG_r({
2b8b4781 2735 trie_words = newAV();
a3621e74 2736 });
a3621e74 2737
3b89859a 2738 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, GV_ADD);
316ebaf2 2739 assert(re_trie_maxbuff);
a3621e74 2740 if (!SvIOK(re_trie_maxbuff)) {
0111c4fd 2741 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
a3621e74 2742 }
df826430 2743 DEBUG_TRIE_COMPILE_r({
6ad9a8ab 2744 Perl_re_indentf( aTHX_
cb41e5d6
YO
2745 "make_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
2746 depth+1,
88f063b4 2747 REG_NODE_NUM(startbranch), REG_NODE_NUM(first),
538e84ed 2748 REG_NODE_NUM(last), REG_NODE_NUM(tail), (int)depth);
3dab1dad 2749 });
538e84ed 2750
7f69552c
YO
2751 /* Find the node we are going to overwrite */
2752 if ( first == startbranch && OP( last ) != BRANCH ) {
2753 /* whole branch chain */
2754 convert = first;
2755 } else {
2756 /* branch sub-chain */
2757 convert = NEXTOPER( first );
2758 }
538e84ed 2759
a3621e74
YO
2760 /* -- First loop and Setup --
2761
2762 We first traverse the branches and scan each word to determine if it
2763 contains widechars, and how many unique chars there are, this is
2764 important as we have to build a table with at least as many columns as we
2765 have unique chars.
2766
2767 We use an array of integers to represent the character codes 0..255
538e84ed
KW
2768 (trie->charmap) and we use a an HV* to store Unicode characters. We use
2769 the native representation of the character value as the key and IV's for
2770 the coded index.
a3621e74
YO
2771
2772 *TODO* If we keep track of how many times each character is used we can
2773 remap the columns so that the table compression later on is more
3b753521 2774 efficient in terms of memory by ensuring the most common value is in the
a3621e74
YO
2775 middle and the least common are on the outside. IMO this would be better
2776 than a most to least common mapping as theres a decent chance the most
2777 common letter will share a node with the least common, meaning the node
486ec47a 2778 will not be compressible. With a middle is most common approach the worst
a3621e74
YO
2779 case is when we have the least common nodes twice.
2780
2781 */
2782
a3621e74 2783 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
df826430 2784 regnode *noper = NEXTOPER( cur );
944e05e3
YO
2785 const U8 *uc;
2786 const U8 *e;
bc031a7d 2787 int foldlen = 0;
07be1b83 2788 U32 wordlen = 0; /* required init */
bc031a7d
KW
2789 STRLEN minchars = 0;
2790 STRLEN maxchars = 0;
538e84ed
KW
2791 bool set_bit = trie->bitmap ? 1 : 0; /*store the first char in the
2792 bitmap?*/
a3621e74 2793
3dab1dad 2794 if (OP(noper) == NOTHING) {
20ed8c88
YO
2795 /* skip past a NOTHING at the start of an alternation
2796 * eg, /(?:)a|(?:b)/ should be the same as /a|b/
ca902fb8
YO
2797 *
2798 * If the next node is not something we are supposed to process
2799 * we will just ignore it due to the condition guarding the
2800 * next block.
20ed8c88 2801 */
ca902fb8 2802
df826430 2803 regnode *noper_next= regnext(noper);
944e05e3
YO
2804 if (noper_next < tail)
2805 noper= noper_next;
2806 }
2807
f6b4b99d
KW
2808 if ( noper < tail
2809 && ( OP(noper) == flags
3f2416ae
KW
2810 || (flags == EXACT && OP(noper) == EXACT_REQ8)
2811 || (flags == EXACTFU && ( OP(noper) == EXACTFU_REQ8
0ea669f4 2812 || OP(noper) == EXACTFUP))))
f6b4b99d 2813 {
944e05e3
YO
2814 uc= (U8*)STRING(noper);
2815 e= uc + STR_LEN(noper);
2816 } else {
2817 trie->minlen= 0;
2818 continue;
3dab1dad 2819 }
df826430 2820
944e05e3 2821
fab2782b 2822 if ( set_bit ) { /* bitmap only alloced when !(UTF&&Folding) */
02daf0ab
YO
2823 TRIE_BITMAP_SET(trie,*uc); /* store the raw first byte
2824 regardless of encoding */
0ea669f4 2825 if (OP( noper ) == EXACTFUP) {
fab2782b 2826 /* false positives are ok, so just set this */
0dc4a61d 2827 TRIE_BITMAP_SET(trie, LATIN_SMALL_LETTER_SHARP_S);
fab2782b
YO
2828 }
2829 }
dca5fc4c 2830
bc031a7d
KW
2831 for ( ; uc < e ; uc += len ) { /* Look at each char in the current
2832 branch */
3dab1dad 2833 TRIE_CHARCOUNT(trie)++;
a3621e74 2834 TRIE_READ_CHAR;
645de4ce 2835
bc031a7d
KW
2836 /* TRIE_READ_CHAR returns the current character, or its fold if /i
2837 * is in effect. Under /i, this character can match itself, or
2838 * anything that folds to it. If not under /i, it can match just
2839 * itself. Most folds are 1-1, for example k, K, and KELVIN SIGN
2840 * all fold to k, and all are single characters. But some folds
2841 * expand to more than one character, so for example LATIN SMALL
2842 * LIGATURE FFI folds to the three character sequence 'ffi'. If
2843 * the string beginning at 'uc' is 'ffi', it could be matched by
2844 * three characters, or just by the one ligature character. (It
2845 * could also be matched by two characters: LATIN SMALL LIGATURE FF
2846 * followed by 'i', or by 'f' followed by LATIN SMALL LIGATURE FI).
2847 * (Of course 'I' and/or 'F' instead of 'i' and 'f' can also
2848 * match.) The trie needs to know the minimum and maximum number
2849 * of characters that could match so that it can use size alone to
2850 * quickly reject many match attempts. The max is simple: it is
2851 * the number of folded characters in this branch (since a fold is
2852 * never shorter than what folds to it. */
2853
2854 maxchars++;
2855
2856 /* And the min is equal to the max if not under /i (indicated by
2857 * 'folder' being NULL), or there are no multi-character folds. If
2858 * there is a multi-character fold, the min is incremented just
2859 * once, for the character that folds to the sequence. Each
2860 * character in the sequence needs to be added to the list below of
2861 * characters in the trie, but we count only the first towards the
2862 * min number of characters needed. This is done through the
2863 * variable 'foldlen', which is returned by the macros that look
2864 * for these sequences as the number of bytes the sequence
2865 * occupies. Each time through the loop, we decrement 'foldlen' by
2866 * how many bytes the current char occupies. Only when it reaches
2867 * 0 do we increment 'minchars' or look for another multi-character
2868 * sequence. */
2869 if (folder == NULL) {
2870 minchars++;
2871 }
2872 else if (foldlen > 0) {
2873 foldlen -= (UTF) ? UTF8SKIP(uc) : 1;
645de4ce
KW
2874 }
2875 else {
bc031a7d
KW
2876 minchars++;
2877
2878 /* See if *uc is the beginning of a multi-character fold. If
2879 * so, we decrement the length remaining to look at, to account
2880 * for the current character this iteration. (We can use 'uc'
2881 * instead of the fold returned by TRIE_READ_CHAR because for
2882 * non-UTF, the latin1_safe macro is smart enough to account
2883 * for all the unfolded characters, and because for UTF, the
2884 * string will already have been folded earlier in the
2885 * compilation process */
2886 if (UTF) {
2887 if ((foldlen = is_MULTI_CHAR_FOLD_utf8_safe(uc, e))) {
2888 foldlen -= UTF8SKIP(uc);
645de4ce
KW
2889 }
2890 }
bc031a7d
KW
2891 else if ((foldlen = is_MULTI_CHAR_FOLD_latin1_safe(uc, e))) {
2892 foldlen--;
2893 }
645de4ce 2894 }
bc031a7d
KW
2895
2896 /* The current character (and any potential folds) should be added
2897 * to the possible matching characters for this position in this
2898 * branch */
a3621e74 2899 if ( uvc < 256 ) {
fab2782b
YO
2900 if ( folder ) {
2901 U8 folded= folder[ (U8) uvc ];
2902 if ( !trie->charmap[ folded ] ) {
2903 trie->charmap[ folded ]=( ++trie->uniquecharcount );
2904 TRIE_STORE_REVCHAR( folded );
2905 }
2906 }
a3621e74
YO
2907 if ( !trie->charmap[ uvc ] ) {
2908 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
fab2782b 2909 TRIE_STORE_REVCHAR( uvc );
a3621e74 2910 }
02daf0ab 2911 if ( set_bit ) {
62012aee
KW
2912 /* store the codepoint in the bitmap, and its folded
2913 * equivalent. */
8bcafbf4 2914 TRIE_BITMAP_SET_FOLDED(trie, uvc, folder);
02daf0ab
YO
2915 set_bit = 0; /* We've done our bit :-) */
2916 }
a3621e74 2917 } else {
bc031a7d
KW
2918
2919 /* XXX We could come up with the list of code points that fold
2920 * to this using PL_utf8_foldclosures, except not for
2921 * multi-char folds, as there may be multiple combinations
2922 * there that could work, which needs to wait until runtime to
2923 * resolve (The comment about LIGATURE FFI above is such an
2924 * example */
2925
a3621e74 2926 SV** svpp;
55eed653
NC
2927 if ( !widecharmap )
2928 widecharmap = newHV();
a3621e74 2929
55eed653 2930 svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );
a3621e74
YO
2931
2932 if ( !svpp )
147e3846 2933 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%" UVXf, uvc );
a3621e74
YO
2934
2935 if ( !SvTRUE( *svpp ) ) {
2936 sv_setiv( *svpp, ++trie->uniquecharcount );
fab2782b 2937 TRIE_STORE_REVCHAR(uvc);
a3621e74
YO
2938 }
2939 }
bc031a7d
KW
2940 } /* end loop through characters in this branch of the trie */
2941
2942 /* We take the min and max for this branch and combine to find the min
2943 * and max for all branches processed so far */
3dab1dad 2944 if( cur == first ) {
bc031a7d
KW
2945 trie->minlen = minchars;
2946 trie->maxlen = maxchars;
2947 } else if (minchars < trie->minlen) {
2948 trie->minlen = minchars;
2949 } else if (maxchars > trie->maxlen) {
2950 trie->maxlen = maxchars;
fab2782b 2951 }
a3621e74
YO
2952 } /* end first pass */
2953 DEBUG_TRIE_COMPILE_r(
6ad9a8ab 2954 Perl_re_indentf( aTHX_
cb41e5d6
YO
2955 "TRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
2956 depth+1,
55eed653 2957 ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
be8e71aa
YO
2958 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
2959 (int)trie->minlen, (int)trie->maxlen )
a3621e74 2960 );
a3621e74
YO
2961
2962 /*
2963 We now know what we are dealing with in terms of unique chars and
2964 string sizes so we can calculate how much memory a naive
0111c4fd
RGS
2965 representation using a flat table will take. If it's over a reasonable
2966 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
a3621e74
YO
2967 conservative but potentially much slower representation using an array
2968 of lists.
2969
2970 At the end we convert both representations into the same compressed
2971 form that will be used in regexec.c for matching with. The latter
2972 is a form that cannot be used to construct with but has memory
2973 properties similar to the list form and access properties similar
2974 to the table form making it both suitable for fast searches and
2975 small enough that its feasable to store for the duration of a program.
2976
2977 See the comment in the code where the compressed table is produced
2978 inplace from the flat tabe representation for an explanation of how
2979 the compression works.
2980
2981 */
2982
2983
2e64971a
DM
2984 Newx(prev_states, TRIE_CHARCOUNT(trie) + 2, U32);
2985 prev_states[1] = 0;
2986
538e84ed
KW
2987 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1)
2988 > SvIV(re_trie_maxbuff) )
2989 {
a3621e74
YO
2990 /*
2991 Second Pass -- Array Of Lists Representation
2992
2993 Each state will be represented by a list of charid:state records
2994 (reg_trie_trans_le) the first such element holds the CUR and LEN
2995 points of the allocated array. (See defines above).
2996
2997 We build the initial structure using the lists, and then convert
2998 it into the compressed table form which allows faster lookups
2999 (but cant be modified once converted).
a3621e74
YO
3000 */
3001
a3621e74
YO
3002 STRLEN transcount = 1;
3003
6ad9a8ab 3004 DEBUG_TRIE_COMPILE_MORE_r( Perl_re_indentf( aTHX_ "Compiling trie using list compiler\n",
cb41e5d6 3005 depth+1));
686b73d4 3006
c944940b
JH
3007 trie->states = (reg_trie_state *)
3008 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
3009 sizeof(reg_trie_state) );
a3621e74
YO
3010 TRIE_LIST_NEW(1);
3011 next_alloc = 2;
3012
3013 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
3014
df826430 3015 regnode *noper = NEXTOPER( cur );
c445ea15
AL
3016 U32 state = 1; /* required init */
3017 U16 charid = 0; /* sanity init */
07be1b83 3018 U32 wordlen = 0; /* required init */
c445ea15 3019
df826430
YO
3020 if (OP(noper) == NOTHING) {
3021 regnode *noper_next= regnext(noper);
944e05e3
YO
3022 if (noper_next < tail)
3023 noper= noper_next;
ca902fb8
YO
3024 /* we will undo this assignment if noper does not
3025 * point at a trieable type in the else clause of
3026 * the following statement. */
df826430
YO
3027 }
3028
f6b4b99d
KW
3029 if ( noper < tail
3030 && ( OP(noper) == flags
3f2416ae
KW
3031 || (flags == EXACT && OP(noper) == EXACT_REQ8)
3032 || (flags == EXACTFU && ( OP(noper) == EXACTFU_REQ8
0ea669f4 3033 || OP(noper) == EXACTFUP))))
f6b4b99d 3034 {
944e05e3
YO
3035 const U8 *uc= (U8*)STRING(noper);
3036 const U8 *e= uc + STR_LEN(noper);
3037
786e8c11 3038 for ( ; uc < e ; uc += len ) {
c445ea15 3039
786e8c11 3040 TRIE_READ_CHAR;
c445ea15 3041
786e8c11
YO
3042 if ( uvc < 256 ) {
3043 charid = trie->charmap[ uvc ];
c445ea15 3044 } else {
538e84ed
KW
3045 SV** const svpp = hv_fetch( widecharmap,
3046 (char*)&uvc,
3047 sizeof( UV ),
3048 0);
786e8c11
YO
3049 if ( !svpp ) {
3050 charid = 0;
3051 } else {
3052 charid=(U16)SvIV( *svpp );
3053 }
c445ea15 3054 }
538e84ed
KW
3055 /* charid is now 0 if we dont know the char read, or
3056 * nonzero if we do */
786e8c11 3057 if ( charid ) {
a3621e74 3058
786e8c11
YO
3059 U16 check;
3060 U32 newstate = 0;
a3621e74 3061
786e8c11
YO
3062 charid--;
3063 if ( !trie->states[ state ].trans.list ) {
3064 TRIE_LIST_NEW( state );
c445ea15 3065 }
538e84ed
KW
3066 for ( check = 1;
3067 check <= TRIE_LIST_USED( state );
3068 check++ )
3069 {
3070 if ( TRIE_LIST_ITEM( state, check ).forid
3071 == charid )
3072 {
786e8c11
YO
3073 newstate = TRIE_LIST_ITEM( state, check ).newstate;
3074 break;
3075 }
3076 }
3077 if ( ! newstate ) {
3078 newstate = next_alloc++;
2e64971a 3079 prev_states[newstate] = state;
786e8c11
YO
3080 TRIE_LIST_PUSH( state, charid, newstate );
3081 transcount++;
3082 }
3083 state = newstate;
3084 } else {
147e3846 3085 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %" IVdf, uvc );
c445ea15 3086 }
a28509cc 3087 }
ca902fb8
YO
3088 } else {
3089 /* If we end up here it is because we skipped past a NOTHING, but did not end up
3090 * on a trieable type. So we need to reset noper back to point at the first regop
3091 * in the branch before we call TRIE_HANDLE_WORD()
3092 */
3093 noper= NEXTOPER(cur);
3094 }
3dab1dad 3095 TRIE_HANDLE_WORD(state);
a3621e74
YO
3096
3097 } /* end second pass */
3098
1e2e3d02 3099 /* next alloc is the NEXT state to be allocated */
538e84ed 3100 trie->statecount = next_alloc;
c944940b
JH
3101 trie->states = (reg_trie_state *)
3102 PerlMemShared_realloc( trie->states,
3103 next_alloc
3104 * sizeof(reg_trie_state) );
a3621e74 3105
3dab1dad 3106 /* and now dump it out before we compress it */
2b8b4781
NC
3107 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
3108 revcharmap, next_alloc,
3109 depth+1)
1e2e3d02 3110 );
a3621e74 3111
c944940b
JH
3112 trie->trans = (reg_trie_trans *)
3113 PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
a3621e74
YO
3114 {
3115 U32 state;
a3621e74
YO
3116 U32 tp = 0;
3117 U32 zp = 0;
3118
3119
3120 for( state=1 ; state < next_alloc ; state ++ ) {
3121 U32 base=0;
3122
3123 /*
3124 DEBUG_TRIE_COMPILE_MORE_r(
6ad9a8ab 3125 Perl_re_printf( aTHX_ "tp: %d zp: %d ",tp,zp)
a3621e74
YO
3126 );
3127 */
3128
3129 if (trie->states[state].trans.list) {
3130 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
3131 U16 maxid=minid;
a28509cc 3132 U16 idx;
a3621e74
YO
3133
3134 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
c445ea15
AL
3135 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
3136 if ( forid < minid ) {
3137 minid=forid;
3138 } else if ( forid > maxid ) {
3139 maxid=forid;
3140 }
a3621e74
YO
3141 }
3142 if ( transcount < tp + maxid - minid + 1) {
3143 transcount *= 2;
c944940b
JH
3144 trie->trans = (reg_trie_trans *)
3145 PerlMemShared_realloc( trie->trans,
446bd890
NC
3146 transcount
3147 * sizeof(reg_trie_trans) );
538e84ed
KW
3148 Zero( trie->trans + (transcount / 2),
3149 transcount / 2,
3150 reg_trie_trans );
a3621e74
YO
3151 }
3152 base = trie->uniquecharcount + tp - minid;
3153 if ( maxid == minid ) {
3154 U32 set = 0;
3155 for ( ; zp < tp ; zp++ ) {
3156 if ( ! trie->trans[ zp ].next ) {
3157 base = trie->uniquecharcount + zp - minid;
538e84ed
KW
3158 trie->trans[ zp ].next = TRIE_LIST_ITEM( state,
3159 1).newstate;
a3621e74
YO
3160 trie->trans[ zp ].check = state;
3161 set = 1;
3162 break;
3163 }
3164 }
3165 if ( !set ) {
538e84ed
KW
3166 trie->trans[ tp ].next = TRIE_LIST_ITEM( state,
3167 1).newstate;
a3621e74
YO
3168 trie->trans[ tp ].check = state;
3169 tp++;
3170 zp = tp;
3171 }
3172 } else {
3173 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
538e84ed
KW
3174 const U32 tid = base
3175 - trie->uniquecharcount
3176 + TRIE_LIST_ITEM( state, idx ).forid;
3177 trie->trans[ tid ].next = TRIE_LIST_ITEM( state,
3178 idx ).newstate;
a3621e74
YO
3179 trie->trans[ tid ].check = state;
3180 }
3181 tp += ( maxid - minid + 1 );
3182 }
3183 Safefree(trie->states[ state ].trans.list);
3184 }
3185 /*
3186 DEBUG_TRIE_COMPILE_MORE_r(
6ad9a8ab 3187 Perl_re_printf( aTHX_ " base: %d\n",base);
a3621e74
YO
3188 );
3189 */
3190 trie->states[ state ].trans.base=base;
3191 }
cc601c31 3192 trie->lasttrans = tp + 1;
a3621e74
YO
3193 }
3194 } else {
3195 /*
3196 Second Pass -- Flat Table Representation.
3197
b423522f
KW
3198 we dont use the 0 slot of either trans[] or states[] so we add 1 to
3199 each. We know that we will need Charcount+1 trans at most to store
3200 the data (one row per char at worst case) So we preallocate both
3201 structures assuming worst case.
a3621e74
YO
3202
3203 We then construct the trie using only the .next slots of the entry
3204 structs.
3205
b423522f
KW
3206 We use the .check field of the first entry of the node temporarily
3207 to make compression both faster and easier by keeping track of how
3208 many non zero fields are in the node.
a3621e74
YO
3209
3210 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
3211 transition.
3212
b423522f
KW
3213 There are two terms at use here: state as a TRIE_NODEIDX() which is
3214 a number representing the first entry of the node, and state as a
3215 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1)
3216 and TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3)
3217 if there are 2 entrys per node. eg:
a3621e74
YO
3218
3219 A B A B
3220 1. 2 4 1. 3 7
3221 2. 0 3 3. 0 5
3222 3. 0 0 5. 0 0
3223 4. 0 0 7. 0 0
3224
b423522f
KW
3225 The table is internally in the right hand, idx form. However as we
3226 also have to deal with the states array which is indexed by nodenum
3227 we have to use TRIE_NODENUM() to convert.
a3621e74
YO
3228
3229 */
6ad9a8ab 3230 DEBUG_TRIE_COMPILE_MORE_r( Perl_re_indentf( aTHX_ "Compiling trie using table compiler\n",
cb41e5d6 3231 depth+1));
3dab1dad 3232
c944940b
JH
3233 trie->trans = (reg_trie_trans *)
3234 PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
3235 * trie->uniquecharcount + 1,
3236 sizeof(reg_trie_trans) );
3237 trie->states = (reg_trie_state *)
3238 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
3239 sizeof(reg_trie_state) );
a3621e74
YO
3240 next_alloc = trie->uniquecharcount + 1;
3241
3dab1dad 3242
a3621e74
YO
3243 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
3244
df826430 3245 regnode *noper = NEXTOPER( cur );
a3621e74
YO
3246
3247 U32 state = 1; /* required init */
3248
3249 U16 charid = 0; /* sanity init */
3250 U32 accept_state = 0; /* sanity init */
a3621e74 3251
07be1b83 3252 U32 wordlen = 0; /* required init */
a3621e74 3253
df826430
YO
3254 if (OP(noper) == NOTHING) {
3255 regnode *noper_next= regnext(noper);
944e05e3
YO
3256 if (noper_next < tail)
3257 noper= noper_next;
ca902fb8
YO
3258 /* we will undo this assignment if noper does not
3259 * point at a trieable type in the else clause of
3260 * the following statement. */
df826430 3261 }
fab2782b 3262
f6b4b99d
KW
3263 if ( noper < tail
3264 && ( OP(noper) == flags
3f2416ae
KW
3265 || (flags == EXACT && OP(noper) == EXACT_REQ8)
3266 || (flags == EXACTFU && ( OP(noper) == EXACTFU_REQ8
0ea669f4 3267 || OP(noper) == EXACTFUP))))
f6b4b99d 3268 {
944e05e3
YO
3269 const U8 *uc= (U8*)STRING(noper);
3270 const U8 *e= uc + STR_LEN(noper);
3271
786e8c11 3272 for ( ; uc < e ; uc += len ) {
a3621e74 3273
786e8c11 3274 TRIE_READ_CHAR;
a3621e74 3275
786e8c11
YO
3276 if ( uvc < 256 ) {
3277 charid = trie->charmap[ uvc ];
3278 } else {
538e84ed
KW
3279 SV* const * const svpp = hv_fetch( widecharmap,
3280 (char*)&uvc,
3281 sizeof( UV ),
3282 0);
786e8c11 3283 charid = svpp ? (U16)SvIV(*svpp) : 0;
a3621e74 3284 }
786e8c11
YO
3285 if ( charid ) {
3286 charid--;
3287 if ( !trie->trans[ state + charid ].next ) {
3288 trie->trans[ state + charid ].next = next_alloc;
3289 trie->trans[ state ].check++;
2e64971a
DM
3290 prev_states[TRIE_NODENUM(next_alloc)]
3291 = TRIE_NODENUM(state);
786e8c11
YO
3292 next_alloc += trie->uniquecharcount;
3293 }
3294 state = trie->trans[ state + charid ].next;
3295 } else {
147e3846 3296 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %" IVdf, uvc );
786e8c11 3297 }
538e84ed
KW
3298 /* charid is now 0 if we dont know the char read, or
3299 * nonzero if we do */
a3621e74 3300 }
ca902fb8
YO
3301 } else {
3302 /* If we end up here it is because we skipped past a NOTHING, but did not end up
3303 * on a trieable type. So we need to reset noper back to point at the first regop
3304 * in the branch before we call TRIE_HANDLE_WORD().
3305 */
3306 noper= NEXTOPER(cur);
a3621e74 3307 }
3dab1dad
YO
3308 accept_state = TRIE_NODENUM( state );
3309 TRIE_HANDLE_WORD(accept_state);
a3621e74
YO
3310
3311 } /* end second pass */
3312
3dab1dad 3313 /* and now dump it out before we compress it */
2b8b4781
NC
3314 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
3315 revcharmap,
3316 next_alloc, depth+1));
a3621e74 3317
a3621e74
YO
3318 {
3319 /*
3320 * Inplace compress the table.*
3321
3322 For sparse data sets the table constructed by the trie algorithm will
3323 be mostly 0/FAIL transitions or to put it another way mostly empty.
3324 (Note that leaf nodes will not contain any transitions.)
3325
3326 This algorithm compresses the tables by eliminating most such
3327 transitions, at the cost of a modest bit of extra work during lookup:
3328
3329 - Each states[] entry contains a .base field which indicates the
3330 index in the state[] array wheres its transition data is stored.
3331
3b753521 3332 - If .base is 0 there are no valid transitions from that node.
a3621e74
YO
3333
3334 - If .base is nonzero then charid is added to it to find an entry in
3335 the trans array.
3336
3337 -If trans[states[state].base+charid].check!=state then the
3338 transition is taken to be a 0/Fail transition. Thus if there are fail
3339 transitions at the front of the node then the .base offset will point
3340 somewhere inside the previous nodes data (or maybe even into a node
3341 even earlier), but the .check field determines if the transition is
3342 valid.
3343
786e8c11 3344 XXX - wrong maybe?
a3621e74 3345 The following process inplace converts the table to the compressed
3b753521 3346 table: We first do not compress the root node 1,and mark all its
a3621e74 3347 .check pointers as 1 and set its .base pointer as 1 as well. This
3b753521
FN
3348 allows us to do a DFA construction from the compressed table later,
3349 and ensures that any .base pointers we calculate later are greater
3350 than 0.
a3621e74
YO
3351
3352 - We set 'pos' to indicate the first entry of the second node.
3353
3354 - We then iterate over the columns of the node, finding the first and
3355 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
3356 and set the .check pointers accordingly, and advance pos
3357 appropriately and repreat for the next node. Note that when we copy
3358 the next pointers we have to convert them from the original
3359 NODEIDX form to NODENUM form as the former is not valid post
3360 compression.
3361
3362 - If a node has no transitions used we mark its base as 0 and do not
3363 advance the pos pointer.
3364
3365 - If a node only has one transition we use a second pointer into the
3366 structure to fill in allocated fail transitions from other states.
3367 This pointer is independent of the main pointer and scans forward
3368 looking for null transitions that are allocated to a state. When it
3369 finds one it writes the single transition into the "hole". If the
786e8c11 3370 pointer doesnt find one the single transition is appended as normal.
a3621e74
YO
3371
3372 - Once compressed we can Renew/realloc the structures to release the
3373 excess space.
3374
3375 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
3376 specifically Fig 3.47 and the associated pseudocode.
3377
3378 demq
3379 */
a3b680e6 3380 const U32 laststate = TRIE_NODENUM( next_alloc );
a28509cc 3381 U32 state, charid;
a3621e74 3382 U32 pos = 0, zp=0;
1e2e3d02 3383 trie->statecount = laststate;
a3621e74
YO
3384
3385 for ( state = 1 ; state < laststate ; state++ ) {
3386 U8 flag = 0;
a28509cc
AL
3387 const U32 stateidx = TRIE_NODEIDX( state );
3388 const U32 o_used = trie->trans[ stateidx ].check;
3389 U32 used = trie->trans[ stateidx ].check;
a3621e74
YO
3390 trie->trans[ stateidx ].check = 0;
3391
538e84ed
KW
3392 for ( charid = 0;
3393 used && charid < trie->uniquecharcount;
3394 charid++ )
3395 {
a3621e74
YO
3396 if ( flag || trie->trans[ stateidx + charid ].next ) {
3397 if ( trie->trans[ stateidx + charid ].next ) {
3398 if (o_used == 1) {
3399 for ( ; zp < pos ; zp++ ) {
3400 if ( ! trie->trans[ zp ].next ) {
3401 break;
3402 }
3403 }
538e84ed
KW
3404 trie->states[ state ].trans.base
3405 = zp
3406 + trie->uniquecharcount
3407 - charid ;
3408 trie->trans[ zp ].next
3409 = SAFE_TRIE_NODENUM( trie->trans[ stateidx
3410 + charid ].next );
a3621e74
YO
3411 trie->trans[ zp ].check = state;
3412 if ( ++zp > pos ) pos = zp;