This is a live mirror of the Perl 5 development currently hosted at https://github.com/perl/perl5
regcomp.c: Defer setting regnode operand
[perl5.git] / regcomp.c
CommitLineData
a0d0e21e
LW
1/* regcomp.c
2 */
3
4/*
4ac71550
TC
5 * 'A fair jaw-cracker dwarf-language must be.' --Samwise Gamgee
6 *
7 * [p.285 of _The Lord of the Rings_, II/iii: "The Ring Goes South"]
a0d0e21e
LW
8 */
9
61296642
DM
10/* This file contains functions for compiling a regular expression. See
11 * also regexec.c which funnily enough, contains functions for executing
166f8a29 12 * a regular expression.
e4a054ea
DM
13 *
14 * This file is also copied at build time to ext/re/re_comp.c, where
15 * it's built with -DPERL_EXT_RE_BUILD -DPERL_EXT_RE_DEBUG -DPERL_EXT.
16 * This causes the main functions to be compiled under new names and with
17 * debugging support added, which makes "use re 'debug'" work.
166f8a29
DM
18 */
19
a687059c
LW
20/* NOTE: this is derived from Henry Spencer's regexp code, and should not
21 * confused with the original package (see point 3 below). Thanks, Henry!
22 */
23
24/* Additional note: this code is very heavily munged from Henry's version
25 * in places. In some spots I've traded clarity for efficiency, so don't
26 * blame Henry for some of the lack of readability.
27 */
28
e50aee73 29/* The names of the functions have been changed from regcomp and
3b753521 30 * regexec to pregcomp and pregexec in order to avoid conflicts
e50aee73
AD
31 * with the POSIX routines of the same names.
32*/
33
b9d5759e 34#ifdef PERL_EXT_RE_BUILD
54df2634 35#include "re_top.h"
b81d288d 36#endif
56953603 37
a687059c 38/*
e50aee73 39 * pregcomp and pregexec -- regsub and regerror are not used in perl
a687059c
LW
40 *
41 * Copyright (c) 1986 by University of Toronto.
42 * Written by Henry Spencer. Not derived from licensed software.
43 *
44 * Permission is granted to anyone to use this software for any
45 * purpose on any computer system, and to redistribute it freely,
46 * subject to the following restrictions:
47 *
48 * 1. The author is not responsible for the consequences of use of
49 * this software, no matter how awful, even if they arise
50 * from defects in it.
51 *
52 * 2. The origin of this software must not be misrepresented, either
53 * by explicit claim or by omission.
54 *
55 * 3. Altered versions must be plainly marked as such, and must not
56 * be misrepresented as being the original software.
57 *
58 *
59 **** Alterations to Henry's code are...
60 ****
4bb101f2 61 **** Copyright (C) 1991, 1992, 1993, 1994, 1995, 1996, 1997, 1998, 1999,
1129b882
NC
62 **** 2000, 2001, 2002, 2003, 2004, 2005, 2006, 2007, 2008
63 **** by Larry Wall and others
a687059c 64 ****
9ef589d8
LW
65 **** You may distribute under the terms of either the GNU General Public
66 **** License or the Artistic License, as specified in the README file.
67
a687059c
LW
68 *
69 * Beware that some of this code is subtly aware of the way operator
70 * precedence is structured in regular expressions. Serious changes in
71 * regular-expression syntax might require a total rethink.
72 */
73#include "EXTERN.h"
864dbfa3 74#define PERL_IN_REGCOMP_C
a687059c 75#include "perl.h"
d06ea78c 76
acfe0abc 77#ifndef PERL_IN_XSUB_RE
d06ea78c
GS
78# include "INTERN.h"
79#endif
c277df42
IZ
80
81#define REG_COMP_C
54df2634
NC
82#ifdef PERL_IN_XSUB_RE
83# include "re_comp.h"
afda64e8 84EXTERN_C const struct regexp_engine my_reg_engine;
54df2634
NC
85#else
86# include "regcomp.h"
87#endif
a687059c 88
f7e03a10 89#include "dquote_inline.h"
b992490d 90#include "invlist_inline.h"
1b0f46bf 91#include "unicode_constants.h"
04e98a4d 92
538e84ed
KW
93#define HAS_NONLATIN1_FOLD_CLOSURE(i) \
94 _HAS_NONLATIN1_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)
f12c0118
KW
95#define HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(i) \
96 _HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)
26faadbd 97#define IS_NON_FINAL_FOLD(c) _IS_NON_FINAL_FOLD_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
2c61f163 98#define IS_IN_SOME_FOLD_L1(c) _IS_IN_SOME_FOLD_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
94dc5c2d 99
a687059c
LW
100#ifndef STATIC
101#define STATIC static
102#endif
103
3f910e62
YO
104/* this is a chain of data about sub patterns we are processing that
105 need to be handled separately/specially in study_chunk. Its so
106 we can simulate recursion without losing state. */
107struct scan_frame;
108typedef struct scan_frame {
109 regnode *last_regnode; /* last node to process in this frame */
110 regnode *next_regnode; /* next node to process when last is reached */
111 U32 prev_recursed_depth;
112 I32 stopparen; /* what stopparen do we use */
3f910e62
YO
113
114 struct scan_frame *this_prev_frame; /* this previous frame */
115 struct scan_frame *prev_frame; /* previous frame */
116 struct scan_frame *next_frame; /* next frame */
117} scan_frame;
118
f4ae5a27
KW
119/* Certain characters are output as a sequence with the first being a
120 * backslash. */
eb7486de 121#define isBACKSLASHED_PUNCT(c) strchr("-[]\\^", c)
f4ae5a27
KW
122
123
09b2b2e6 124struct RExC_state_t {
514a91f1
DM
125 U32 flags; /* RXf_* are we folding, multilining? */
126 U32 pm_flags; /* PMf_* stuff from the calling PMOP */
830247a4 127 char *precomp; /* uncompiled string. */
711b303b 128 char *precomp_end; /* pointer to end of uncompiled string. */
288b8c02 129 REGEXP *rx_sv; /* The SV that is the regexp. */
f8fc2ecf 130 regexp *rx; /* perl core regexp structure */
538e84ed
KW
131 regexp_internal *rxi; /* internal data for regexp object
132 pprivate field */
fac92740 133 char *start; /* Start of input for compile */
830247a4
IZ
134 char *end; /* End of input for compile */
135 char *parse; /* Input-scan pointer. */
51684e3c
KW
136 char *copy_start; /* start of copy of input within
137 constructed parse string */
138 char *copy_start_in_input; /* Position in input string
139 corresponding to copy_start */
ea3daa5d 140 SSize_t whilem_seen; /* number of WHILEM in this expr */
fac92740 141 regnode *emit_start; /* Start of emitted-code area */
538e84ed
KW
142 regnode *emit_bound; /* First regnode outside of the
143 allocated space */
f55b7b26 144 regnode_offset emit; /* Code-emit pointer */
830247a4
IZ
145 I32 naughty; /* How bad is this pattern? */
146 I32 sawback; /* Did we see \1, ...? */
147 U32 seen;
88f063b4
KW
148 SSize_t size; /* Number of regnode equivalents in
149 pattern */
e7f023ff 150 I32 npar; /* Capture buffer count, (OPEN) plus
538e84ed
KW
151 one. ("par" 0 is the whole
152 pattern)*/
1533d1d2
KW
153 I32 total_par; /* Capture buffer count after parse
154 completed, (OPEN) plus one. ("par" 0
155 is the whole pattern)*/
538e84ed
KW
156 I32 nestroot; /* root parens we are in - used by
157 accept */
830247a4
IZ
158 I32 extralen;
159 I32 seen_zerolen;
f55b7b26
KW
160 regnode_offset *open_parens; /* offsets to open parens */
161 regnode_offset *close_parens; /* offsets to close parens */
d5a00e4a 162 regnode *end_op; /* END node in program */
02daf0ab
YO
163 I32 utf8; /* whether the pattern is utf8 or not */
164 I32 orig_utf8; /* whether the pattern was originally in utf8 */
165 /* XXX use this for future optimisation of case
166 * where pattern must be upgraded to utf8. */
e40e74fe
KW
167 I32 uni_semantics; /* If a d charset modifier should use unicode
168 rules, even if the pattern is not in
169 utf8 */
81714fb9 170 HV *paren_names; /* Paren names */
538e84ed 171
40d049e4 172 regnode **recurse; /* Recurse regops */
88f063b4 173 I32 recurse_count; /* Number of recurse regops we have generated */
4286711a 174 U8 *study_chunk_recursed; /* bitmap of which subs we have moved
538e84ed 175 through */
09a65838 176 U32 study_chunk_recursed_bytes; /* bytes in bitmap */
b57e4118 177 I32 in_lookbehind;
4624b182 178 I32 contains_locale;
bb3f3ed2 179 I32 override_recoding;
b6d67071
KW
180#ifdef EBCDIC
181 I32 recode_x_to_native;
182#endif
9d53c457 183 I32 in_multi_char_class;
1acab4c5 184 struct reg_code_blocks *code_blocks;/* positions of literal (?{})
68e2671b 185 within pattern */
b1603ef8 186 int code_index; /* next code_blocks[] slot */
ee273784 187 SSize_t maxlen; /* mininum possible number of chars in string to match */
3f910e62
YO
188 scan_frame *frame_head;
189 scan_frame *frame_last;
190 U32 frame_count;
7eec73eb 191 AV *warn_text;
dc6d7f5c 192#ifdef ADD_TO_REGEXEC
830247a4
IZ
193 char *starttry; /* -Dr: where regtry was called. */
194#define RExC_starttry (pRExC_state->starttry)
195#endif
d24ca0c5 196 SV *runtime_code_qr; /* qr with the runtime code blocks */
3dab1dad 197#ifdef DEBUGGING
be8e71aa 198 const char *lastparse;
3dab1dad 199 I32 lastnum;
1f1031fe 200 AV *paren_name_list; /* idx -> name */
d9a72fcc 201 U32 study_chunk_recursed_count;
c9f0d54c
YO
202 SV *mysv1;
203 SV *mysv2;
88f063b4 204
3dab1dad
YO
205#define RExC_lastparse (pRExC_state->lastparse)
206#define RExC_lastnum (pRExC_state->lastnum)
1f1031fe 207#define RExC_paren_name_list (pRExC_state->paren_name_list)
d9a72fcc 208#define RExC_study_chunk_recursed_count (pRExC_state->study_chunk_recursed_count)
c9f0d54c
YO
209#define RExC_mysv (pRExC_state->mysv1)
210#define RExC_mysv1 (pRExC_state->mysv1)
211#define RExC_mysv2 (pRExC_state->mysv2)
212
3dab1dad 213#endif
512c0f5a 214 bool seen_unfolded_sharp_s;
911bd04e 215 bool strict;
da7cf1cc 216 bool study_started;
034602eb 217 bool in_script_run;
f55b7b26 218 bool pass1;
09b2b2e6 219};
830247a4 220
e2509266 221#define RExC_flags (pRExC_state->flags)
514a91f1 222#define RExC_pm_flags (pRExC_state->pm_flags)
830247a4 223#define RExC_precomp (pRExC_state->precomp)
51684e3c
KW
224#define RExC_copy_start_in_input (pRExC_state->copy_start_in_input)
225#define RExC_copy_start_in_constructed (pRExC_state->copy_start)
711b303b 226#define RExC_precomp_end (pRExC_state->precomp_end)
288b8c02 227#define RExC_rx_sv (pRExC_state->rx_sv)
830247a4 228#define RExC_rx (pRExC_state->rx)
f8fc2ecf 229#define RExC_rxi (pRExC_state->rxi)
fac92740 230#define RExC_start (pRExC_state->start)
830247a4
IZ
231#define RExC_end (pRExC_state->end)
232#define RExC_parse (pRExC_state->parse)
233#define RExC_whilem_seen (pRExC_state->whilem_seen)
512c0f5a
KW
234
235/* Set during the sizing pass when there is a LATIN SMALL LETTER SHARP S in any
236 * EXACTF node, hence was parsed under /di rules. If later in the parse,
237 * something forces the pattern into using /ui rules, the sharp s should be
238 * folded into the sequence 'ss', which takes up more space than previously
239 * calculated. This means that the sizing pass needs to be restarted. (The
240 * node also becomes an EXACTFU_SS.) For all other characters, an EXACTF node
241 * that gets converted to /ui (and EXACTFU) occupies the same amount of space,
242 * so there is no need to resize [perl #125990]. */
243#define RExC_seen_unfolded_sharp_s (pRExC_state->seen_unfolded_sharp_s)
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)
f55b7b26 250#define RExC_pass1 (pRExC_state->pass1)
fac92740 251#define RExC_emit_start (pRExC_state->emit_start)
3b57cd43 252#define RExC_emit_bound (pRExC_state->emit_bound)
830247a4
IZ
253#define RExC_sawback (pRExC_state->sawback)
254#define RExC_seen (pRExC_state->seen)
255#define RExC_size (pRExC_state->size)
ee273784 256#define RExC_maxlen (pRExC_state->maxlen)
830247a4 257#define RExC_npar (pRExC_state->npar)
1533d1d2 258#define RExC_total_parens (pRExC_state->total_par)
e2e6a0f1 259#define RExC_nestroot (pRExC_state->nestroot)
830247a4
IZ
260#define RExC_extralen (pRExC_state->extralen)
261#define RExC_seen_zerolen (pRExC_state->seen_zerolen)
1aa99e6b 262#define RExC_utf8 (pRExC_state->utf8)
e40e74fe 263#define RExC_uni_semantics (pRExC_state->uni_semantics)
02daf0ab 264#define RExC_orig_utf8 (pRExC_state->orig_utf8)
40d049e4
YO
265#define RExC_open_parens (pRExC_state->open_parens)
266#define RExC_close_parens (pRExC_state->close_parens)
5bd2d46e 267#define RExC_end_op (pRExC_state->end_op)
81714fb9 268#define RExC_paren_names (pRExC_state->paren_names)
40d049e4
YO
269#define RExC_recurse (pRExC_state->recurse)
270#define RExC_recurse_count (pRExC_state->recurse_count)
09a65838 271#define RExC_study_chunk_recursed (pRExC_state->study_chunk_recursed)
538e84ed
KW
272#define RExC_study_chunk_recursed_bytes \
273 (pRExC_state->study_chunk_recursed_bytes)
b57e4118 274#define RExC_in_lookbehind (pRExC_state->in_lookbehind)
4624b182 275#define RExC_contains_locale (pRExC_state->contains_locale)
b6d67071
KW
276#ifdef EBCDIC
277# define RExC_recode_x_to_native (pRExC_state->recode_x_to_native)
278#endif
9d53c457 279#define RExC_in_multi_char_class (pRExC_state->in_multi_char_class)
3f910e62
YO
280#define RExC_frame_head (pRExC_state->frame_head)
281#define RExC_frame_last (pRExC_state->frame_last)
282#define RExC_frame_count (pRExC_state->frame_count)
67cdf558 283#define RExC_strict (pRExC_state->strict)
da7cf1cc 284#define RExC_study_started (pRExC_state->study_started)
7eec73eb 285#define RExC_warn_text (pRExC_state->warn_text)
034602eb 286#define RExC_in_script_run (pRExC_state->in_script_run)
cb9ad9c2 287#define RExC_use_BRANCHJ (!SIZE_ONLY && RExC_extralen)
830247a4 288
99807a43
HS
289/* Heuristic check on the complexity of the pattern: if TOO_NAUGHTY, we set
290 * a flag to disable back-off on the fixed/floating substrings - if it's
291 * a high complexity pattern we assume the benefit of avoiding a full match
292 * is worth the cost of checking for the substrings even if they rarely help.
293 */
294#define RExC_naughty (pRExC_state->naughty)
295#define TOO_NAUGHTY (10)
296#define MARK_NAUGHTY(add) \
297 if (RExC_naughty < TOO_NAUGHTY) \
298 RExC_naughty += (add)
299#define MARK_NAUGHTY_EXP(exp, add) \
300 if (RExC_naughty < TOO_NAUGHTY) \
301 RExC_naughty += RExC_naughty / (exp) + (add)
cde0cee5 302
a687059c
LW
303#define ISMULT1(c) ((c) == '*' || (c) == '+' || (c) == '?')
304#define ISMULT2(s) ((*s) == '*' || (*s) == '+' || (*s) == '?' || \
412f55bb 305 ((*s) == '{' && regcurly(s)))
a687059c
LW
306
307/*
308 * Flags to be passed up and down.
309 */
a687059c 310#define WORST 0 /* Worst case. */
a3b492c3 311#define HASWIDTH 0x01 /* Known to match non-null strings. */
fda99bee 312
e64f369d 313/* Simple enough to be STAR/PLUS operand; in an EXACTish node must be a single
2fd92675
KW
314 * character. (There needs to be a case: in the switch statement in regexec.c
315 * for any node marked SIMPLE.) Note that this is not the same thing as
316 * REGNODE_SIMPLE */
fda99bee 317#define SIMPLE 0x02
e64f369d 318#define SPSTART 0x04 /* Starts with * or + */
8d9c2815
NC
319#define POSTPONED 0x08 /* (?1),(?&name), (??{...}) or similar */
320#define TRYAGAIN 0x10 /* Weeded out a declaration. */
c36361b2
KW
321#define RESTART_PARSE 0x20 /* Need to redo the parse */
322#define NEED_UTF8 0x40 /* In conjunction with RESTART_PARSE, need to
323 calcuate sizes as UTF-8 */
a687059c 324
3dab1dad
YO
325#define REG_NODE_NUM(x) ((x) ? (int)((x)-RExC_emit_start) : -1)
326
07be1b83
YO
327/* whether trie related optimizations are enabled */
328#if PERL_ENABLE_EXTENDED_TRIE_OPTIMISATION
329#define TRIE_STUDY_OPT
786e8c11 330#define FULL_TRIE_STUDY
07be1b83
YO
331#define TRIE_STCLASS
332#endif
1de06328
YO
333
334
40d049e4
YO
335
336#define PBYTE(u8str,paren) ((U8*)(u8str))[(paren) >> 3]
337#define PBITVAL(paren) (1 << ((paren) & 7))
338#define PAREN_TEST(u8str,paren) ( PBYTE(u8str,paren) & PBITVAL(paren))
339#define PAREN_SET(u8str,paren) PBYTE(u8str,paren) |= PBITVAL(paren)
340#define PAREN_UNSET(u8str,paren) PBYTE(u8str,paren) &= (~PBITVAL(paren))
341
82a6ada4 342#define REQUIRE_UTF8(flagp) STMT_START { \
8d9c2815 343 if (!UTF) { \
82a6ada4 344 assert(PASS1); \
c36361b2 345 *flagp = RESTART_PARSE|NEED_UTF8; \
f55b7b26 346 return 0; \
8d9c2815 347 } \
82a6ada4 348 } STMT_END
40d049e4 349
512c0f5a
KW
350/* Change from /d into /u rules, and restart the parse if we've already seen
351 * something whose size would increase as a result, by setting *flagp and
352 * returning 'restart_retval'. RExC_uni_semantics is a flag that indicates
142ece99 353 * we've changed to /u during the parse. */
512c0f5a
KW
354#define REQUIRE_UNI_RULES(flagp, restart_retval) \
355 STMT_START { \
356 if (DEPENDS_SEMANTICS) { \
357 assert(PASS1); \
358 set_regex_charset(&RExC_flags, REGEX_UNICODE_CHARSET); \
359 RExC_uni_semantics = 1; \
360 if (RExC_seen_unfolded_sharp_s) { \
c36361b2 361 *flagp |= RESTART_PARSE; \
512c0f5a
KW
362 return restart_retval; \
363 } \
364 } \
365 } STMT_END
366
fd7fc6ae 367/* Executes a return statement with the value 'X', if 'flags' contains any of
c36361b2 368 * 'RESTART_PARSE', 'NEED_UTF8', or 'extra'. If so, *flagp is set to those
fd7fc6ae
YO
369 * flags */
370#define RETURN_X_ON_RESTART_OR_FLAGS(X, flags, flagp, extra) \
371 STMT_START { \
c36361b2
KW
372 if ((flags) & (RESTART_PARSE|NEED_UTF8|(extra))) { \
373 *(flagp) = (flags) & (RESTART_PARSE|NEED_UTF8|(extra)); \
fd7fc6ae
YO
374 return X; \
375 } \
376 } STMT_END
377
5d555bb8 378#define RETURN_FAIL_ON_RESTART_OR_FLAGS(flags,flagp,extra) \
f55b7b26 379 RETURN_X_ON_RESTART_OR_FLAGS(0,flags,flagp,extra)
fd7fc6ae
YO
380
381#define RETURN_X_ON_RESTART(X, flags,flagp) \
382 RETURN_X_ON_RESTART_OR_FLAGS( X, flags, flagp, 0)
383
384
5d555bb8 385#define RETURN_FAIL_ON_RESTART_FLAGP_OR_FLAGS(flagp,extra) \
c36361b2 386 if (*(flagp) & (RESTART_PARSE|(extra))) return 0
fd7fc6ae 387
c36361b2 388#define MUST_RESTART(flags) ((flags) & (RESTART_PARSE))
fd7fc6ae 389
5d555bb8 390#define RETURN_FAIL_ON_RESTART(flags,flagp) \
f55b7b26 391 RETURN_X_ON_RESTART(0, flags,flagp)
5d555bb8 392#define RETURN_FAIL_ON_RESTART_FLAGP(flagp) \
88f063b4 393 RETURN_FAIL_ON_RESTART_FLAGP_OR_FLAGS(flagp, 0)
fd7fc6ae 394
f19b1a63
KW
395/* This converts the named class defined in regcomp.h to its equivalent class
396 * number defined in handy.h. */
397#define namedclass_to_classnum(class) ((int) ((class) / 2))
398#define classnum_to_namedclass(classnum) ((classnum) * 2)
399
de92f5e6
KW
400#define _invlist_union_complement_2nd(a, b, output) \
401 _invlist_union_maybe_complement_2nd(a, b, TRUE, output)
402#define _invlist_intersection_complement_2nd(a, b, output) \
403 _invlist_intersection_maybe_complement_2nd(a, b, TRUE, output)
404
1de06328
YO
405/* About scan_data_t.
406
407 During optimisation we recurse through the regexp program performing
408 various inplace (keyhole style) optimisations. In addition study_chunk
409 and scan_commit populate this data structure with information about
538e84ed 410 what strings MUST appear in the pattern. We look for the longest
3b753521 411 string that must appear at a fixed location, and we look for the
1de06328
YO
412 longest string that may appear at a floating location. So for instance
413 in the pattern:
538e84ed 414
1de06328 415 /FOO[xX]A.*B[xX]BAR/
538e84ed 416
1de06328
YO
417 Both 'FOO' and 'A' are fixed strings. Both 'B' and 'BAR' are floating
418 strings (because they follow a .* construct). study_chunk will identify
419 both FOO and BAR as being the longest fixed and floating strings respectively.
538e84ed 420
1de06328 421 The strings can be composites, for instance
538e84ed 422
1de06328 423 /(f)(o)(o)/
538e84ed 424
1de06328 425 will result in a composite fixed substring 'foo'.
538e84ed 426
1de06328 427 For each string some basic information is maintained:
538e84ed 428
6ea77169 429 - min_offset
1de06328
YO
430 This is the position the string must appear at, or not before.
431 It also implicitly (when combined with minlenp) tells us how many
3b753521
FN
432 characters must match before the string we are searching for.
433 Likewise when combined with minlenp and the length of the string it
538e84ed 434 tells us how many characters must appear after the string we have
1de06328 435 found.
538e84ed 436
1de06328
YO
437 - max_offset
438 Only used for floating strings. This is the rightmost point that
ea3daa5d 439 the string can appear at. If set to SSize_t_MAX it indicates that the
1de06328 440 string can occur infinitely far to the right.
6e12f832 441 For fixed strings, it is equal to min_offset.
538e84ed 442
1de06328 443 - minlenp
2d608413
KW
444 A pointer to the minimum number of characters of the pattern that the
445 string was found inside. This is important as in the case of positive
538e84ed 446 lookahead or positive lookbehind we can have multiple patterns
1de06328 447 involved. Consider
538e84ed 448
1de06328 449 /(?=FOO).*F/
538e84ed 450
1de06328
YO
451 The minimum length of the pattern overall is 3, the minimum length
452 of the lookahead part is 3, but the minimum length of the part that
538e84ed 453 will actually match is 1. So 'FOO's minimum length is 3, but the
1de06328 454 minimum length for the F is 1. This is important as the minimum length
538e84ed 455 is used to determine offsets in front of and behind the string being
1de06328 456 looked for. Since strings can be composites this is the length of the
486ec47a 457 pattern at the time it was committed with a scan_commit. Note that
1de06328 458 the length is calculated by study_chunk, so that the minimum lengths
538e84ed 459 are not known until the full pattern has been compiled, thus the
1de06328 460 pointer to the value.
538e84ed 461
1de06328 462 - lookbehind
538e84ed 463
1de06328 464 In the case of lookbehind the string being searched for can be
538e84ed 465 offset past the start point of the final matching string.
1de06328
YO
466 If this value was just blithely removed from the min_offset it would
467 invalidate some of the calculations for how many chars must match
468 before or after (as they are derived from min_offset and minlen and
538e84ed 469 the length of the string being searched for).
1de06328
YO
470 When the final pattern is compiled and the data is moved from the
471 scan_data_t structure into the regexp structure the information
538e84ed
KW
472 about lookbehind is factored in, with the information that would
473 have been lost precalculated in the end_shift field for the
1de06328
YO
474 associated string.
475
476 The fields pos_min and pos_delta are used to store the minimum offset
538e84ed 477 and the delta to the maximum offset at the current point in the pattern.
1de06328
YO
478
479*/
2c2d71f5 480
cb6aef8b
DM
481struct scan_data_substrs {
482 SV *str; /* longest substring found in pattern */
483 SSize_t min_offset; /* earliest point in string it can appear */
484 SSize_t max_offset; /* latest point in string it can appear */
485 SSize_t *minlenp; /* pointer to the minlen relevant to the string */
486 SSize_t lookbehind; /* is the pos of the string modified by LB */
11683ecb 487 I32 flags; /* per substring SF_* and SCF_* flags */
cb6aef8b
DM
488};
489
2c2d71f5 490typedef struct scan_data_t {
1de06328
YO
491 /*I32 len_min; unused */
492 /*I32 len_delta; unused */
49f55535 493 SSize_t pos_min;
ea3daa5d 494 SSize_t pos_delta;
2c2d71f5 495 SV *last_found;
ea3daa5d 496 SSize_t last_end; /* min value, <0 unless valid. */
49f55535 497 SSize_t last_start_min;
ea3daa5d 498 SSize_t last_start_max;
2099df82
DM
499 U8 cur_is_floating; /* whether the last_* values should be set as
500 * the next fixed (0) or floating (1)
501 * substring */
6ea77169 502
2099df82
DM
503 /* [0] is longest fixed substring so far, [1] is longest float so far */
504 struct scan_data_substrs substrs[2];
6ea77169 505
11683ecb 506 I32 flags; /* common SF_* and SCF_* flags */
2c2d71f5 507 I32 whilem_c;
49f55535 508 SSize_t *last_closep;
b8f7bb16 509 regnode_ssc *start_class;
2c2d71f5
JH
510} scan_data_t;
511
a687059c 512/*
e50aee73 513 * Forward declarations for pregcomp()'s friends.
a687059c 514 */
a0d0e21e 515
6ea77169 516static const scan_data_t zero_scan_data = {
a2ca2017 517 0, 0, NULL, 0, 0, 0, 0,
6ea77169 518 {
11683ecb
DM
519 { NULL, 0, 0, 0, 0, 0 },
520 { NULL, 0, 0, 0, 0, 0 },
6ea77169
DM
521 },
522 0, 0, NULL, NULL
523};
c277df42 524
11683ecb
DM
525/* study flags */
526
07be1b83
YO
527#define SF_BEFORE_SEOL 0x0001
528#define SF_BEFORE_MEOL 0x0002
11683ecb 529#define SF_BEFORE_EOL (SF_BEFORE_SEOL|SF_BEFORE_MEOL)
c277df42 530
07be1b83
YO
531#define SF_IS_INF 0x0040
532#define SF_HAS_PAR 0x0080
533#define SF_IN_PAR 0x0100
534#define SF_HAS_EVAL 0x0200
bc604ad8
KW
535
536
537/* SCF_DO_SUBSTR is the flag that tells the regexp analyzer to track the
538 * longest substring in the pattern. When it is not set the optimiser keeps
539 * track of position, but does not keep track of the actual strings seen,
540 *
541 * So for instance /foo/ will be parsed with SCF_DO_SUBSTR being true, but
542 * /foo/i will not.
543 *
544 * Similarly, /foo.*(blah|erm|huh).*fnorble/ will have "foo" and "fnorble"
545 * parsed with SCF_DO_SUBSTR on, but while processing the (...) it will be
546 * turned off because of the alternation (BRANCH). */
07be1b83 547#define SCF_DO_SUBSTR 0x0400
bc604ad8 548
653099ff
GS
549#define SCF_DO_STCLASS_AND 0x0800
550#define SCF_DO_STCLASS_OR 0x1000
551#define SCF_DO_STCLASS (SCF_DO_STCLASS_AND|SCF_DO_STCLASS_OR)
e1901655 552#define SCF_WHILEM_VISITED_POS 0x2000
c277df42 553
786e8c11 554#define SCF_TRIE_RESTUDY 0x4000 /* Do restudy? */
538e84ed 555#define SCF_SEEN_ACCEPT 0x8000
688e0391 556#define SCF_TRIE_DOING_RESTUDY 0x10000
4286711a
YO
557#define SCF_IN_DEFINE 0x20000
558
559
560
07be1b83 561
43fead97 562#define UTF cBOOL(RExC_utf8)
00b27cfc
KW
563
564/* The enums for all these are ordered so things work out correctly */
a62b1201 565#define LOC (get_regex_charset(RExC_flags) == REGEX_LOCALE_CHARSET)
538e84ed
KW
566#define DEPENDS_SEMANTICS (get_regex_charset(RExC_flags) \
567 == REGEX_DEPENDS_CHARSET)
00b27cfc 568#define UNI_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_UNICODE_CHARSET)
538e84ed
KW
569#define AT_LEAST_UNI_SEMANTICS (get_regex_charset(RExC_flags) \
570 >= REGEX_UNICODE_CHARSET)
571#define ASCII_RESTRICTED (get_regex_charset(RExC_flags) \
572 == REGEX_ASCII_RESTRICTED_CHARSET)
573#define AT_LEAST_ASCII_RESTRICTED (get_regex_charset(RExC_flags) \
574 >= REGEX_ASCII_RESTRICTED_CHARSET)
575#define ASCII_FOLD_RESTRICTED (get_regex_charset(RExC_flags) \
576 == REGEX_ASCII_MORE_RESTRICTED_CHARSET)
a62b1201 577
43fead97 578#define FOLD cBOOL(RExC_flags & RXf_PMf_FOLD)
a0ed51b3 579
f2c2a6ab
KW
580/* For programs that want to be strictly Unicode compatible by dying if any
581 * attempt is made to match a non-Unicode code point against a Unicode
582 * property. */
583#define ALWAYS_WARN_SUPER ckDEAD(packWARN(WARN_NON_UNICODE))
584
93733859 585#define OOB_NAMEDCLASS -1
b8c5462f 586
8e661ac5
KW
587/* There is no code point that is out-of-bounds, so this is problematic. But
588 * its only current use is to initialize a variable that is always set before
589 * looked at. */
590#define OOB_UNICODE 0xDEADBEEF
591
a0ed51b3 592#define CHR_SVLEN(sv) (UTF ? sv_len_utf8(sv) : SvCUR(sv))
a0ed51b3 593
8615cb43 594
b45f050a
JF
595/* length of regex to show in messages that don't mark a position within */
596#define RegexLengthToShowInErrorMessages 127
597
598/*
599 * If MARKER[12] are adjusted, be sure to adjust the constants at the top
600 * of t/op/regmesg.t, the tests in t/op/re_tests, and those in
601 * op/pragma/warn/regcomp.
602 */
7253e4e3
RK
603#define MARKER1 "<-- HERE" /* marker as it appears in the description */
604#define MARKER2 " <-- HERE " /* marker as it appears within the regex */
b81d288d 605
538e84ed 606#define REPORT_LOCATION " in regex; marked by " MARKER1 \
147e3846 607 " in m/%" UTF8f MARKER2 "%" UTF8f "/"
b45f050a 608
285b5ca0
KW
609/* The code in this file in places uses one level of recursion with parsing
610 * rebased to an alternate string constructed by us in memory. This can take
611 * the form of something that is completely different from the input, or
612 * something that uses the input as part of the alternate. In the first case,
613 * there should be no possibility of an error, as we are in complete control of
51684e3c
KW
614 * the alternate string. But in the second case we don't completely control
615 * the input portion, so there may be errors in that. Here's an example:
285b5ca0
KW
616 * /[abc\x{DF}def]/ui
617 * is handled specially because \x{df} folds to a sequence of more than one
51684e3c 618 * character: 'ss'. What is done is to create and parse an alternate string,
285b5ca0
KW
619 * which looks like this:
620 * /(?:\x{DF}|[abc\x{DF}def])/ui
621 * where it uses the input unchanged in the middle of something it constructs,
622 * which is a branch for the DF outside the character class, and clustering
623 * parens around the whole thing. (It knows enough to skip the DF inside the
624 * class while in this substitute parse.) 'abc' and 'def' may have errors that
625 * need to be reported. The general situation looks like this:
626 *
51684e3c 627 * |<------- identical ------>|
285b5ca0 628 * sI tI xI eI
51684e3c 629 * Input: ---------------------------------------------------------------
285b5ca0
KW
630 * Constructed: ---------------------------------------------------
631 * sC tC xC eC EC
51684e3c 632 * |<------- identical ------>|
285b5ca0 633 *
51684e3c
KW
634 * sI..eI is the portion of the input pattern we are concerned with here.
635 * sC..EC is the constructed substitute parse string.
636 * sC..tC is constructed by us
637 * tC..eC is an exact duplicate of the portion of the input pattern tI..eI.
638 * In the diagram, these are vertically aligned.
639 * eC..EC is also constructed by us.
640 * xC is the position in the substitute parse string where we found a
641 * problem.
642 * xI is the position in the original pattern corresponding to xC.
285b5ca0 643 *
51684e3c
KW
644 * We want to display a message showing the real input string. Thus we need to
645 * translate from xC to xI. We know that xC >= tC, since the portion of the
646 * string sC..tC has been constructed by us, and so shouldn't have errors. We
647 * get:
648 * xI = tI + (xC - tC)
285b5ca0 649 *
51684e3c
KW
650 * When the substitute parse is constructed, the code needs to set:
651 * RExC_start (sC)
652 * RExC_end (eC)
653 * RExC_copy_start_in_input (tI)
654 * RExC_copy_start_in_constructed (tC)
655 * and restore them when done.
285b5ca0 656 *
51684e3c
KW
657 * During normal processing of the input pattern, both
658 * 'RExC_copy_start_in_input' and 'RExC_copy_start_in_constructed' are set to
659 * sI, so that xC equals xI.
285b5ca0
KW
660 */
661
51684e3c
KW
662#define sI RExC_precomp
663#define eI RExC_precomp_end
664#define sC RExC_start
665#define eC RExC_end
666#define tI RExC_copy_start_in_input
667#define tC RExC_copy_start_in_constructed
668#define xI(xC) (tI + (xC - tC))
669#define xI_offset(xC) (xI(xC) - sI)
285b5ca0
KW
670
671#define REPORT_LOCATION_ARGS(xC) \
672 UTF8fARG(UTF, \
51684e3c 673 (xI(xC) > eI) /* Don't run off end */ \
285b5ca0 674 ? eC - sC /* Length before the <--HERE */ \
bc929b61
KW
675 : ((xI_offset(xC) >= 0) \
676 ? xI_offset(xC) \
677 : (Perl_croak(aTHX_ "panic: %s: %d: negative offset: %" \
678 IVdf " trying to output message for " \
679 " pattern %.*s", \
51684e3c 680 __FILE__, __LINE__, (IV) xI_offset(xC), \
bc929b61 681 ((int) (eC - sC)), sC), 0)), \
51684e3c 682 sI), /* The input pattern printed up to the <--HERE */ \
285b5ca0 683 UTF8fARG(UTF, \
51684e3c
KW
684 (xI(xC) > eI) ? 0 : eI - xI(xC), /* Length after <--HERE */ \
685 (xI(xC) > eI) ? eI : xI(xC)) /* pattern after <--HERE */
c1d900c3 686
8a6d8ec6
HS
687/* Used to point after bad bytes for an error message, but avoid skipping
688 * past a nul byte. */
689#define SKIP_IF_CHAR(s) (!*(s) ? 0 : UTF ? UTF8SKIP(s) : 1)
690
ed68fe11
KW
691/* Set up to clean up after our imminent demise */
692#define PREPARE_TO_DIE \
693 STMT_START { \
694 if (RExC_rx_sv) \
695 SAVEFREESV(RExC_rx_sv); \
696 } STMT_END
697
b45f050a
JF
698/*
699 * Calls SAVEDESTRUCTOR_X if needed, then calls Perl_croak with the given
700 * arg. Show regex, up to a maximum length. If it's too long, chop and add
701 * "...".
702 */
58e23c8d 703#define _FAIL(code) STMT_START { \
bfed75c6 704 const char *ellipses = ""; \
88f063b4 705 IV len = RExC_precomp_end - RExC_precomp; \
ccb2c380 706 \
ed68fe11 707 PREPARE_TO_DIE; \
ccb2c380
MP
708 if (len > RegexLengthToShowInErrorMessages) { \
709 /* chop 10 shorter than the max, to ensure meaning of "..." */ \
710 len = RegexLengthToShowInErrorMessages - 10; \
711 ellipses = "..."; \
712 } \
58e23c8d 713 code; \
ccb2c380 714} STMT_END
8615cb43 715
58e23c8d 716#define FAIL(msg) _FAIL( \
147e3846 717 Perl_croak(aTHX_ "%s in regex m/%" UTF8f "%s/", \
c1d900c3 718 msg, UTF8fARG(UTF, len, RExC_precomp), ellipses))
58e23c8d
YO
719
720#define FAIL2(msg,arg) _FAIL( \
147e3846 721 Perl_croak(aTHX_ msg " in regex m/%" UTF8f "%s/", \
c1d900c3 722 arg, UTF8fARG(UTF, len, RExC_precomp), ellipses))
58e23c8d 723
b45f050a 724/*
b45f050a
JF
725 * Simple_vFAIL -- like FAIL, but marks the current location in the scan
726 */
ccb2c380 727#define Simple_vFAIL(m) STMT_START { \
ccb2c380 728 Perl_croak(aTHX_ "%s" REPORT_LOCATION, \
d528642a 729 m, REPORT_LOCATION_ARGS(RExC_parse)); \
ccb2c380 730} STMT_END
b45f050a
JF
731
732/*
733 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL()
734 */
ccb2c380 735#define vFAIL(m) STMT_START { \
ed68fe11 736 PREPARE_TO_DIE; \
ccb2c380
MP
737 Simple_vFAIL(m); \
738} STMT_END
b45f050a
JF
739
740/*
741 * Like Simple_vFAIL(), but accepts two arguments.
742 */
ccb2c380 743#define Simple_vFAIL2(m,a1) STMT_START { \
d528642a
KW
744 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, \
745 REPORT_LOCATION_ARGS(RExC_parse)); \
ccb2c380 746} STMT_END
b45f050a
JF
747
748/*
749 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL2().
750 */
ccb2c380 751#define vFAIL2(m,a1) STMT_START { \
ed68fe11 752 PREPARE_TO_DIE; \
ccb2c380
MP
753 Simple_vFAIL2(m, a1); \
754} STMT_END
b45f050a
JF
755
756
757/*
758 * Like Simple_vFAIL(), but accepts three arguments.
759 */
ccb2c380 760#define Simple_vFAIL3(m, a1, a2) STMT_START { \
c1d900c3 761 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, \
d528642a 762 REPORT_LOCATION_ARGS(RExC_parse)); \
ccb2c380 763} STMT_END
b45f050a
JF
764
765/*
766 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL3().
767 */
ccb2c380 768#define vFAIL3(m,a1,a2) STMT_START { \
ed68fe11 769 PREPARE_TO_DIE; \
ccb2c380
MP
770 Simple_vFAIL3(m, a1, a2); \
771} STMT_END
b45f050a
JF
772
773/*
774 * Like Simple_vFAIL(), but accepts four arguments.
775 */
ccb2c380 776#define Simple_vFAIL4(m, a1, a2, a3) STMT_START { \
d528642a
KW
777 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, a3, \
778 REPORT_LOCATION_ARGS(RExC_parse)); \
ccb2c380 779} STMT_END
b45f050a 780
95db3ffa 781#define vFAIL4(m,a1,a2,a3) STMT_START { \
ed68fe11 782 PREPARE_TO_DIE; \
95db3ffa
KW
783 Simple_vFAIL4(m, a1, a2, a3); \
784} STMT_END
785
946095af 786/* A specialized version of vFAIL2 that works with UTF8f */
d528642a 787#define vFAIL2utf8f(m, a1) STMT_START { \
ed68fe11 788 PREPARE_TO_DIE; \
d528642a
KW
789 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, \
790 REPORT_LOCATION_ARGS(RExC_parse)); \
946095af
BF
791} STMT_END
792
3ba22297 793#define vFAIL3utf8f(m, a1, a2) STMT_START { \
ed68fe11 794 PREPARE_TO_DIE; \
3ba22297 795 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, \
d528642a 796 REPORT_LOCATION_ARGS(RExC_parse)); \
3ba22297
KW
797} STMT_END
798
d62e749e
KW
799/* Setting this to NULL is a signal to not output warnings */
800#define TURN_OFF_WARNINGS_IN_SUBSTITUTE_PARSE RExC_copy_start_in_constructed = NULL
801#define RESTORE_WARNINGS RExC_copy_start_in_constructed = RExC_precomp
802
338190a1
KW
803/* Outputting warnings is generally deferred until the 2nd pass. This is
804 * because the first pass can be restarted, for example if the pattern has to
805 * be converted to UTF-8. If a warning had already been output earlier in the
806 * pass, it would be re-output after the restart. Pass 2 is never restarted,
807 * so the problem simply goes away if we defer the output to that pass. See
d62e749e
KW
808 * [perl #122671]. 'RExC_copy_start_in_constructed' being NULL is a flag to
809 * not generate any warnings */
55bcf857 810#define TO_OUTPUT_WARNINGS(loc) \
d62e749e 811 (PASS2 && RExC_copy_start_in_constructed)
55bcf857 812
f8220fe2
KW
813#define _WARN_HELPER(loc, warns, code) \
814 STMT_START { \
d62e749e
KW
815 if (! RExC_copy_start_in_constructed) { \
816 Perl_croak( aTHX_ "panic! %s: %d: Tried to warn when none" \
817 " expected at '%s'", \
818 __FILE__, __LINE__, loc); \
819 } \
55bcf857 820 if (TO_OUTPUT_WARNINGS(loc)) { \
338190a1
KW
821 code; \
822 } \
f8220fe2 823 } STMT_END
946095af 824
5e0a247b 825/* m is not necessarily a "literal string", in this macro */
f8220fe2
KW
826#define reg_warn_non_literal_string(loc, m) \
827 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
828 Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
d528642a 829 "%s" REPORT_LOCATION, \
f8220fe2 830 m, REPORT_LOCATION_ARGS(loc)))
5e0a247b 831
f8220fe2
KW
832#define ckWARNreg(loc,m) \
833 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
834 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
d528642a 835 m REPORT_LOCATION, \
f8220fe2 836 REPORT_LOCATION_ARGS(loc)))
ccb2c380 837
f8220fe2
KW
838#define vWARN(loc, m) \
839 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
840 Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
d528642a 841 m REPORT_LOCATION, \
f8220fe2 842 REPORT_LOCATION_ARGS(loc))) \
b927b7e9 843
f8220fe2
KW
844#define vWARN_dep(loc, m) \
845 _WARN_HELPER(loc, packWARN(WARN_DEPRECATED), \
846 Perl_warner(aTHX_ packWARN(WARN_DEPRECATED), \
d528642a 847 m REPORT_LOCATION, \
f8220fe2 848 REPORT_LOCATION_ARGS(loc)))
0d6106aa 849
f8220fe2
KW
850#define ckWARNdep(loc,m) \
851 _WARN_HELPER(loc, packWARN(WARN_DEPRECATED), \
852 Perl_ck_warner_d(aTHX_ packWARN(WARN_DEPRECATED), \
d528642a 853 m REPORT_LOCATION, \
f8220fe2 854 REPORT_LOCATION_ARGS(loc)))
147508a2 855
f8220fe2
KW
856#define ckWARNregdep(loc,m) \
857 _WARN_HELPER(loc, packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
858 Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, \
d528642a
KW
859 WARN_REGEXP), \
860 m REPORT_LOCATION, \
f8220fe2 861 REPORT_LOCATION_ARGS(loc)))
ccb2c380 862
f8220fe2
KW
863#define ckWARN2reg_d(loc,m, a1) \
864 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
865 Perl_ck_warner_d(aTHX_ packWARN(WARN_REGEXP), \
d528642a 866 m REPORT_LOCATION, \
f8220fe2 867 a1, REPORT_LOCATION_ARGS(loc)))
2335b3d3 868
f8220fe2
KW
869#define ckWARN2reg(loc, m, a1) \
870 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
871 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
d528642a 872 m REPORT_LOCATION, \
f8220fe2 873 a1, REPORT_LOCATION_ARGS(loc)))
ccb2c380 874
f8220fe2
KW
875#define vWARN3(loc, m, a1, a2) \
876 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
877 Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
d528642a 878 m REPORT_LOCATION, \
f8220fe2 879 a1, a2, REPORT_LOCATION_ARGS(loc)))
ccb2c380 880
f8220fe2
KW
881#define ckWARN3reg(loc, m, a1, a2) \
882 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
883 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
d528642a
KW
884 m REPORT_LOCATION, \
885 a1, a2, \
f8220fe2 886 REPORT_LOCATION_ARGS(loc)))
668c081a 887
f8220fe2
KW
888#define vWARN4(loc, m, a1, a2, a3) \
889 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
890 Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
d528642a
KW
891 m REPORT_LOCATION, \
892 a1, a2, a3, \
f8220fe2 893 REPORT_LOCATION_ARGS(loc)))
ccb2c380 894
f8220fe2
KW
895#define ckWARN4reg(loc, m, a1, a2, a3) \
896 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
897 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
d528642a
KW
898 m REPORT_LOCATION, \
899 a1, a2, a3, \
f8220fe2 900 REPORT_LOCATION_ARGS(loc)))
668c081a 901
f8220fe2
KW
902#define vWARN5(loc, m, a1, a2, a3, a4) \
903 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
904 Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
d528642a
KW
905 m REPORT_LOCATION, \
906 a1, a2, a3, a4, \
f8220fe2 907 REPORT_LOCATION_ARGS(loc)))
9d1d55b5 908
aea5f181
KW
909#define ckWARNexperimental(loc, class, m) \
910 _WARN_HELPER(loc, packWARN(class), \
911 Perl_ck_warner_d(aTHX_ packWARN(class), \
912 m REPORT_LOCATION, \
913 REPORT_LOCATION_ARGS(loc)))
914
1a322bb4
KW
915/* Convert between a pointer to a node and its offset from the beginning of the
916 * program */
917#define REGNODE_p(offset) (RExC_emit_start + (offset))
918#define REGNODE_OFFSET(node) ((node) - RExC_emit_start)
919
538e84ed 920/* Macros for recording node offsets. 20001227 mjd@plover.com
fac92740
MJD
921 * Nodes are numbered 1, 2, 3, 4. Node #n's position is recorded in
922 * element 2*n-1 of the array. Element #2n holds the byte length node #n.
923 * Element 0 holds the number n.
07be1b83 924 * Position is 1 indexed.
fac92740 925 */
7122b237 926#ifndef RE_TRACK_PATTERN_OFFSETS
d7c442ce 927#define Set_Node_Offset_To_R(offset,byte)
7122b237
YO
928#define Set_Node_Offset(node,byte)
929#define Set_Cur_Node_Offset
930#define Set_Node_Length_To_R(node,len)
931#define Set_Node_Length(node,len)
6a86c6ad 932#define Set_Node_Cur_Length(node,start)
538e84ed
KW
933#define Node_Offset(n)
934#define Node_Length(n)
7122b237
YO
935#define Set_Node_Offset_Length(node,offset,len)
936#define ProgLen(ri) ri->u.proglen
937#define SetProgLen(ri,x) ri->u.proglen = x
938#else
939#define ProgLen(ri) ri->u.offsets[0]
940#define SetProgLen(ri,x) ri->u.offsets[0] = x
d7c442ce 941#define Set_Node_Offset_To_R(offset,byte) STMT_START { \
ccb2c380
MP
942 if (! SIZE_ONLY) { \
943 MJD_OFFSET_DEBUG(("** (%d) offset of node %d is %d.\n", \
d7c442ce
KW
944 __LINE__, (int)(offset), (int)(byte))); \
945 if((offset) < 0) { \
538e84ed 946 Perl_croak(aTHX_ "value of node is %d in Offset macro", \
d7c442ce 947 (int)(offset)); \
ccb2c380 948 } else { \
d7c442ce 949 RExC_offsets[2*(offset)-1] = (byte); \
ccb2c380
MP
950 } \
951 } \
952} STMT_END
953
88f063b4 954#define Set_Node_Offset(node,byte) \
1a322bb4 955 Set_Node_Offset_To_R(REGNODE_OFFSET(node), (byte)-RExC_start)
ccb2c380
MP
956#define Set_Cur_Node_Offset Set_Node_Offset(RExC_emit, RExC_parse)
957
958#define Set_Node_Length_To_R(node,len) STMT_START { \
959 if (! SIZE_ONLY) { \
960 MJD_OFFSET_DEBUG(("** (%d) size of node %d is %d.\n", \
551405c4 961 __LINE__, (int)(node), (int)(len))); \
ccb2c380 962 if((node) < 0) { \
538e84ed
KW
963 Perl_croak(aTHX_ "value of node is %d in Length macro", \
964 (int)(node)); \
ccb2c380
MP
965 } else { \
966 RExC_offsets[2*(node)] = (len); \
967 } \
968 } \
969} STMT_END
970
971#define Set_Node_Length(node,len) \
1a322bb4 972 Set_Node_Length_To_R(REGNODE_OFFSET(node), len)
6a86c6ad
NC
973#define Set_Node_Cur_Length(node, start) \
974 Set_Node_Length(node, RExC_parse - start)
fac92740
MJD
975
976/* Get offsets and lengths */
1a322bb4
KW
977#define Node_Offset(n) (RExC_offsets[2*(REGNODE_OFFSET(n))-1])
978#define Node_Length(n) (RExC_offsets[2*(REGNODE_OFFSET(n))])
fac92740 979
07be1b83 980#define Set_Node_Offset_Length(node,offset,len) STMT_START { \
1a322bb4
KW
981 Set_Node_Offset_To_R(REGNODE_OFFSET(node), (offset)); \
982 Set_Node_Length_To_R(REGNODE_OFFSET(node), (len)); \
07be1b83 983} STMT_END
7122b237 984#endif
07be1b83
YO
985
986#if PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS
987#define EXPERIMENTAL_INPLACESCAN
f427392e 988#endif /*PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS*/
07be1b83 989
cb41e5d6
YO
990#ifdef DEBUGGING
991int
6ad9a8ab 992Perl_re_printf(pTHX_ const char *fmt, ...)
cb41e5d6
YO
993{
994 va_list ap;
995 int result;
996 PerlIO *f= Perl_debug_log;
997 PERL_ARGS_ASSERT_RE_PRINTF;
998 va_start(ap, fmt);
999 result = PerlIO_vprintf(f, fmt, ap);
1000 va_end(ap);
1001 return result;
1002}
1003
1004int
7b031478 1005Perl_re_indentf(pTHX_ const char *fmt, U32 depth, ...)
cb41e5d6
YO
1006{
1007 va_list ap;
1008 int result;
1009 PerlIO *f= Perl_debug_log;
1010 PERL_ARGS_ASSERT_RE_INDENTF;
1011 va_start(ap, depth);
daeb874b 1012 PerlIO_printf(f, "%*s", ( (int)depth % 20 ) * 2, "");
cb41e5d6
YO
1013 result = PerlIO_vprintf(f, fmt, ap);
1014 va_end(ap);
1015 return result;
1016}
1017#endif /* DEBUGGING */
1018
7b031478 1019#define DEBUG_RExC_seen() \
538e84ed 1020 DEBUG_OPTIMISE_MORE_r({ \
88f063b4 1021 Perl_re_printf( aTHX_ "RExC_seen: "); \
538e84ed 1022 \
e384d5c1 1023 if (RExC_seen & REG_ZERO_LEN_SEEN) \
88f063b4 1024 Perl_re_printf( aTHX_ "REG_ZERO_LEN_SEEN "); \
538e84ed 1025 \
e384d5c1 1026 if (RExC_seen & REG_LOOKBEHIND_SEEN) \
88f063b4 1027 Perl_re_printf( aTHX_ "REG_LOOKBEHIND_SEEN "); \
538e84ed 1028 \
e384d5c1 1029 if (RExC_seen & REG_GPOS_SEEN) \
88f063b4 1030 Perl_re_printf( aTHX_ "REG_GPOS_SEEN "); \
538e84ed 1031 \
e384d5c1 1032 if (RExC_seen & REG_RECURSE_SEEN) \
88f063b4 1033 Perl_re_printf( aTHX_ "REG_RECURSE_SEEN "); \
538e84ed 1034 \
7b031478 1035 if (RExC_seen & REG_TOP_LEVEL_BRANCHES_SEEN) \
88f063b4 1036 Perl_re_printf( aTHX_ "REG_TOP_LEVEL_BRANCHES_SEEN "); \
538e84ed 1037 \
e384d5c1 1038 if (RExC_seen & REG_VERBARG_SEEN) \
88f063b4 1039 Perl_re_printf( aTHX_ "REG_VERBARG_SEEN "); \
538e84ed 1040 \
e384d5c1 1041 if (RExC_seen & REG_CUTGROUP_SEEN) \
88f063b4 1042 Perl_re_printf( aTHX_ "REG_CUTGROUP_SEEN "); \
538e84ed 1043 \
e384d5c1 1044 if (RExC_seen & REG_RUN_ON_COMMENT_SEEN) \
88f063b4 1045 Perl_re_printf( aTHX_ "REG_RUN_ON_COMMENT_SEEN "); \
538e84ed 1046 \
e384d5c1 1047 if (RExC_seen & REG_UNFOLDED_MULTI_SEEN) \
88f063b4 1048 Perl_re_printf( aTHX_ "REG_UNFOLDED_MULTI_SEEN "); \
538e84ed 1049 \
7b031478 1050 if (RExC_seen & REG_UNBOUNDED_QUANTIFIER_SEEN) \
88f063b4 1051 Perl_re_printf( aTHX_ "REG_UNBOUNDED_QUANTIFIER_SEEN "); \
ee273784 1052 \
88f063b4 1053 Perl_re_printf( aTHX_ "\n"); \
9e9ecfdf
YO
1054 });
1055
fdfb4f21 1056#define DEBUG_SHOW_STUDY_FLAG(flags,flag) \
6ad9a8ab 1057 if ((flags) & flag) Perl_re_printf( aTHX_ "%s ", #flag)
fdfb4f21 1058
fdfb4f21 1059
f5a36d78
DM
1060#ifdef DEBUGGING
1061static void
1062S_debug_show_study_flags(pTHX_ U32 flags, const char *open_str,
1063 const char *close_str)
1064{
1065 if (!flags)
1066 return;
1067
1068 Perl_re_printf( aTHX_ "%s", open_str);
11683ecb
DM
1069 DEBUG_SHOW_STUDY_FLAG(flags, SF_BEFORE_SEOL);
1070 DEBUG_SHOW_STUDY_FLAG(flags, SF_BEFORE_MEOL);
f5a36d78
DM
1071 DEBUG_SHOW_STUDY_FLAG(flags, SF_IS_INF);
1072 DEBUG_SHOW_STUDY_FLAG(flags, SF_HAS_PAR);
1073 DEBUG_SHOW_STUDY_FLAG(flags, SF_IN_PAR);
1074 DEBUG_SHOW_STUDY_FLAG(flags, SF_HAS_EVAL);
1075 DEBUG_SHOW_STUDY_FLAG(flags, SCF_DO_SUBSTR);
1076 DEBUG_SHOW_STUDY_FLAG(flags, SCF_DO_STCLASS_AND);
1077 DEBUG_SHOW_STUDY_FLAG(flags, SCF_DO_STCLASS_OR);
1078 DEBUG_SHOW_STUDY_FLAG(flags, SCF_DO_STCLASS);
1079 DEBUG_SHOW_STUDY_FLAG(flags, SCF_WHILEM_VISITED_POS);
1080 DEBUG_SHOW_STUDY_FLAG(flags, SCF_TRIE_RESTUDY);
1081 DEBUG_SHOW_STUDY_FLAG(flags, SCF_SEEN_ACCEPT);
1082 DEBUG_SHOW_STUDY_FLAG(flags, SCF_TRIE_DOING_RESTUDY);
1083 DEBUG_SHOW_STUDY_FLAG(flags, SCF_IN_DEFINE);
1084 Perl_re_printf( aTHX_ "%s", close_str);
1085}
1086
1087
1088static void
1089S_debug_studydata(pTHX_ const char *where, scan_data_t *data,
1090 U32 depth, int is_inf)
1091{
1092 GET_RE_DEBUG_FLAGS_DECL;
1093
1094 DEBUG_OPTIMISE_MORE_r({
1095 if (!data)
1096 return;
1097 Perl_re_indentf(aTHX_ "%s: Pos:%" IVdf "/%" IVdf " Flags: 0x%" UVXf,
1098 depth,
1099 where,
1100 (IV)data->pos_min,
1101 (IV)data->pos_delta,
1102 (UV)data->flags
1103 );
1104
1105 S_debug_show_study_flags(aTHX_ data->flags," [","]");
fdfb4f21 1106
f5a36d78
DM
1107 Perl_re_printf( aTHX_
1108 " Whilem_c: %" IVdf " Lcp: %" IVdf " %s",
1109 (IV)data->whilem_c,
1110 (IV)(data->last_closep ? *((data)->last_closep) : -1),
1111 is_inf ? "INF " : ""
1112 );
1113
11683ecb 1114 if (data->last_found) {
37b6262f 1115 int i;
11683ecb
DM
1116 Perl_re_printf(aTHX_
1117 "Last:'%s' %" IVdf ":%" IVdf "/%" IVdf,
1118 SvPVX_const(data->last_found),
1119 (IV)data->last_end,
1120 (IV)data->last_start_min,
1121 (IV)data->last_start_max
1122 );
1123
2099df82 1124 for (i = 0; i < 2; i++) {
37b6262f
DM
1125 Perl_re_printf(aTHX_
1126 " %s%s: '%s' @ %" IVdf "/%" IVdf,
2099df82
DM
1127 data->cur_is_floating == i ? "*" : "",
1128 i ? "Float" : "Fixed",
37b6262f
DM
1129 SvPVX_const(data->substrs[i].str),
1130 (IV)data->substrs[i].min_offset,
1131 (IV)data->substrs[i].max_offset
1132 );
1133 S_debug_show_study_flags(aTHX_ data->substrs[i].flags," [","]");
1134 }
11683ecb 1135 }
f5a36d78
DM
1136
1137 Perl_re_printf( aTHX_ "\n");
1138 });
1139}
1140
1141
1142static void
1143S_debug_peep(pTHX_ const char *str, const RExC_state_t *pRExC_state,
1144 regnode *scan, U32 depth, U32 flags)
1145{
1146 GET_RE_DEBUG_FLAGS_DECL;
1147
1148 DEBUG_OPTIMISE_r({
1149 regnode *Next;
1150
1151 if (!scan)
1152 return;
1153 Next = regnext(scan);
1154 regprop(RExC_rx, RExC_mysv, scan, NULL, pRExC_state);
1155 Perl_re_indentf( aTHX_ "%s>%3d: %s (%d)",
1156 depth,
1157 str,
1158 REG_NODE_NUM(scan), SvPV_nolen_const(RExC_mysv),
1159 Next ? (REG_NODE_NUM(Next)) : 0 );
1160 S_debug_show_study_flags(aTHX_ flags," [ ","]");
1161 Perl_re_printf( aTHX_ "\n");
1162 });
1163}
1164
1165
1166# define DEBUG_STUDYDATA(where, data, depth, is_inf) \
1167 S_debug_studydata(aTHX_ where, data, depth, is_inf)
1168
1169# define DEBUG_PEEP(str, scan, depth, flags) \
1170 S_debug_peep(aTHX_ str, pRExC_state, scan, depth, flags)
1171
1172#else
1173# define DEBUG_STUDYDATA(where, data, depth, is_inf) NOOP
1174# define DEBUG_PEEP(str, scan, depth, flags) NOOP
1175#endif
1de06328 1176
cb41e5d6 1177
c6871b76
KW
1178/* =========================================================
1179 * BEGIN edit_distance stuff.
1180 *
1181 * This calculates how many single character changes of any type are needed to
1182 * transform a string into another one. It is taken from version 3.1 of
1183 *
1184 * https://metacpan.org/pod/Text::Levenshtein::Damerau::XS
1185 */
1186
1187/* Our unsorted dictionary linked list. */
1188/* Note we use UVs, not chars. */
1189
1190struct dictionary{
1191 UV key;
1192 UV value;
1193 struct dictionary* next;
1194};
1195typedef struct dictionary item;
1196
1197
1198PERL_STATIC_INLINE item*
88f063b4 1199push(UV key, item* curr)
c6871b76
KW
1200{
1201 item* head;
d09f14bf 1202 Newx(head, 1, item);
c6871b76
KW
1203 head->key = key;
1204 head->value = 0;
1205 head->next = curr;
1206 return head;
1207}
1208
1209
1210PERL_STATIC_INLINE item*
1211find(item* head, UV key)
1212{
1213 item* iterator = head;
1214 while (iterator){
1215 if (iterator->key == key){
1216 return iterator;
1217 }
1218 iterator = iterator->next;
1219 }
1220
1221 return NULL;
1222}
1223
1224PERL_STATIC_INLINE item*
88f063b4 1225uniquePush(item* head, UV key)
c6871b76
KW
1226{
1227 item* iterator = head;
1228
1229 while (iterator){
1230 if (iterator->key == key) {
1231 return head;
1232 }
1233 iterator = iterator->next;
1234 }
1235
88f063b4 1236 return push(key, head);
c6871b76
KW
1237}
1238
1239PERL_STATIC_INLINE void
1240dict_free(item* head)
1241{
1242 item* iterator = head;
1243
1244 while (iterator) {
1245 item* temp = iterator;
1246 iterator = iterator->next;
1247 Safefree(temp);
1248 }
1249
1250 head = NULL;
1251}
1252
1253/* End of Dictionary Stuff */
1254
1255/* All calculations/work are done here */
1256STATIC int
1257S_edit_distance(const UV* src,
1258 const UV* tgt,
1259 const STRLEN x, /* length of src[] */
1260 const STRLEN y, /* length of tgt[] */
1261 const SSize_t maxDistance
1262)
1263{
1264 item *head = NULL;
88f063b4 1265 UV swapCount, swapScore, targetCharCount, i, j;
c6871b76
KW
1266 UV *scores;
1267 UV score_ceil = x + y;
1268
1269 PERL_ARGS_ASSERT_EDIT_DISTANCE;
1270
1271 /* intialize matrix start values */
d09f14bf 1272 Newx(scores, ( (x + 2) * (y + 2)), UV);
c6871b76
KW
1273 scores[0] = score_ceil;
1274 scores[1 * (y + 2) + 0] = score_ceil;
1275 scores[0 * (y + 2) + 1] = score_ceil;
1276 scores[1 * (y + 2) + 1] = 0;
88f063b4 1277 head = uniquePush(uniquePush(head, src[0]), tgt[0]);
c6871b76
KW
1278
1279 /* work loops */
1280 /* i = src index */
1281 /* j = tgt index */
1282 for (i=1;i<=x;i++) {
1283 if (i < x)
88f063b4 1284 head = uniquePush(head, src[i]);
c6871b76
KW
1285 scores[(i+1) * (y + 2) + 1] = i;
1286 scores[(i+1) * (y + 2) + 0] = score_ceil;
1287 swapCount = 0;
1288
1289 for (j=1;j<=y;j++) {
1290 if (i == 1) {
1291 if(j < y)
88f063b4 1292 head = uniquePush(head, tgt[j]);
c6871b76
KW
1293 scores[1 * (y + 2) + (j + 1)] = j;
1294 scores[0 * (y + 2) + (j + 1)] = score_ceil;
1295 }
1296
88f063b4 1297 targetCharCount = find(head, tgt[j-1])->value;
c6871b76
KW
1298 swapScore = scores[targetCharCount * (y + 2) + swapCount] + i - targetCharCount - 1 + j - swapCount;
1299
1300 if (src[i-1] != tgt[j-1]){
1301 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));
1302 }
1303 else {
1304 swapCount = j;
1305 scores[(i+1) * (y + 2) + (j + 1)] = MIN(scores[i * (y + 2) + j], swapScore);
1306 }
1307 }
1308
88f063b4 1309 find(head, src[i-1])->value = i;
c6871b76
KW
1310 }
1311
1312 {
1313 IV score = scores[(x+1) * (y + 2) + (y + 1)];
1314 dict_free(head);
1315 Safefree(scores);
1316 return (maxDistance != 0 && maxDistance < score)?(-1):score;
1317 }
1318}
1319
1320/* END of edit_distance() stuff
1321 * ========================================================= */
1322
8e35b056
KW
1323/* is c a control character for which we have a mnemonic? */
1324#define isMNEMONIC_CNTRL(c) _IS_MNEMONIC_CNTRL_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
1325
549b4e78
KW
1326STATIC const char *
1327S_cntrl_to_mnemonic(const U8 c)
1328{
1329 /* Returns the mnemonic string that represents character 'c', if one
1330 * exists; NULL otherwise. The only ones that exist for the purposes of
1331 * this routine are a few control characters */
1332
1333 switch (c) {
1334 case '\a': return "\\a";
1335 case '\b': return "\\b";
1336 case ESC_NATIVE: return "\\e";
1337 case '\f': return "\\f";
1338 case '\n': return "\\n";
1339 case '\r': return "\\r";
1340 case '\t': return "\\t";
1341 }
1342
1343 return NULL;
1344}
1345
653099ff 1346/* Mark that we cannot extend a found fixed substring at this point.
37b6262f 1347 Update the longest found anchored substring or the longest found
653099ff
GS
1348 floating substrings if needed. */
1349
4327152a 1350STATIC void
ea3daa5d
FC
1351S_scan_commit(pTHX_ const RExC_state_t *pRExC_state, scan_data_t *data,
1352 SSize_t *minlenp, int is_inf)
c277df42 1353{
e1ec3a88 1354 const STRLEN l = CHR_SVLEN(data->last_found);
2099df82 1355 SV * const longest_sv = data->substrs[data->cur_is_floating].str;
a2ca2017 1356 const STRLEN old_l = CHR_SVLEN(longest_sv);
1de06328 1357 GET_RE_DEBUG_FLAGS_DECL;
b81d288d 1358
7918f24d
NC
1359 PERL_ARGS_ASSERT_SCAN_COMMIT;
1360
c277df42 1361 if ((l >= old_l) && ((l > old_l) || (data->flags & SF_BEFORE_EOL))) {
2099df82 1362 const U8 i = data->cur_is_floating;
a2ca2017 1363 SvSetMagicSV(longest_sv, data->last_found);
37b6262f
DM
1364 data->substrs[i].min_offset = l ? data->last_start_min : data->pos_min;
1365
2099df82
DM
1366 if (!i) /* fixed */
1367 data->substrs[0].max_offset = data->substrs[0].min_offset;
1368 else { /* float */
1369 data->substrs[1].max_offset = (l
646e8787
DM
1370 ? data->last_start_max
1371 : (data->pos_delta > SSize_t_MAX - data->pos_min
ea3daa5d
FC
1372 ? SSize_t_MAX
1373 : data->pos_min + data->pos_delta));
1374 if (is_inf
2099df82
DM
1375 || (STRLEN)data->substrs[1].max_offset > (STRLEN)SSize_t_MAX)
1376 data->substrs[1].max_offset = SSize_t_MAX;
37b6262f 1377 }
11683ecb 1378
37b6262f
DM
1379 if (data->flags & SF_BEFORE_EOL)
1380 data->substrs[i].flags |= (data->flags & SF_BEFORE_EOL);
1381 else
1382 data->substrs[i].flags &= ~SF_BEFORE_EOL;
1383 data->substrs[i].minlenp = minlenp;
1384 data->substrs[i].lookbehind = 0;
c277df42 1385 }
37b6262f 1386
c277df42 1387 SvCUR_set(data->last_found, 0);
0eda9292 1388 {
a28509cc 1389 SV * const sv = data->last_found;
097eb12c
AL
1390 if (SvUTF8(sv) && SvMAGICAL(sv)) {
1391 MAGIC * const mg = mg_find(sv, PERL_MAGIC_utf8);
1392 if (mg)
1393 mg->mg_len = 0;
1394 }
0eda9292 1395 }
c277df42
IZ
1396 data->last_end = -1;
1397 data->flags &= ~SF_BEFORE_EOL;
f5a36d78 1398 DEBUG_STUDYDATA("commit", data, 0, is_inf);
c277df42
IZ
1399}
1400
cdd87c1d
KW
1401/* An SSC is just a regnode_charclass_posix with an extra field: the inversion
1402 * list that describes which code points it matches */
1403
653099ff 1404STATIC void
3420edd7 1405S_ssc_anything(pTHX_ regnode_ssc *ssc)
653099ff 1406{
cdd87c1d
KW
1407 /* Set the SSC 'ssc' to match an empty string or any code point */
1408
557bd3fb 1409 PERL_ARGS_ASSERT_SSC_ANYTHING;
7918f24d 1410
71068078 1411 assert(is_ANYOF_SYNTHETIC(ssc));
3fffb88a 1412
0854ea0b
KW
1413 /* mortalize so won't leak */
1414 ssc->invlist = sv_2mortal(_add_range_to_invlist(NULL, 0, UV_MAX));
93e92956 1415 ANYOF_FLAGS(ssc) |= SSC_MATCHES_EMPTY_STRING; /* Plus matches empty */
653099ff
GS
1416}
1417
653099ff 1418STATIC int
dc3bf405 1419S_ssc_is_anything(const regnode_ssc *ssc)
653099ff 1420{
c144baaa
KW
1421 /* Returns TRUE if the SSC 'ssc' can match the empty string and any code
1422 * point; FALSE otherwise. Thus, this is used to see if using 'ssc' buys
1423 * us anything: if the function returns TRUE, 'ssc' hasn't been restricted
1424 * in any way, so there's no point in using it */
cdd87c1d
KW
1425
1426 UV start, end;
1427 bool ret;
653099ff 1428
557bd3fb 1429 PERL_ARGS_ASSERT_SSC_IS_ANYTHING;
7918f24d 1430
71068078 1431 assert(is_ANYOF_SYNTHETIC(ssc));
cdd87c1d 1432
93e92956 1433 if (! (ANYOF_FLAGS(ssc) & SSC_MATCHES_EMPTY_STRING)) {
cdd87c1d
KW
1434 return FALSE;
1435 }
1436
1437 /* See if the list consists solely of the range 0 - Infinity */
1438 invlist_iterinit(ssc->invlist);
1439 ret = invlist_iternext(ssc->invlist, &start, &end)
1440 && start == 0
1441 && end == UV_MAX;
1442
1443 invlist_iterfinish(ssc->invlist);
1444
1445 if (ret) {
1446 return TRUE;
1447 }
1448
1449 /* If e.g., both \w and \W are set, matches everything */
e0e1be5f 1450 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
cdd87c1d
KW
1451 int i;
1452 for (i = 0; i < ANYOF_POSIXL_MAX; i += 2) {
1453 if (ANYOF_POSIXL_TEST(ssc, i) && ANYOF_POSIXL_TEST(ssc, i+1)) {
1454 return TRUE;
1455 }
1456 }
1457 }
1458
1459 return FALSE;
653099ff
GS
1460}
1461
653099ff 1462STATIC void
cdd87c1d 1463S_ssc_init(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc)
653099ff 1464{
cdd87c1d
KW
1465 /* Initializes the SSC 'ssc'. This includes setting it to match an empty
1466 * string, any code point, or any posix class under locale */
1467
557bd3fb 1468 PERL_ARGS_ASSERT_SSC_INIT;
7918f24d 1469
557bd3fb 1470 Zero(ssc, 1, regnode_ssc);
71068078 1471 set_ANYOF_SYNTHETIC(ssc);
93e92956 1472 ARG_SET(ssc, ANYOF_ONLY_HAS_BITMAP);
3420edd7 1473 ssc_anything(ssc);
cdd87c1d 1474
2f306ab9
KW
1475 /* If any portion of the regex is to operate under locale rules that aren't
1476 * fully known at compile time, initialization includes it. The reason
1477 * this isn't done for all regexes is that the optimizer was written under
1478 * the assumption that locale was all-or-nothing. Given the complexity and
1479 * lack of documentation in the optimizer, and that there are inadequate
1480 * test cases for locale, many parts of it may not work properly, it is
1481 * safest to avoid locale unless necessary. */
cdd87c1d
KW
1482 if (RExC_contains_locale) {
1483 ANYOF_POSIXL_SETALL(ssc);
cdd87c1d
KW
1484 }
1485 else {
1486 ANYOF_POSIXL_ZERO(ssc);
1487 }
653099ff
GS
1488}
1489
b423522f 1490STATIC int
dc3bf405
BF
1491S_ssc_is_cp_posixl_init(const RExC_state_t *pRExC_state,
1492 const regnode_ssc *ssc)
b423522f
KW
1493{
1494 /* Returns TRUE if the SSC 'ssc' is in its initial state with regard only
1495 * to the list of code points matched, and locale posix classes; hence does
1496 * not check its flags) */
1497
1498 UV start, end;
1499 bool ret;
1500
1501 PERL_ARGS_ASSERT_SSC_IS_CP_POSIXL_INIT;
1502
71068078 1503 assert(is_ANYOF_SYNTHETIC(ssc));
b423522f
KW
1504
1505 invlist_iterinit(ssc->invlist);
1506 ret = invlist_iternext(ssc->invlist, &start, &end)
1507 && start == 0
1508 && end == UV_MAX;
1509
1510 invlist_iterfinish(ssc->invlist);
1511
1512 if (! ret) {
1513 return FALSE;
1514 }
1515
e0e1be5f 1516 if (RExC_contains_locale && ! ANYOF_POSIXL_SSC_TEST_ALL_SET(ssc)) {
31f05a37 1517 return FALSE;
b423522f
KW
1518 }
1519
1520 return TRUE;
1521}
1522
1523STATIC SV*
1524S_get_ANYOF_cp_list_for_ssc(pTHX_ const RExC_state_t *pRExC_state,
5c0f85ef 1525 const regnode_charclass* const node)
b423522f
KW
1526{
1527 /* Returns a mortal inversion list defining which code points are matched
1528 * by 'node', which is of type ANYOF. Handles complementing the result if
1529 * appropriate. If some code points aren't knowable at this time, the
31f05a37
KW
1530 * returned list must, and will, contain every code point that is a
1531 * possibility. */
b423522f 1532
e2506fa7 1533 SV* invlist = NULL;
1ee208c4 1534 SV* only_utf8_locale_invlist = NULL;
b423522f
KW
1535 unsigned int i;
1536 const U32 n = ARG(node);
31f05a37 1537 bool new_node_has_latin1 = FALSE;
b423522f
KW
1538
1539 PERL_ARGS_ASSERT_GET_ANYOF_CP_LIST_FOR_SSC;
1540
1541 /* Look at the data structure created by S_set_ANYOF_arg() */
93e92956 1542 if (n != ANYOF_ONLY_HAS_BITMAP) {
b423522f
KW
1543 SV * const rv = MUTABLE_SV(RExC_rxi->data->data[n]);
1544 AV * const av = MUTABLE_AV(SvRV(rv));
1545 SV **const ary = AvARRAY(av);
1546 assert(RExC_rxi->data->what[n] == 's');
1547
1548 if (ary[1] && ary[1] != &PL_sv_undef) { /* Has compile-time swash */
28118b9c 1549 invlist = sv_2mortal(invlist_clone(_get_swash_invlist(ary[1]), NULL));
b423522f
KW
1550 }
1551 else if (ary[0] && ary[0] != &PL_sv_undef) {
1552
1553 /* Here, no compile-time swash, and there are things that won't be
1554 * known until runtime -- we have to assume it could be anything */
e2506fa7 1555 invlist = sv_2mortal(_new_invlist(1));
b423522f
KW
1556 return _add_range_to_invlist(invlist, 0, UV_MAX);
1557 }
1ee208c4 1558 else if (ary[3] && ary[3] != &PL_sv_undef) {
b423522f
KW
1559
1560 /* Here no compile-time swash, and no run-time only data. Use the
1561 * node's inversion list */
28118b9c 1562 invlist = sv_2mortal(invlist_clone(ary[3], NULL));
1ee208c4
KW
1563 }
1564
1565 /* Get the code points valid only under UTF-8 locales */
037715a6 1566 if ((ANYOF_FLAGS(node) & ANYOFL_FOLD)
1ee208c4
KW
1567 && ary[2] && ary[2] != &PL_sv_undef)
1568 {
1569 only_utf8_locale_invlist = ary[2];
b423522f
KW
1570 }
1571 }
1572
e2506fa7
KW
1573 if (! invlist) {
1574 invlist = sv_2mortal(_new_invlist(0));
1575 }
1576
dcb20b36
KW
1577 /* An ANYOF node contains a bitmap for the first NUM_ANYOF_CODE_POINTS
1578 * code points, and an inversion list for the others, but if there are code
1579 * points that should match only conditionally on the target string being
1580 * UTF-8, those are placed in the inversion list, and not the bitmap.
1581 * Since there are circumstances under which they could match, they are
1582 * included in the SSC. But if the ANYOF node is to be inverted, we have
1583 * to exclude them here, so that when we invert below, the end result
1584 * actually does include them. (Think about "\xe0" =~ /[^\xc0]/di;). We
1585 * have to do this here before we add the unconditionally matched code
1586 * points */
b423522f
KW
1587 if (ANYOF_FLAGS(node) & ANYOF_INVERT) {
1588 _invlist_intersection_complement_2nd(invlist,
1589 PL_UpperLatin1,
1590 &invlist);
1591 }
1592
1593 /* Add in the points from the bit map */
dcb20b36 1594 for (i = 0; i < NUM_ANYOF_CODE_POINTS; i++) {
b423522f 1595 if (ANYOF_BITMAP_TEST(node, i)) {
6f8848d5
TC
1596 unsigned int start = i++;
1597
1598 for (; i < NUM_ANYOF_CODE_POINTS && ANYOF_BITMAP_TEST(node, i); ++i) {
1599 /* empty */
1600 }
1601 invlist = _add_range_to_invlist(invlist, start, i-1);
31f05a37 1602 new_node_has_latin1 = TRUE;
b423522f
KW
1603 }
1604 }
1605
1606 /* If this can match all upper Latin1 code points, have to add them
ac33c516
KW
1607 * as well. But don't add them if inverting, as when that gets done below,
1608 * it would exclude all these characters, including the ones it shouldn't
1609 * that were added just above */
1610 if (! (ANYOF_FLAGS(node) & ANYOF_INVERT) && OP(node) == ANYOFD
f240c685
KW
1611 && (ANYOF_FLAGS(node) & ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER))
1612 {
b423522f
KW
1613 _invlist_union(invlist, PL_UpperLatin1, &invlist);
1614 }
1615
1616 /* Similarly for these */
93e92956 1617 if (ANYOF_FLAGS(node) & ANYOF_MATCHES_ALL_ABOVE_BITMAP) {
e0a1ff7a 1618 _invlist_union_complement_2nd(invlist, PL_InBitmap, &invlist);
b423522f
KW
1619 }
1620
1621 if (ANYOF_FLAGS(node) & ANYOF_INVERT) {
1622 _invlist_invert(invlist);
1623 }
037715a6 1624 else if (new_node_has_latin1 && ANYOF_FLAGS(node) & ANYOFL_FOLD) {
31f05a37
KW
1625
1626 /* Under /li, any 0-255 could fold to any other 0-255, depending on the
1627 * locale. We can skip this if there are no 0-255 at all. */
1628 _invlist_union(invlist, PL_Latin1, &invlist);
1629 }
1630
1ee208c4
KW
1631 /* Similarly add the UTF-8 locale possible matches. These have to be
1632 * deferred until after the non-UTF-8 locale ones are taken care of just
1633 * above, or it leads to wrong results under ANYOF_INVERT */
1634 if (only_utf8_locale_invlist) {
31f05a37 1635 _invlist_union_maybe_complement_2nd(invlist,
1ee208c4 1636 only_utf8_locale_invlist,
31f05a37
KW
1637 ANYOF_FLAGS(node) & ANYOF_INVERT,
1638 &invlist);
1639 }
b423522f
KW
1640
1641 return invlist;
1642}
1643
1051e1c4 1644/* These two functions currently do the exact same thing */
557bd3fb 1645#define ssc_init_zero ssc_init
653099ff 1646
cdd87c1d
KW
1647#define ssc_add_cp(ssc, cp) ssc_add_range((ssc), (cp), (cp))
1648#define ssc_match_all_cp(ssc) ssc_add_range(ssc, 0, UV_MAX)
1649
557bd3fb 1650/* 'AND' a given class with another one. Can create false positives. 'ssc'
93e92956
KW
1651 * should not be inverted. 'and_with->flags & ANYOF_MATCHES_POSIXL' should be
1652 * 0 if 'and_with' is a regnode_charclass instead of a regnode_ssc. */
cdd87c1d 1653
653099ff 1654STATIC void
b423522f 1655S_ssc_and(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc,
7dcac5f6 1656 const regnode_charclass *and_with)
653099ff 1657{
cdd87c1d
KW
1658 /* Accumulate into SSC 'ssc' its 'AND' with 'and_with', which is either
1659 * another SSC or a regular ANYOF class. Can create false positives. */
40d049e4 1660
a0dd4231
KW
1661 SV* anded_cp_list;
1662 U8 anded_flags;
1e6ade67 1663
cdd87c1d 1664 PERL_ARGS_ASSERT_SSC_AND;
653099ff 1665
71068078 1666 assert(is_ANYOF_SYNTHETIC(ssc));
a0dd4231
KW
1667
1668 /* 'and_with' is used as-is if it too is an SSC; otherwise have to extract
1669 * the code point inversion list and just the relevant flags */
71068078 1670 if (is_ANYOF_SYNTHETIC(and_with)) {
7dcac5f6 1671 anded_cp_list = ((regnode_ssc *)and_with)->invlist;
a0dd4231 1672 anded_flags = ANYOF_FLAGS(and_with);
e9b08962
KW
1673
1674 /* XXX This is a kludge around what appears to be deficiencies in the
1675 * optimizer. If we make S_ssc_anything() add in the WARN_SUPER flag,
1676 * there are paths through the optimizer where it doesn't get weeded
1677 * out when it should. And if we don't make some extra provision for
1678 * it like the code just below, it doesn't get added when it should.
1679 * This solution is to add it only when AND'ing, which is here, and
1680 * only when what is being AND'ed is the pristine, original node
1681 * matching anything. Thus it is like adding it to ssc_anything() but
1682 * only when the result is to be AND'ed. Probably the same solution
1683 * could be adopted for the same problem we have with /l matching,
1684 * which is solved differently in S_ssc_init(), and that would lead to
1685 * fewer false positives than that solution has. But if this solution
1686 * creates bugs, the consequences are only that a warning isn't raised
1687 * that should be; while the consequences for having /l bugs is
1688 * incorrect matches */
7dcac5f6 1689 if (ssc_is_anything((regnode_ssc *)and_with)) {
f240c685 1690 anded_flags |= ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER;
e9b08962 1691 }
a0dd4231
KW
1692 }
1693 else {
5c0f85ef 1694 anded_cp_list = get_ANYOF_cp_list_for_ssc(pRExC_state, and_with);
f240c685
KW
1695 if (OP(and_with) == ANYOFD) {
1696 anded_flags = ANYOF_FLAGS(and_with) & ANYOF_COMMON_FLAGS;
1697 }
1698 else {
1699 anded_flags = ANYOF_FLAGS(and_with)
1700 &( ANYOF_COMMON_FLAGS
108316fb
KW
1701 |ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER
1702 |ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP);
d1c40ef5
KW
1703 if (ANYOFL_UTF8_LOCALE_REQD(ANYOF_FLAGS(and_with))) {
1704 anded_flags &=
1705 ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
1706 }
f240c685 1707 }
a0dd4231
KW
1708 }
1709
1710 ANYOF_FLAGS(ssc) &= anded_flags;
cdd87c1d
KW
1711
1712 /* Below, C1 is the list of code points in 'ssc'; P1, its posix classes.
1713 * C2 is the list of code points in 'and-with'; P2, its posix classes.
1714 * 'and_with' may be inverted. When not inverted, we have the situation of
1715 * computing:
1716 * (C1 | P1) & (C2 | P2)
1717 * = (C1 & (C2 | P2)) | (P1 & (C2 | P2))
1718 * = ((C1 & C2) | (C1 & P2)) | ((P1 & C2) | (P1 & P2))
1719 * <= ((C1 & C2) | P2)) | ( P1 | (P1 & P2))
1720 * <= ((C1 & C2) | P1 | P2)
1721 * Alternatively, the last few steps could be:
1722 * = ((C1 & C2) | (C1 & P2)) | ((P1 & C2) | (P1 & P2))
1723 * <= ((C1 & C2) | C1 ) | ( C2 | (P1 & P2))
1724 * <= (C1 | C2 | (P1 & P2))
1725 * We favor the second approach if either P1 or P2 is non-empty. This is
1726 * because these components are a barrier to doing optimizations, as what
1727 * they match cannot be known until the moment of matching as they are
1728 * dependent on the current locale, 'AND"ing them likely will reduce or
1729 * eliminate them.
1730 * But we can do better if we know that C1,P1 are in their initial state (a
1731 * frequent occurrence), each matching everything:
1732 * (<everything>) & (C2 | P2) = C2 | P2
1733 * Similarly, if C2,P2 are in their initial state (again a frequent
1734 * occurrence), the result is a no-op
1735 * (C1 | P1) & (<everything>) = C1 | P1
1736 *
1737 * Inverted, we have
1738 * (C1 | P1) & ~(C2 | P2) = (C1 | P1) & (~C2 & ~P2)
1739 * = (C1 & (~C2 & ~P2)) | (P1 & (~C2 & ~P2))
1740 * <= (C1 & ~C2) | (P1 & ~P2)
1741 * */
1aa99e6b 1742
a0dd4231 1743 if ((ANYOF_FLAGS(and_with) & ANYOF_INVERT)
71068078 1744 && ! is_ANYOF_SYNTHETIC(and_with))
a0dd4231 1745 {
cdd87c1d 1746 unsigned int i;
8951c461 1747
cdd87c1d
KW
1748 ssc_intersection(ssc,
1749 anded_cp_list,
1750 FALSE /* Has already been inverted */
1751 );
c6b76537 1752
cdd87c1d
KW
1753 /* If either P1 or P2 is empty, the intersection will be also; can skip
1754 * the loop */
93e92956 1755 if (! (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL)) {
cdd87c1d
KW
1756 ANYOF_POSIXL_ZERO(ssc);
1757 }
e0e1be5f 1758 else if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
cdd87c1d
KW
1759
1760 /* Note that the Posix class component P from 'and_with' actually
1761 * looks like:
1762 * P = Pa | Pb | ... | Pn
1763 * where each component is one posix class, such as in [\w\s].
1764 * Thus
1765 * ~P = ~(Pa | Pb | ... | Pn)
1766 * = ~Pa & ~Pb & ... & ~Pn
1767 * <= ~Pa | ~Pb | ... | ~Pn
1768 * The last is something we can easily calculate, but unfortunately
1769 * is likely to have many false positives. We could do better
1770 * in some (but certainly not all) instances if two classes in
1771 * P have known relationships. For example
1772 * :lower: <= :alpha: <= :alnum: <= \w <= :graph: <= :print:
1773 * So
1774 * :lower: & :print: = :lower:
1775 * And similarly for classes that must be disjoint. For example,
1776 * since \s and \w can have no elements in common based on rules in
1777 * the POSIX standard,
1778 * \w & ^\S = nothing
1779 * Unfortunately, some vendor locales do not meet the Posix
1780 * standard, in particular almost everything by Microsoft.
1781 * The loop below just changes e.g., \w into \W and vice versa */
1782
1ee208c4 1783 regnode_charclass_posixl temp;
cdd87c1d
KW
1784 int add = 1; /* To calculate the index of the complement */
1785
b1234259 1786 Zero(&temp, 1, regnode_charclass_posixl);
cdd87c1d
KW
1787 ANYOF_POSIXL_ZERO(&temp);
1788 for (i = 0; i < ANYOF_MAX; i++) {
1789 assert(i % 2 != 0
7dcac5f6
KW
1790 || ! ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i)
1791 || ! ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i + 1));
cdd87c1d 1792
7dcac5f6 1793 if (ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i)) {
cdd87c1d
KW
1794 ANYOF_POSIXL_SET(&temp, i + add);
1795 }
1796 add = 0 - add; /* 1 goes to -1; -1 goes to 1 */
1797 }
1798 ANYOF_POSIXL_AND(&temp, ssc);
c6b76537 1799
cdd87c1d
KW
1800 } /* else ssc already has no posixes */
1801 } /* else: Not inverted. This routine is a no-op if 'and_with' is an SSC
1802 in its initial state */
71068078 1803 else if (! is_ANYOF_SYNTHETIC(and_with)
7dcac5f6 1804 || ! ssc_is_cp_posixl_init(pRExC_state, (regnode_ssc *)and_with))
cdd87c1d
KW
1805 {
1806 /* But if 'ssc' is in its initial state, the result is just 'and_with';
1807 * copy it over 'ssc' */
1808 if (ssc_is_cp_posixl_init(pRExC_state, ssc)) {
71068078 1809 if (is_ANYOF_SYNTHETIC(and_with)) {
cdd87c1d
KW
1810 StructCopy(and_with, ssc, regnode_ssc);
1811 }
1812 else {
1813 ssc->invlist = anded_cp_list;
1814 ANYOF_POSIXL_ZERO(ssc);
93e92956 1815 if (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL) {
7dcac5f6 1816 ANYOF_POSIXL_OR((regnode_charclass_posixl*) and_with, ssc);
cdd87c1d
KW
1817 }
1818 }
1819 }
e0e1be5f 1820 else if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)
93e92956 1821 || (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL))
cdd87c1d
KW
1822 {
1823 /* One or the other of P1, P2 is non-empty. */
93e92956 1824 if (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL) {
1ea8b7fe
KW
1825 ANYOF_POSIXL_AND((regnode_charclass_posixl*) and_with, ssc);
1826 }
cdd87c1d
KW
1827 ssc_union(ssc, anded_cp_list, FALSE);
1828 }
1829 else { /* P1 = P2 = empty */
1830 ssc_intersection(ssc, anded_cp_list, FALSE);
1831 }
137165a6 1832 }
653099ff
GS
1833}
1834
653099ff 1835STATIC void
cdd87c1d 1836S_ssc_or(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc,
7dcac5f6 1837 const regnode_charclass *or_with)
653099ff 1838{
cdd87c1d
KW
1839 /* Accumulate into SSC 'ssc' its 'OR' with 'or_with', which is either
1840 * another SSC or a regular ANYOF class. Can create false positives if
1841 * 'or_with' is to be inverted. */
7918f24d 1842
a0dd4231
KW
1843 SV* ored_cp_list;
1844 U8 ored_flags;
c6b76537 1845
cdd87c1d 1846 PERL_ARGS_ASSERT_SSC_OR;
c6b76537 1847
71068078 1848 assert(is_ANYOF_SYNTHETIC(ssc));
a0dd4231
KW
1849
1850 /* 'or_with' is used as-is if it too is an SSC; otherwise have to extract
1851 * the code point inversion list and just the relevant flags */
71068078 1852 if (is_ANYOF_SYNTHETIC(or_with)) {
7dcac5f6 1853 ored_cp_list = ((regnode_ssc*) or_with)->invlist;
a0dd4231
KW
1854 ored_flags = ANYOF_FLAGS(or_with);
1855 }
1856 else {
5c0f85ef 1857 ored_cp_list = get_ANYOF_cp_list_for_ssc(pRExC_state, or_with);
eff8b7dc 1858 ored_flags = ANYOF_FLAGS(or_with) & ANYOF_COMMON_FLAGS;
f240c685
KW
1859 if (OP(or_with) != ANYOFD) {
1860 ored_flags
1861 |= ANYOF_FLAGS(or_with)
108316fb
KW
1862 & ( ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER
1863 |ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP);
d1c40ef5
KW
1864 if (ANYOFL_UTF8_LOCALE_REQD(ANYOF_FLAGS(or_with))) {
1865 ored_flags |=
1866 ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
1867 }
f240c685 1868 }
a0dd4231
KW
1869 }
1870
1871 ANYOF_FLAGS(ssc) |= ored_flags;
cdd87c1d
KW
1872
1873 /* Below, C1 is the list of code points in 'ssc'; P1, its posix classes.
1874 * C2 is the list of code points in 'or-with'; P2, its posix classes.
1875 * 'or_with' may be inverted. When not inverted, we have the simple
1876 * situation of computing:
1877 * (C1 | P1) | (C2 | P2) = (C1 | C2) | (P1 | P2)
1878 * If P1|P2 yields a situation with both a class and its complement are
1879 * set, like having both \w and \W, this matches all code points, and we
1880 * can delete these from the P component of the ssc going forward. XXX We
1881 * might be able to delete all the P components, but I (khw) am not certain
1882 * about this, and it is better to be safe.
1883 *
1884 * Inverted, we have
1885 * (C1 | P1) | ~(C2 | P2) = (C1 | P1) | (~C2 & ~P2)
1886 * <= (C1 | P1) | ~C2
1887 * <= (C1 | ~C2) | P1
1888 * (which results in actually simpler code than the non-inverted case)
1889 * */
9826f543 1890
a0dd4231 1891 if ((ANYOF_FLAGS(or_with) & ANYOF_INVERT)
71068078 1892 && ! is_ANYOF_SYNTHETIC(or_with))
a0dd4231 1893 {
cdd87c1d 1894 /* We ignore P2, leaving P1 going forward */
1ea8b7fe 1895 } /* else Not inverted */
93e92956 1896 else if (ANYOF_FLAGS(or_with) & ANYOF_MATCHES_POSIXL) {
7dcac5f6 1897 ANYOF_POSIXL_OR((regnode_charclass_posixl*)or_with, ssc);
e0e1be5f 1898 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
cdd87c1d
KW
1899 unsigned int i;
1900 for (i = 0; i < ANYOF_MAX; i += 2) {
1901 if (ANYOF_POSIXL_TEST(ssc, i) && ANYOF_POSIXL_TEST(ssc, i + 1))
1902 {
1903 ssc_match_all_cp(ssc);
1904 ANYOF_POSIXL_CLEAR(ssc, i);
1905 ANYOF_POSIXL_CLEAR(ssc, i+1);
cdd87c1d
KW
1906 }
1907 }
1908 }
1aa99e6b 1909 }
cdd87c1d
KW
1910
1911 ssc_union(ssc,
1912 ored_cp_list,
1913 FALSE /* Already has been inverted */
1914 );
653099ff
GS
1915}
1916
b423522f
KW
1917PERL_STATIC_INLINE void
1918S_ssc_union(pTHX_ regnode_ssc *ssc, SV* const invlist, const bool invert2nd)
1919{
1920 PERL_ARGS_ASSERT_SSC_UNION;
1921
71068078 1922 assert(is_ANYOF_SYNTHETIC(ssc));
b423522f
KW
1923
1924 _invlist_union_maybe_complement_2nd(ssc->invlist,
1925 invlist,
1926 invert2nd,
1927 &ssc->invlist);
1928}
1929
1930PERL_STATIC_INLINE void
1931S_ssc_intersection(pTHX_ regnode_ssc *ssc,
1932 SV* const invlist,
1933 const bool invert2nd)
1934{
1935 PERL_ARGS_ASSERT_SSC_INTERSECTION;
1936
71068078 1937 assert(is_ANYOF_SYNTHETIC(ssc));
b423522f
KW
1938
1939 _invlist_intersection_maybe_complement_2nd(ssc->invlist,
1940 invlist,
1941 invert2nd,
1942 &ssc->invlist);
1943}
1944
1945PERL_STATIC_INLINE void
1946S_ssc_add_range(pTHX_ regnode_ssc *ssc, const UV start, const UV end)
1947{
1948 PERL_ARGS_ASSERT_SSC_ADD_RANGE;
1949
71068078 1950 assert(is_ANYOF_SYNTHETIC(ssc));
b423522f
KW
1951
1952 ssc->invlist = _add_range_to_invlist(ssc->invlist, start, end);
1953}
1954
1955PERL_STATIC_INLINE void
1956S_ssc_cp_and(pTHX_ regnode_ssc *ssc, const UV cp)
1957{
1958 /* AND just the single code point 'cp' into the SSC 'ssc' */
1959
1960 SV* cp_list = _new_invlist(2);
1961
1962 PERL_ARGS_ASSERT_SSC_CP_AND;
1963
71068078 1964 assert(is_ANYOF_SYNTHETIC(ssc));
b423522f
KW
1965
1966 cp_list = add_cp_to_invlist(cp_list, cp);
1967 ssc_intersection(ssc, cp_list,
1968 FALSE /* Not inverted */
1969 );
1970 SvREFCNT_dec_NN(cp_list);
1971}
1972
1973PERL_STATIC_INLINE void
dc3bf405 1974S_ssc_clear_locale(regnode_ssc *ssc)
b423522f
KW
1975{
1976 /* Set the SSC 'ssc' to not match any locale things */
b423522f
KW
1977 PERL_ARGS_ASSERT_SSC_CLEAR_LOCALE;
1978
71068078 1979 assert(is_ANYOF_SYNTHETIC(ssc));
b423522f
KW
1980
1981 ANYOF_POSIXL_ZERO(ssc);
1982 ANYOF_FLAGS(ssc) &= ~ANYOF_LOCALE_FLAGS;
1983}
1984
b35552de
KW
1985#define NON_OTHER_COUNT NON_OTHER_COUNT_FOR_USE_ONLY_BY_REGCOMP_DOT_C
1986
1987STATIC bool
1988S_is_ssc_worth_it(const RExC_state_t * pRExC_state, const regnode_ssc * ssc)
1989{
1990 /* The synthetic start class is used to hopefully quickly winnow down
1991 * places where a pattern could start a match in the target string. If it
1992 * doesn't really narrow things down that much, there isn't much point to
1993 * having the overhead of using it. This function uses some very crude
1994 * heuristics to decide if to use the ssc or not.
1995 *
1996 * It returns TRUE if 'ssc' rules out more than half what it considers to
1997 * be the "likely" possible matches, but of course it doesn't know what the
1998 * actual things being matched are going to be; these are only guesses
1999 *
2000 * For /l matches, it assumes that the only likely matches are going to be
2001 * in the 0-255 range, uniformly distributed, so half of that is 127
2002 * For /a and /d matches, it assumes that the likely matches will be just
2003 * the ASCII range, so half of that is 63
2004 * For /u and there isn't anything matching above the Latin1 range, it
2005 * assumes that that is the only range likely to be matched, and uses
2006 * half that as the cut-off: 127. If anything matches above Latin1,
2007 * it assumes that all of Unicode could match (uniformly), except for
2008 * non-Unicode code points and things in the General Category "Other"
2009 * (unassigned, private use, surrogates, controls and formats). This
2010 * is a much large number. */
2011
b35552de
KW
2012 U32 count = 0; /* Running total of number of code points matched by
2013 'ssc' */
2014 UV start, end; /* Start and end points of current range in inversion
2015 list */
72400949
KW
2016 const U32 max_code_points = (LOC)
2017 ? 256
2018 : (( ! UNI_SEMANTICS
2019 || invlist_highest(ssc->invlist) < 256)
2020 ? 128
2021 : NON_OTHER_COUNT);
2022 const U32 max_match = max_code_points / 2;
b35552de
KW
2023
2024 PERL_ARGS_ASSERT_IS_SSC_WORTH_IT;
2025
2026 invlist_iterinit(ssc->invlist);
2027 while (invlist_iternext(ssc->invlist, &start, &end)) {
72400949
KW
2028 if (start >= max_code_points) {
2029 break;
b35552de 2030 }
72400949 2031 end = MIN(end, max_code_points - 1);
b35552de 2032 count += end - start + 1;
72400949 2033 if (count >= max_match) {
b35552de
KW
2034 invlist_iterfinish(ssc->invlist);
2035 return FALSE;
2036 }
2037 }
2038
2039 return TRUE;
2040}
2041
2042
b423522f
KW
2043STATIC void
2044S_ssc_finalize(pTHX_ RExC_state_t *pRExC_state, regnode_ssc *ssc)
2045{
2046 /* The inversion list in the SSC is marked mortal; now we need a more
2047 * permanent copy, which is stored the same way that is done in a regular
dcb20b36
KW
2048 * ANYOF node, with the first NUM_ANYOF_CODE_POINTS code points in a bit
2049 * map */
b423522f 2050
28118b9c 2051 SV* invlist = invlist_clone(ssc->invlist, NULL);
b423522f
KW
2052
2053 PERL_ARGS_ASSERT_SSC_FINALIZE;
2054
71068078 2055 assert(is_ANYOF_SYNTHETIC(ssc));
b423522f 2056
a0dd4231 2057 /* The code in this file assumes that all but these flags aren't relevant
93e92956
KW
2058 * to the SSC, except SSC_MATCHES_EMPTY_STRING, which should be cleared
2059 * by the time we reach here */
f240c685
KW
2060 assert(! (ANYOF_FLAGS(ssc)
2061 & ~( ANYOF_COMMON_FLAGS
108316fb
KW
2062 |ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER
2063 |ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP)));
a0dd4231 2064
b423522f
KW
2065 populate_ANYOF_from_invlist( (regnode *) ssc, &invlist);
2066
1ee208c4
KW
2067 set_ANYOF_arg(pRExC_state, (regnode *) ssc, invlist,
2068 NULL, NULL, NULL, FALSE);
b423522f 2069
85c8e306
KW
2070 /* Make sure is clone-safe */
2071 ssc->invlist = NULL;
2072
e0e1be5f 2073 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
93e92956 2074 ANYOF_FLAGS(ssc) |= ANYOF_MATCHES_POSIXL;
d156f5cb 2075 OP(ssc) = ANYOFPOSIXL;
e0e1be5f 2076 }
d156f5cb 2077 else if (RExC_contains_locale) {
b2e90ddf
KW
2078 OP(ssc) = ANYOFL;
2079 }
2080
1462525b 2081 assert(! (ANYOF_FLAGS(ssc) & ANYOF_LOCALE_FLAGS) || RExC_contains_locale);
b423522f
KW
2082}
2083
a3621e74
YO
2084#define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
2085#define TRIE_LIST_CUR(state) ( TRIE_LIST_ITEM( state, 0 ).forid )
2086#define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
538e84ed
KW
2087#define TRIE_LIST_USED(idx) ( trie->states[state].trans.list \
2088 ? (TRIE_LIST_CUR( idx ) - 1) \
2089 : 0 )
a3621e74 2090
3dab1dad
YO
2091
2092#ifdef DEBUGGING
07be1b83 2093/*
2b8b4781
NC
2094 dump_trie(trie,widecharmap,revcharmap)
2095 dump_trie_interim_list(trie,widecharmap,revcharmap,next_alloc)
2096 dump_trie_interim_table(trie,widecharmap,revcharmap,next_alloc)
3dab1dad
YO
2097
2098 These routines dump out a trie in a somewhat readable format.
07be1b83
YO
2099 The _interim_ variants are used for debugging the interim
2100 tables that are used to generate the final compressed
2101 representation which is what dump_trie expects.
2102
486ec47a 2103 Part of the reason for their existence is to provide a form
3dab1dad 2104 of documentation as to how the different representations function.
07be1b83
YO
2105
2106*/
3dab1dad
YO
2107
2108/*
3dab1dad
YO
2109 Dumps the final compressed table form of the trie to Perl_debug_log.
2110 Used for debugging make_trie().
2111*/
b9a59e08 2112
3dab1dad 2113STATIC void
2b8b4781
NC
2114S_dump_trie(pTHX_ const struct _reg_trie_data *trie, HV *widecharmap,
2115 AV *revcharmap, U32 depth)
3dab1dad
YO
2116{
2117 U32 state;
ab3bbdeb 2118 SV *sv=sv_newmortal();
55eed653 2119 int colwidth= widecharmap ? 6 : 4;
2e64971a 2120 U16 word;
3dab1dad
YO
2121 GET_RE_DEBUG_FLAGS_DECL;
2122
7918f24d 2123 PERL_ARGS_ASSERT_DUMP_TRIE;
ab3bbdeb 2124
6ad9a8ab 2125 Perl_re_indentf( aTHX_ "Char : %-6s%-6s%-4s ",
1e37780e 2126 depth+1, "Match","Base","Ofs" );
3dab1dad
YO
2127
2128 for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
2b8b4781 2129 SV ** const tmp = av_fetch( revcharmap, state, 0);
3dab1dad 2130 if ( tmp ) {
6ad9a8ab 2131 Perl_re_printf( aTHX_ "%*s",
ab3bbdeb 2132 colwidth,
538e84ed 2133 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
ab3bbdeb
YO
2134 PL_colors[0], PL_colors[1],
2135 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
538e84ed
KW
2136 PERL_PV_ESCAPE_FIRSTCHAR
2137 )
ab3bbdeb 2138 );
3dab1dad
YO
2139 }
2140 }
1e37780e
YO
2141 Perl_re_printf( aTHX_ "\n");
2142 Perl_re_indentf( aTHX_ "State|-----------------------", depth+1);
3dab1dad
YO
2143
2144 for( state = 0 ; state < trie->uniquecharcount ; state++ )
6ad9a8ab
YO
2145 Perl_re_printf( aTHX_ "%.*s", colwidth, "--------");
2146 Perl_re_printf( aTHX_ "\n");
3dab1dad 2147
1e2e3d02 2148 for( state = 1 ; state < trie->statecount ; state++ ) {
be8e71aa 2149 const U32 base = trie->states[ state ].trans.base;
3dab1dad 2150
147e3846 2151 Perl_re_indentf( aTHX_ "#%4" UVXf "|", depth+1, (UV)state);
3dab1dad
YO
2152
2153 if ( trie->states[ state ].wordnum ) {
1e37780e 2154 Perl_re_printf( aTHX_ " W%4X", trie->states[ state ].wordnum );
3dab1dad 2155 } else {
6ad9a8ab 2156 Perl_re_printf( aTHX_ "%6s", "" );
3dab1dad
YO
2157 }
2158
147e3846 2159 Perl_re_printf( aTHX_ " @%4" UVXf " ", (UV)base );
3dab1dad
YO
2160
2161 if ( base ) {
2162 U32 ofs = 0;
2163
2164 while( ( base + ofs < trie->uniquecharcount ) ||
2165 ( base + ofs - trie->uniquecharcount < trie->lasttrans
538e84ed
KW
2166 && trie->trans[ base + ofs - trie->uniquecharcount ].check
2167 != state))
3dab1dad
YO
2168 ofs++;
2169
147e3846 2170 Perl_re_printf( aTHX_ "+%2" UVXf "[ ", (UV)ofs);
3dab1dad
YO
2171
2172 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
538e84ed
KW
2173 if ( ( base + ofs >= trie->uniquecharcount )
2174 && ( base + ofs - trie->uniquecharcount
2175 < trie->lasttrans )
2176 && trie->trans[ base + ofs
2177 - trie->uniquecharcount ].check == state )
3dab1dad 2178 {
147e3846 2179 Perl_re_printf( aTHX_ "%*" UVXf, colwidth,
1e37780e
YO
2180 (UV)trie->trans[ base + ofs - trie->uniquecharcount ].next
2181 );
3dab1dad 2182 } else {
88f063b4 2183 Perl_re_printf( aTHX_ "%*s", colwidth," ." );
3dab1dad
YO
2184 }
2185 }
2186
6ad9a8ab 2187 Perl_re_printf( aTHX_ "]");
3dab1dad
YO
2188
2189 }
6ad9a8ab 2190 Perl_re_printf( aTHX_ "\n" );
3dab1dad 2191 }
6ad9a8ab 2192 Perl_re_indentf( aTHX_ "word_info N:(prev,len)=",
cb41e5d6 2193 depth);
2e64971a 2194 for (word=1; word <= trie->wordcount; word++) {
6ad9a8ab 2195 Perl_re_printf( aTHX_ " %d:(%d,%d)",
2e64971a
DM
2196 (int)word, (int)(trie->wordinfo[word].prev),
2197 (int)(trie->wordinfo[word].len));
2198 }
6ad9a8ab 2199 Perl_re_printf( aTHX_ "\n" );
538e84ed 2200}
3dab1dad 2201/*
3dab1dad 2202 Dumps a fully constructed but uncompressed trie in list form.
538e84ed 2203 List tries normally only are used for construction when the number of
3dab1dad
YO
2204 possible chars (trie->uniquecharcount) is very high.
2205 Used for debugging make_trie().
2206*/
2207STATIC void
55eed653 2208S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie,
2b8b4781
NC
2209 HV *widecharmap, AV *revcharmap, U32 next_alloc,
2210 U32 depth)
3dab1dad
YO
2211{
2212 U32 state;
ab3bbdeb 2213 SV *sv=sv_newmortal();
55eed653 2214 int colwidth= widecharmap ? 6 : 4;
3dab1dad 2215 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
2216
2217 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_LIST;
2218
3dab1dad 2219 /* print out the table precompression. */
6ad9a8ab 2220 Perl_re_indentf( aTHX_ "State :Word | Transition Data\n",
cb41e5d6 2221 depth+1 );
6ad9a8ab 2222 Perl_re_indentf( aTHX_ "%s",
cb41e5d6 2223 depth+1, "------:-----+-----------------\n" );
538e84ed 2224
3dab1dad
YO
2225 for( state=1 ; state < next_alloc ; state ++ ) {
2226 U16 charid;
538e84ed 2227
147e3846 2228 Perl_re_indentf( aTHX_ " %4" UVXf " :",
cb41e5d6 2229 depth+1, (UV)state );
3dab1dad 2230 if ( ! trie->states[ state ].wordnum ) {
6ad9a8ab 2231 Perl_re_printf( aTHX_ "%5s| ","");
3dab1dad 2232 } else {
6ad9a8ab 2233 Perl_re_printf( aTHX_ "W%4x| ",
3dab1dad
YO
2234 trie->states[ state ].wordnum
2235 );
2236 }
2237 for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
538e84ed 2238 SV ** const tmp = av_fetch( revcharmap,
88f063b4 2239 TRIE_LIST_ITEM(state, charid).forid, 0);
ab3bbdeb 2240 if ( tmp ) {
147e3846 2241 Perl_re_printf( aTHX_ "%*s:%3X=%4" UVXf " | ",
ab3bbdeb 2242 colwidth,
538e84ed
KW
2243 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp),
2244 colwidth,
2245 PL_colors[0], PL_colors[1],
2246 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0)
2247 | PERL_PV_ESCAPE_FIRSTCHAR
ab3bbdeb 2248 ) ,
88f063b4
KW
2249 TRIE_LIST_ITEM(state, charid).forid,
2250 (UV)TRIE_LIST_ITEM(state, charid).newstate
1e2e3d02 2251 );
538e84ed 2252 if (!(charid % 10))
6ad9a8ab 2253 Perl_re_printf( aTHX_ "\n%*s| ",
664e119d 2254 (int)((depth * 2) + 14), "");
1e2e3d02 2255 }
ab3bbdeb 2256 }
6ad9a8ab 2257 Perl_re_printf( aTHX_ "\n");
3dab1dad 2258 }
538e84ed 2259}
3dab1dad
YO
2260
2261/*
3dab1dad 2262 Dumps a fully constructed but uncompressed trie in table form.
538e84ed
KW
2263 This is the normal DFA style state transition table, with a few
2264 twists to facilitate compression later.
3dab1dad
YO
2265 Used for debugging make_trie().
2266*/
2267STATIC void
55eed653 2268S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie,
2b8b4781
NC
2269 HV *widecharmap, AV *revcharmap, U32 next_alloc,
2270 U32 depth)
3dab1dad
YO
2271{
2272 U32 state;
2273 U16 charid;
ab3bbdeb 2274 SV *sv=sv_newmortal();
55eed653 2275 int colwidth= widecharmap ? 6 : 4;
3dab1dad 2276 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
2277
2278 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_TABLE;
538e84ed 2279
3dab1dad
YO
2280 /*
2281 print out the table precompression so that we can do a visual check
2282 that they are identical.
2283 */
538e84ed 2284
6ad9a8ab 2285 Perl_re_indentf( aTHX_ "Char : ", depth+1 );
3dab1dad
YO
2286
2287 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
2b8b4781 2288 SV ** const tmp = av_fetch( revcharmap, charid, 0);
3dab1dad 2289 if ( tmp ) {
6ad9a8ab 2290 Perl_re_printf( aTHX_ "%*s",
ab3bbdeb 2291 colwidth,
538e84ed 2292 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
ab3bbdeb
YO
2293 PL_colors[0], PL_colors[1],
2294 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
538e84ed
KW
2295 PERL_PV_ESCAPE_FIRSTCHAR
2296 )
ab3bbdeb 2297 );
3dab1dad
YO
2298 }
2299 }
2300
4aaafc03
YO
2301 Perl_re_printf( aTHX_ "\n");
2302 Perl_re_indentf( aTHX_ "State+-", depth+1 );
3dab1dad
YO
2303
2304 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
6ad9a8ab 2305 Perl_re_printf( aTHX_ "%.*s", colwidth,"--------");
3dab1dad
YO
2306 }
2307
6ad9a8ab 2308 Perl_re_printf( aTHX_ "\n" );
3dab1dad
YO
2309
2310 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
2311
147e3846 2312 Perl_re_indentf( aTHX_ "%4" UVXf " : ",
cb41e5d6 2313 depth+1,
3dab1dad
YO
2314 (UV)TRIE_NODENUM( state ) );
2315
2316 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
ab3bbdeb
YO
2317 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
2318 if (v)
147e3846 2319 Perl_re_printf( aTHX_ "%*" UVXf, colwidth, v );
ab3bbdeb 2320 else
6ad9a8ab 2321 Perl_re_printf( aTHX_ "%*s", colwidth, "." );
3dab1dad
YO
2322 }
2323 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
147e3846 2324 Perl_re_printf( aTHX_ " (%4" UVXf ")\n",
538e84ed 2325 (UV)trie->trans[ state ].check );
3dab1dad 2326 } else {
147e3846 2327 Perl_re_printf( aTHX_ " (%4" UVXf ") W%4X\n",
538e84ed 2328 (UV)trie->trans[ state ].check,
3dab1dad
YO
2329 trie->states[ TRIE_NODENUM( state ) ].wordnum );
2330 }
2331 }
07be1b83 2332}
3dab1dad
YO
2333
2334#endif
2335
2e64971a 2336
786e8c11
YO
2337/* make_trie(startbranch,first,last,tail,word_count,flags,depth)
2338 startbranch: the first branch in the whole branch sequence
2339 first : start branch of sequence of branch-exact nodes.
2340 May be the same as startbranch
2341 last : Thing following the last branch.
2342 May be the same as tail.
2343 tail : item following the branch sequence
2344 count : words in the sequence
a4525e78 2345 flags : currently the OP() type we will be building one of /EXACT(|F|FA|FU|FU_SS|L|FLU8)/
786e8c11 2346 depth : indent depth
3dab1dad 2347
786e8c11 2348Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
07be1b83 2349
786e8c11
YO
2350A trie is an N'ary tree where the branches are determined by digital
2351decomposition of the key. IE, at the root node you look up the 1st character and
2352follow that branch repeat until you find the end of the branches. Nodes can be
2353marked as "accepting" meaning they represent a complete word. Eg:
07be1b83 2354
786e8c11 2355 /he|she|his|hers/
72f13be8 2356
786e8c11
YO
2357would convert into the following structure. Numbers represent states, letters
2358following numbers represent valid transitions on the letter from that state, if
2359the number is in square brackets it represents an accepting state, otherwise it
2360will be in parenthesis.
07be1b83 2361
786e8c11
YO
2362 +-h->+-e->[3]-+-r->(8)-+-s->[9]
2363 | |
2364 | (2)
2365 | |
2366 (1) +-i->(6)-+-s->[7]
2367 |
2368 +-s->(3)-+-h->(4)-+-e->[5]
07be1b83 2369
786e8c11
YO
2370 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
2371
2372This shows that when matching against the string 'hers' we will begin at state 1
2373read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
2374then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
2375is also accepting. Thus we know that we can match both 'he' and 'hers' with a
2376single traverse. We store a mapping from accepting to state to which word was
2377matched, and then when we have multiple possibilities we try to complete the
b8fda935 2378rest of the regex in the order in which they occurred in the alternation.
786e8c11
YO
2379
2380The only prior NFA like behaviour that would be changed by the TRIE support is
2381the silent ignoring of duplicate alternations which are of the form:
2382
2383 / (DUPE|DUPE) X? (?{ ... }) Y /x
2384
4b714af6 2385Thus EVAL blocks following a trie may be called a different number of times with
786e8c11 2386and without the optimisation. With the optimisations dupes will be silently
486ec47a 2387ignored. This inconsistent behaviour of EVAL type nodes is well established as
786e8c11
YO
2388the following demonstrates:
2389
2390 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
2391
2392which prints out 'word' three times, but
2393
2394 'words'=~/(word|word|word)(?{ print $1 })S/
2395
2396which doesnt print it out at all. This is due to other optimisations kicking in.
2397
2398Example of what happens on a structural level:
2399
486ec47a 2400The regexp /(ac|ad|ab)+/ will produce the following debug output:
786e8c11
YO
2401
2402 1: CURLYM[1] {1,32767}(18)
2403 5: BRANCH(8)
2404 6: EXACT <ac>(16)
2405 8: BRANCH(11)
2406 9: EXACT <ad>(16)
2407 11: BRANCH(14)
2408 12: EXACT <ab>(16)
2409 16: SUCCEED(0)
2410 17: NOTHING(18)
2411 18: END(0)
2412
2413This would be optimizable with startbranch=5, first=5, last=16, tail=16
2414and should turn into:
2415
2416 1: CURLYM[1] {1,32767}(18)
2417 5: TRIE(16)
2418 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
2419 <ac>
2420 <ad>
2421 <ab>
2422 16: SUCCEED(0)
2423 17: NOTHING(18)
2424 18: END(0)
2425
2426Cases where tail != last would be like /(?foo|bar)baz/:
2427
2428 1: BRANCH(4)
2429 2: EXACT <foo>(8)
2430 4: BRANCH(7)
2431 5: EXACT <bar>(8)
2432 7: TAIL(8)
2433 8: EXACT <baz>(10)
2434 10: END(0)
2435
2436which would be optimizable with startbranch=1, first=1, last=7, tail=8
2437and would end up looking like:
2438
2439 1: TRIE(8)
2440 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
2441 <foo>
2442 <bar>
2443 7: TAIL(8)
2444 8: EXACT <baz>(10)
2445 10: END(0)
2446
c80e42f3 2447 d = uvchr_to_utf8_flags(d, uv, 0);
786e8c11
YO
2448
2449is the recommended Unicode-aware way of saying
2450
2451 *(d++) = uv;
2452*/
2453
fab2782b 2454#define TRIE_STORE_REVCHAR(val) \
786e8c11 2455 STMT_START { \
73031816 2456 if (UTF) { \
668fcfea 2457 SV *zlopp = newSV(UTF8_MAXBYTES); \
88c9ea1e 2458 unsigned char *flrbbbbb = (unsigned char *) SvPVX(zlopp); \
c80e42f3 2459 unsigned const char *const kapow = uvchr_to_utf8(flrbbbbb, val); \
73031816
NC
2460 SvCUR_set(zlopp, kapow - flrbbbbb); \
2461 SvPOK_on(zlopp); \
2462 SvUTF8_on(zlopp); \
2463 av_push(revcharmap, zlopp); \
2464 } else { \
fab2782b 2465 char ooooff = (char)val; \
73031816
NC
2466 av_push(revcharmap, newSVpvn(&ooooff, 1)); \
2467 } \
2468 } STMT_END
786e8c11 2469
914a25d5
KW
2470/* This gets the next character from the input, folding it if not already
2471 * folded. */
2472#define TRIE_READ_CHAR STMT_START { \
2473 wordlen++; \
2474 if ( UTF ) { \
2475 /* if it is UTF then it is either already folded, or does not need \
2476 * folding */ \
1c1d615a 2477 uvc = valid_utf8_to_uvchr( (const U8*) uc, &len); \
914a25d5
KW
2478 } \
2479 else if (folder == PL_fold_latin1) { \
7d006b13
KW
2480 /* This folder implies Unicode rules, which in the range expressible \
2481 * by not UTF is the lower case, with the two exceptions, one of \
2482 * which should have been taken care of before calling this */ \
2483 assert(*uc != LATIN_SMALL_LETTER_SHARP_S); \
2484 uvc = toLOWER_L1(*uc); \
2485 if (UNLIKELY(uvc == MICRO_SIGN)) uvc = GREEK_SMALL_LETTER_MU; \
2486 len = 1; \
914a25d5
KW
2487 } else { \
2488 /* raw data, will be folded later if needed */ \
2489 uvc = (U32)*uc; \
2490 len = 1; \
2491 } \
786e8c11
YO
2492} STMT_END
2493
2494
2495
2496#define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
2497 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
00195859 2498 U32 ging = TRIE_LIST_LEN( state ) * 2; \
f9003953 2499 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
00195859 2500 TRIE_LIST_LEN( state ) = ging; \
786e8c11
YO
2501 } \
2502 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
2503 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
2504 TRIE_LIST_CUR( state )++; \
2505} STMT_END
07be1b83 2506
786e8c11 2507#define TRIE_LIST_NEW(state) STMT_START { \
d09f14bf 2508 Newx( trie->states[ state ].trans.list, \
786e8c11
YO
2509 4, reg_trie_trans_le ); \
2510 TRIE_LIST_CUR( state ) = 1; \
2511 TRIE_LIST_LEN( state ) = 4; \
2512} STMT_END
07be1b83 2513
786e8c11
YO
2514#define TRIE_HANDLE_WORD(state) STMT_START { \
2515 U16 dupe= trie->states[ state ].wordnum; \
2516 regnode * const noper_next = regnext( noper ); \
2517 \
786e8c11
YO
2518 DEBUG_r({ \
2519 /* store the word for dumping */ \
2520 SV* tmp; \
2521 if (OP(noper) != NOTHING) \
740cce10 2522 tmp = newSVpvn_utf8(STRING(noper), STR_LEN(noper), UTF); \
786e8c11 2523 else \
740cce10 2524 tmp = newSVpvn_utf8( "", 0, UTF ); \
2b8b4781 2525 av_push( trie_words, tmp ); \
786e8c11
YO
2526 }); \
2527 \
2528 curword++; \
2e64971a
DM
2529 trie->wordinfo[curword].prev = 0; \
2530 trie->wordinfo[curword].len = wordlen; \
2531 trie->wordinfo[curword].accept = state; \
786e8c11
YO
2532 \
2533 if ( noper_next < tail ) { \
2534 if (!trie->jump) \
538e84ed
KW
2535 trie->jump = (U16 *) PerlMemShared_calloc( word_count + 1, \
2536 sizeof(U16) ); \
7f69552c 2537 trie->jump[curword] = (U16)(noper_next - convert); \
786e8c11
YO
2538 if (!jumper) \
2539 jumper = noper_next; \
2540 if (!nextbranch) \
2541 nextbranch= regnext(cur); \
2542 } \
2543 \
2544 if ( dupe ) { \
2e64971a
DM
2545 /* It's a dupe. Pre-insert into the wordinfo[].prev */\
2546 /* chain, so that when the bits of chain are later */\
2547 /* linked together, the dups appear in the chain */\
2548 trie->wordinfo[curword].prev = trie->wordinfo[dupe].prev; \
2549 trie->wordinfo[dupe].prev = curword; \
786e8c11
YO
2550 } else { \
2551 /* we haven't inserted this word yet. */ \
2552 trie->states[ state ].wordnum = curword; \
2553 } \
2554} STMT_END
07be1b83 2555
3dab1dad 2556
786e8c11
YO
2557#define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
2558 ( ( base + charid >= ucharcount \
2559 && base + charid < ubound \
2560 && state == trie->trans[ base - ucharcount + charid ].check \
2561 && trie->trans[ base - ucharcount + charid ].next ) \
2562 ? trie->trans[ base - ucharcount + charid ].next \
2563 : ( state==1 ? special : 0 ) \
2564 )
3dab1dad 2565
8bcafbf4
YO
2566#define TRIE_BITMAP_SET_FOLDED(trie, uvc, folder) \
2567STMT_START { \
2568 TRIE_BITMAP_SET(trie, uvc); \
2569 /* store the folded codepoint */ \
2570 if ( folder ) \
2571 TRIE_BITMAP_SET(trie, folder[(U8) uvc ]); \
2572 \
2573 if ( !UTF ) { \
2574 /* store first byte of utf8 representation of */ \
2575 /* variant codepoints */ \
2576 if (! UVCHR_IS_INVARIANT(uvc)) { \
2577 TRIE_BITMAP_SET(trie, UTF8_TWO_BYTE_HI(uvc)); \
2578 } \
2579 } \
2580} STMT_END
786e8c11
YO
2581#define MADE_TRIE 1
2582#define MADE_JUMP_TRIE 2
2583#define MADE_EXACT_TRIE 4
3dab1dad 2584
a3621e74 2585STATIC I32
538e84ed
KW
2586S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch,
2587 regnode *first, regnode *last, regnode *tail,
2588 U32 word_count, U32 flags, U32 depth)
a3621e74
YO
2589{
2590 /* first pass, loop through and scan words */
2591 reg_trie_data *trie;
55eed653 2592 HV *widecharmap = NULL;
2b8b4781 2593 AV *revcharmap = newAV();
a3621e74 2594 regnode *cur;
a3621e74
YO
2595 STRLEN len = 0;
2596 UV uvc = 0;
2597 U16 curword = 0;
2598 U32 next_alloc = 0;
786e8c11
YO
2599 regnode *jumper = NULL;
2600 regnode *nextbranch = NULL;
7f69552c 2601 regnode *convert = NULL;
2e64971a 2602 U32 *prev_states; /* temp array mapping each state to previous one */
a3621e74 2603 /* we just use folder as a flag in utf8 */
1e696034 2604 const U8 * folder = NULL;
a3621e74 2605
3a611511
YO
2606 /* in the below add_data call we are storing either 'tu' or 'tuaa'
2607 * which stands for one trie structure, one hash, optionally followed
2608 * by two arrays */
2b8b4781 2609#ifdef DEBUGGING
3a611511 2610 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("tuaa"));
2b8b4781
NC
2611 AV *trie_words = NULL;
2612 /* along with revcharmap, this only used during construction but both are
2613 * useful during debugging so we store them in the struct when debugging.
8e11feef 2614 */
2b8b4781 2615#else
cf78de0b 2616 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("tu"));
3dab1dad 2617 STRLEN trie_charcount=0;
3dab1dad 2618#endif
2b8b4781 2619 SV *re_trie_maxbuff;
a3621e74 2620 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
2621
2622 PERL_ARGS_ASSERT_MAKE_TRIE;
72f13be8
YO
2623#ifndef DEBUGGING
2624 PERL_UNUSED_ARG(depth);
2625#endif
a3621e74 2626
1e696034 2627 switch (flags) {
a4525e78 2628 case EXACT: case EXACTL: break;
89829bb5 2629 case EXACTFAA:
fab2782b 2630 case EXACTFU_SS:
a4525e78
KW
2631 case EXACTFU:
2632 case EXACTFLU8: folder = PL_fold_latin1; break;
1e696034 2633 case EXACTF: folder = PL_fold; break;
fab2782b 2634 default: Perl_croak( aTHX_ "panic! In trie construction, unknown node type %u %s", (unsigned) flags, PL_reg_name[flags] );
1e696034
KW
2635 }
2636
c944940b 2637 trie = (reg_trie_data *) PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
a3621e74 2638 trie->refcount = 1;
3dab1dad 2639 trie->startstate = 1;
786e8c11 2640 trie->wordcount = word_count;
f8fc2ecf 2641 RExC_rxi->data->data[ data_slot ] = (void*)trie;
c944940b 2642 trie->charmap = (U16 *) PerlMemShared_calloc( 256, sizeof(U16) );
a4525e78 2643 if (flags == EXACT || flags == EXACTL)
c944940b 2644 trie->bitmap = (char *) PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
2e64971a
DM
2645 trie->wordinfo = (reg_trie_wordinfo *) PerlMemShared_calloc(
2646 trie->wordcount+1, sizeof(reg_trie_wordinfo));
2647
a3621e74 2648 DEBUG_r({
2b8b4781 2649 trie_words = newAV();
a3621e74 2650 });
a3621e74 2651
0111c4fd 2652 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
316ebaf2 2653 assert(re_trie_maxbuff);
a3621e74 2654 if (!SvIOK(re_trie_maxbuff)) {
0111c4fd 2655 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
a3621e74 2656 }
df826430 2657 DEBUG_TRIE_COMPILE_r({
6ad9a8ab 2658 Perl_re_indentf( aTHX_
cb41e5d6
YO
2659 "make_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
2660 depth+1,
88f063b4 2661 REG_NODE_NUM(startbranch), REG_NODE_NUM(first),
538e84ed 2662 REG_NODE_NUM(last), REG_NODE_NUM(tail), (int)depth);
3dab1dad 2663 });
538e84ed 2664
7f69552c
YO
2665 /* Find the node we are going to overwrite */
2666 if ( first == startbranch && OP( last ) != BRANCH ) {
2667 /* whole branch chain */
2668 convert = first;
2669 } else {
2670 /* branch sub-chain */
2671 convert = NEXTOPER( first );
2672 }
538e84ed 2673
a3621e74
YO
2674 /* -- First loop and Setup --
2675
2676 We first traverse the branches and scan each word to determine if it
2677 contains widechars, and how many unique chars there are, this is
2678 important as we have to build a table with at least as many columns as we
2679 have unique chars.
2680
2681 We use an array of integers to represent the character codes 0..255
538e84ed
KW
2682 (trie->charmap) and we use a an HV* to store Unicode characters. We use
2683 the native representation of the character value as the key and IV's for
2684 the coded index.
a3621e74
YO
2685
2686 *TODO* If we keep track of how many times each character is used we can
2687 remap the columns so that the table compression later on is more
3b753521 2688 efficient in terms of memory by ensuring the most common value is in the
a3621e74
YO
2689 middle and the least common are on the outside. IMO this would be better
2690 than a most to least common mapping as theres a decent chance the most
2691 common letter will share a node with the least common, meaning the node
486ec47a 2692 will not be compressible. With a middle is most common approach the worst
a3621e74
YO
2693 case is when we have the least common nodes twice.
2694
2695 */
2696
a3621e74 2697 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
df826430 2698 regnode *noper = NEXTOPER( cur );
944e05e3
YO
2699 const U8 *uc;
2700 const U8 *e;
bc031a7d 2701 int foldlen = 0;
07be1b83 2702 U32 wordlen = 0; /* required init */
bc031a7d
KW
2703 STRLEN minchars = 0;
2704 STRLEN maxchars = 0;
538e84ed
KW
2705 bool set_bit = trie->bitmap ? 1 : 0; /*store the first char in the
2706 bitmap?*/
a3621e74 2707
3dab1dad 2708 if (OP(noper) == NOTHING) {
20ed8c88
YO
2709 /* skip past a NOTHING at the start of an alternation
2710 * eg, /(?:)a|(?:b)/ should be the same as /a|b/
2711 */
df826430 2712 regnode *noper_next= regnext(noper);
944e05e3
YO
2713 if (noper_next < tail)
2714 noper= noper_next;
2715 }
2716
dca5fc4c
YO
2717 if ( noper < tail &&
2718 (
2719 OP(noper) == flags ||
2720 (
2721 flags == EXACTFU &&
2722 OP(noper) == EXACTFU_SS
2723 )
2724 )
2725 ) {
944e05e3
YO
2726 uc= (U8*)STRING(noper);
2727 e= uc + STR_LEN(noper);
2728 } else {
2729 trie->minlen= 0;
2730 continue;
3dab1dad 2731 }
df826430 2732
944e05e3 2733
fab2782b 2734 if ( set_bit ) { /* bitmap only alloced when !(UTF&&Folding) */
02daf0ab
YO
2735 TRIE_BITMAP_SET(trie,*uc); /* store the raw first byte
2736 regardless of encoding */
fab2782b
YO
2737 if (OP( noper ) == EXACTFU_SS) {
2738 /* false positives are ok, so just set this */
0dc4a61d 2739 TRIE_BITMAP_SET(trie, LATIN_SMALL_LETTER_SHARP_S);
fab2782b
YO
2740 }
2741 }
dca5fc4c 2742
bc031a7d
KW
2743 for ( ; uc < e ; uc += len ) { /* Look at each char in the current
2744 branch */
3dab1dad 2745 TRIE_CHARCOUNT(trie)++;
a3621e74 2746 TRIE_READ_CHAR;
645de4ce 2747
bc031a7d
KW
2748 /* TRIE_READ_CHAR returns the current character, or its fold if /i
2749 * is in effect. Under /i, this character can match itself, or
2750 * anything that folds to it. If not under /i, it can match just
2751 * itself. Most folds are 1-1, for example k, K, and KELVIN SIGN
2752 * all fold to k, and all are single characters. But some folds
2753 * expand to more than one character, so for example LATIN SMALL
2754 * LIGATURE FFI folds to the three character sequence 'ffi'. If
2755 * the string beginning at 'uc' is 'ffi', it could be matched by
2756 * three characters, or just by the one ligature character. (It
2757 * could also be matched by two characters: LATIN SMALL LIGATURE FF
2758 * followed by 'i', or by 'f' followed by LATIN SMALL LIGATURE FI).
2759 * (Of course 'I' and/or 'F' instead of 'i' and 'f' can also
2760 * match.) The trie needs to know the minimum and maximum number
2761 * of characters that could match so that it can use size alone to
2762 * quickly reject many match attempts. The max is simple: it is
2763 * the number of folded characters in this branch (since a fold is
2764 * never shorter than what folds to it. */
2765
2766 maxchars++;
2767
2768 /* And the min is equal to the max if not under /i (indicated by
2769 * 'folder' being NULL), or there are no multi-character folds. If
2770 * there is a multi-character fold, the min is incremented just
2771 * once, for the character that folds to the sequence. Each
2772 * character in the sequence needs to be added to the list below of
2773 * characters in the trie, but we count only the first towards the
2774 * min number of characters needed. This is done through the
2775 * variable 'foldlen', which is returned by the macros that look
2776 * for these sequences as the number of bytes the sequence
2777 * occupies. Each time through the loop, we decrement 'foldlen' by
2778 * how many bytes the current char occupies. Only when it reaches
2779 * 0 do we increment 'minchars' or look for another multi-character
2780 * sequence. */
2781 if (folder == NULL) {
2782 minchars++;
2783 }
2784 else if (foldlen > 0) {
2785 foldlen -= (UTF) ? UTF8SKIP(uc) : 1;
645de4ce
KW
2786 }
2787 else {
bc031a7d
KW
2788 minchars++;
2789
2790 /* See if *uc is the beginning of a multi-character fold. If
2791 * so, we decrement the length remaining to look at, to account
2792 * for the current character this iteration. (We can use 'uc'
2793 * instead of the fold returned by TRIE_READ_CHAR because for
2794 * non-UTF, the latin1_safe macro is smart enough to account
2795 * for all the unfolded characters, and because for UTF, the
2796 * string will already have been folded earlier in the
2797 * compilation process */
2798 if (UTF) {
2799 if ((foldlen = is_MULTI_CHAR_FOLD_utf8_safe(uc, e))) {
2800 foldlen -= UTF8SKIP(uc);
645de4ce
KW
2801 }
2802 }
bc031a7d
KW
2803 else if ((foldlen = is_MULTI_CHAR_FOLD_latin1_safe(uc, e))) {
2804 foldlen--;
2805 }
645de4ce 2806 }
bc031a7d
KW
2807
2808 /* The current character (and any potential folds) should be added
2809 * to the possible matching characters for this position in this
2810 * branch */
a3621e74 2811 if ( uvc < 256 ) {
fab2782b
YO
2812 if ( folder ) {
2813 U8 folded= folder[ (U8) uvc ];
2814 if ( !trie->charmap[ folded ] ) {
2815 trie->charmap[ folded ]=( ++trie->uniquecharcount );
2816 TRIE_STORE_REVCHAR( folded );
2817 }
2818 }
a3621e74
YO
2819 if ( !trie->charmap[ uvc ] ) {
2820 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
fab2782b 2821 TRIE_STORE_REVCHAR( uvc );
a3621e74 2822 }
02daf0ab 2823 if ( set_bit ) {
62012aee
KW
2824 /* store the codepoint in the bitmap, and its folded
2825 * equivalent. */
8bcafbf4 2826 TRIE_BITMAP_SET_FOLDED(trie, uvc, folder);
02daf0ab
YO
2827 set_bit = 0; /* We've done our bit :-) */
2828 }
a3621e74 2829 } else {
bc031a7d
KW
2830
2831 /* XXX We could come up with the list of code points that fold
2832 * to this using PL_utf8_foldclosures, except not for
2833 * multi-char folds, as there may be multiple combinations
2834 * there that could work, which needs to wait until runtime to
2835 * resolve (The comment about LIGATURE FFI above is such an
2836 * example */
2837
a3621e74 2838 SV** svpp;
55eed653
NC
2839 if ( !widecharmap )
2840 widecharmap = newHV();
a3621e74 2841
55eed653 2842 svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );
a3621e74
YO
2843
2844 if ( !svpp )
147e3846 2845 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%" UVXf, uvc );
a3621e74
YO
2846
2847 if ( !SvTRUE( *svpp ) ) {
2848 sv_setiv( *svpp, ++trie->uniquecharcount );
fab2782b 2849 TRIE_STORE_REVCHAR(uvc);
a3621e74
YO
2850 }
2851 }
bc031a7d
KW
2852 } /* end loop through characters in this branch of the trie */
2853
2854 /* We take the min and max for this branch and combine to find the min
2855 * and max for all branches processed so far */
3dab1dad 2856 if( cur == first ) {
bc031a7d
KW
2857 trie->minlen = minchars;
2858 trie->maxlen = maxchars;
2859 } else if (minchars < trie->minlen) {
2860 trie->minlen = minchars;
2861 } else if (maxchars > trie->maxlen) {
2862 trie->maxlen = maxchars;
fab2782b 2863 }
a3621e74
YO
2864 } /* end first pass */
2865 DEBUG_TRIE_COMPILE_r(
6ad9a8ab 2866 Perl_re_indentf( aTHX_
cb41e5d6
YO
2867 "TRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
2868 depth+1,
55eed653 2869 ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
be8e71aa
YO
2870 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
2871 (int)trie->minlen, (int)trie->maxlen )
a3621e74 2872 );
a3621e74
YO
2873
2874 /*
2875 We now know what we are dealing with in terms of unique chars and
2876 string sizes so we can calculate how much memory a naive
0111c4fd
RGS
2877 representation using a flat table will take. If it's over a reasonable
2878 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
a3621e74
YO
2879 conservative but potentially much slower representation using an array
2880 of lists.
2881
2882 At the end we convert both representations into the same compressed
2883 form that will be used in regexec.c for matching with. The latter
2884 is a form that cannot be used to construct with but has memory
2885 properties similar to the list form and access properties similar
2886 to the table form making it both suitable for fast searches and
2887 small enough that its feasable to store for the duration of a program.
2888
2889 See the comment in the code where the compressed table is produced
2890 inplace from the flat tabe representation for an explanation of how
2891 the compression works.
2892
2893 */
2894
2895
2e64971a
DM
2896 Newx(prev_states, TRIE_CHARCOUNT(trie) + 2, U32);
2897 prev_states[1] = 0;
2898
538e84ed
KW
2899 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1)
2900 > SvIV(re_trie_maxbuff) )
2901 {
a3621e74
YO
2902 /*
2903 Second Pass -- Array Of Lists Representation
2904
2905 Each state will be represented by a list of charid:state records
2906 (reg_trie_trans_le) the first such element holds the CUR and LEN
2907 points of the allocated array. (See defines above).
2908
2909 We build the initial structure using the lists, and then convert
2910 it into the compressed table form which allows faster lookups
2911 (but cant be modified once converted).
a3621e74
YO
2912 */
2913
a3621e74
YO
2914 STRLEN transcount = 1;
2915
6ad9a8ab 2916 DEBUG_TRIE_COMPILE_MORE_r( Perl_re_indentf( aTHX_ "Compiling trie using list compiler\n",
cb41e5d6 2917 depth+1));
686b73d4 2918
c944940b
JH
2919 trie->states = (reg_trie_state *)
2920 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
2921 sizeof(reg_trie_state) );
a3621e74
YO
2922 TRIE_LIST_NEW(1);
2923 next_alloc = 2;
2924
2925 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2926
df826430 2927 regnode *noper = NEXTOPER( cur );
c445ea15
AL
2928 U32 state = 1; /* required init */
2929 U16 charid = 0; /* sanity init */
07be1b83 2930 U32 wordlen = 0; /* required init */
c445ea15 2931
df826430
YO
2932 if (OP(noper) == NOTHING) {
2933 regnode *noper_next= regnext(noper);
944e05e3
YO
2934 if (noper_next < tail)
2935 noper= noper_next;
df826430
YO
2936 }
2937
944e05e3
YO
2938 if ( noper < tail && ( OP(noper) == flags || ( flags == EXACTFU && OP(noper) == EXACTFU_SS ) ) ) {
2939 const U8 *uc= (U8*)STRING(noper);
2940 const U8 *e= uc + STR_LEN(noper);
2941
786e8c11 2942 for ( ; uc < e ; uc += len ) {
c445ea15 2943
786e8c11 2944 TRIE_READ_CHAR;
c445ea15 2945
786e8c11
YO
2946 if ( uvc < 256 ) {
2947 charid = trie->charmap[ uvc ];
c445ea15 2948 } else {
538e84ed
KW
2949 SV** const svpp = hv_fetch( widecharmap,
2950 (char*)&uvc,
2951 sizeof( UV ),
2952 0);
786e8c11
YO
2953 if ( !svpp ) {
2954 charid = 0;
2955 } else {
2956 charid=(U16)SvIV( *svpp );
2957 }
c445ea15 2958 }
538e84ed
KW
2959 /* charid is now 0 if we dont know the char read, or
2960 * nonzero if we do */
786e8c11 2961 if ( charid ) {
a3621e74 2962
786e8c11
YO
2963 U16 check;
2964 U32 newstate = 0;
a3621e74 2965
786e8c11
YO
2966 charid--;
2967 if ( !trie->states[ state ].trans.list ) {
2968 TRIE_LIST_NEW( state );
c445ea15 2969 }
538e84ed
KW
2970 for ( check = 1;
2971 check <= TRIE_LIST_USED( state );
2972 check++ )
2973 {
2974 if ( TRIE_LIST_ITEM( state, check ).forid
2975 == charid )
2976 {
786e8c11
YO
2977 newstate = TRIE_LIST_ITEM( state, check ).newstate;
2978 break;
2979 }
2980 }
2981 if ( ! newstate ) {
2982 newstate = next_alloc++;
2e64971a 2983 prev_states[newstate] = state;
786e8c11
YO
2984 TRIE_LIST_PUSH( state, charid, newstate );
2985 transcount++;
2986 }
2987 state = newstate;
2988 } else {
147e3846 2989 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %" IVdf, uvc );
c445ea15 2990 }
a28509cc 2991 }
c445ea15 2992 }
3dab1dad 2993 TRIE_HANDLE_WORD(state);
a3621e74
YO
2994
2995 } /* end second pass */
2996
1e2e3d02 2997 /* next alloc is the NEXT state to be allocated */
538e84ed 2998 trie->statecount = next_alloc;
c944940b
JH
2999 trie->states = (reg_trie_state *)
3000 PerlMemShared_realloc( trie->states,
3001 next_alloc
3002 * sizeof(reg_trie_state) );
a3621e74 3003
3dab1dad 3004 /* and now dump it out before we compress it */
2b8b4781
NC
3005 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
3006 revcharmap, next_alloc,
3007 depth+1)
1e2e3d02 3008 );
a3621e74 3009
c944940b
JH
3010 trie->trans = (reg_trie_trans *)
3011 PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
a3621e74
YO
3012 {
3013 U32 state;
a3621e74
YO
3014 U32 tp = 0;
3015 U32 zp = 0;
3016
3017
3018 for( state=1 ; state < next_alloc ; state ++ ) {
3019 U32 base=0;
3020
3021 /*
3022 DEBUG_TRIE_COMPILE_MORE_r(
6ad9a8ab 3023 Perl_re_printf( aTHX_ "tp: %d zp: %d ",tp,zp)
a3621e74
YO
3024 );
3025 */
3026
3027 if (trie->states[state].trans.list) {
3028 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
3029 U16 maxid=minid;
a28509cc 3030 U16 idx;
a3621e74
YO
3031
3032 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
c445ea15
AL
3033 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
3034 if ( forid < minid ) {
3035 minid=forid;
3036 } else if ( forid > maxid ) {
3037 maxid=forid;
3038 }
a3621e74
YO
3039 }
3040 if ( transcount < tp + maxid - minid + 1) {
3041 transcount *= 2;
c944940b
JH
3042 trie->trans = (reg_trie_trans *)
3043 PerlMemShared_realloc( trie->trans,
446bd890
NC
3044 transcount
3045 * sizeof(reg_trie_trans) );
538e84ed
KW
3046 Zero( trie->trans + (transcount / 2),
3047 transcount / 2,
3048 reg_trie_trans );
a3621e74
YO
3049 }
3050 base = trie->uniquecharcount + tp - minid;
3051 if ( maxid == minid ) {
3052 U32 set = 0;
3053 for ( ; zp < tp ; zp++ ) {
3054 if ( ! trie->trans[ zp ].next ) {
3055 base = trie->uniquecharcount + zp - minid;
538e84ed
KW
3056 trie->trans[ zp ].next = TRIE_LIST_ITEM( state,
3057 1).newstate;
a3621e74
YO
3058 trie->trans[ zp ].check = state;
3059 set = 1;
3060 break;
3061 }
3062 }
3063 if ( !set ) {
538e84ed
KW
3064 trie->trans[ tp ].next = TRIE_LIST_ITEM( state,
3065 1).newstate;
a3621e74
YO
3066 trie->trans[ tp ].check = state;
3067 tp++;
3068 zp = tp;
3069 }
3070 } else {
3071 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
538e84ed
KW
3072 const U32 tid = base
3073 - trie->uniquecharcount
3074 + TRIE_LIST_ITEM( state, idx ).forid;
3075 trie->trans[ tid ].next = TRIE_LIST_ITEM( state,
3076 idx ).newstate;
a3621e74
YO
3077 trie->trans[ tid ].check = state;
3078 }
3079 tp += ( maxid - minid + 1 );
3080 }
3081 Safefree(trie->states[ state ].trans.list);
3082 }
3083 /*
3084 DEBUG_TRIE_COMPILE_MORE_r(
6ad9a8ab 3085 Perl_re_printf( aTHX_ " base: %d\n",base);
a3621e74
YO
3086 );
3087 */
3088 trie->states[ state ].trans.base=base;
3089 }
cc601c31 3090 trie->lasttrans = tp + 1;
a3621e74
YO
3091 }
3092 } else {
3093 /*
3094 Second Pass -- Flat Table Representation.
3095
b423522f
KW
3096 we dont use the 0 slot of either trans[] or states[] so we add 1 to
3097 each. We know that we will need Charcount+1 trans at most to store
3098 the data (one row per char at worst case) So we preallocate both
3099 structures assuming worst case.
a3621e74
YO
3100
3101 We then construct the trie using only the .next slots of the entry
3102 structs.
3103
b423522f
KW
3104 We use the .check field of the first entry of the node temporarily
3105 to make compression both faster and easier by keeping track of how
3106 many non zero fields are in the node.
a3621e74
YO
3107
3108 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
3109 transition.
3110
b423522f
KW
3111 There are two terms at use here: state as a TRIE_NODEIDX() which is
3112 a number representing the first entry of the node, and state as a
3113 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1)
3114 and TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3)
3115 if there are 2 entrys per node. eg:
a3621e74
YO
3116
3117 A B A B
3118 1. 2 4 1. 3 7
3119 2. 0 3 3. 0 5
3120 3. 0 0 5. 0 0
3121 4. 0 0 7. 0 0
3122
b423522f
KW
3123 The table is internally in the right hand, idx form. However as we
3124 also have to deal with the states array which is indexed by nodenum
3125 we have to use TRIE_NODENUM() to convert.
a3621e74
YO
3126
3127 */
6ad9a8ab 3128 DEBUG_TRIE_COMPILE_MORE_r( Perl_re_indentf( aTHX_ "Compiling trie using table compiler\n",
cb41e5d6 3129 depth+1));
3dab1dad 3130
c944940b
JH
3131 trie->trans = (reg_trie_trans *)
3132 PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
3133 * trie->uniquecharcount + 1,
3134 sizeof(reg_trie_trans) );
3135 trie->states = (reg_trie_state *)
3136 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
3137 sizeof(reg_trie_state) );
a3621e74
YO
3138 next_alloc = trie->uniquecharcount + 1;
3139
3dab1dad 3140
a3621e74
YO
3141 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
3142
df826430 3143 regnode *noper = NEXTOPER( cur );
a3621e74
YO
3144
3145 U32 state = 1; /* required init */
3146
3147 U16 charid = 0; /* sanity init */
3148 U32 accept_state = 0; /* sanity init */
a3621e74 3149
07be1b83 3150 U32 wordlen = 0; /* required init */
a3621e74 3151
df826430
YO
3152 if (OP(noper) == NOTHING) {
3153 regnode *noper_next= regnext(noper);
944e05e3
YO
3154 if (noper_next < tail)
3155 noper= noper_next;
df826430 3156 }
fab2782b 3157
944e05e3
YO
3158 if ( noper < tail && ( OP(noper) == flags || ( flags == EXACTFU && OP(noper) == EXACTFU_SS ) ) ) {
3159 const U8 *uc= (U8*)STRING(noper);
3160 const U8 *e= uc + STR_LEN(noper);
3161
786e8c11 3162 for ( ; uc < e ; uc += len ) {
a3621e74 3163
786e8c11 3164 TRIE_READ_CHAR;
a3621e74 3165
786e8c11
YO
3166 if ( uvc < 256 ) {
3167 charid = trie->charmap[ uvc ];
3168 } else {
538e84ed
KW
3169 SV* const * const svpp = hv_fetch( widecharmap,
3170 (char*)&uvc,
3171 sizeof( UV ),
3172 0);
786e8c11 3173 charid = svpp ? (U16)SvIV(*svpp) : 0;
a3621e74 3174 }
786e8c11
YO
3175 if ( charid ) {
3176 charid--;
3177 if ( !trie->trans[ state + charid ].next ) {
3178 trie->trans[ state + charid ].next = next_alloc;
3179 trie->trans[ state ].check++;
2e64971a
DM
3180 prev_states[TRIE_NODENUM(next_alloc)]
3181 = TRIE_NODENUM(state);
786e8c11
YO
3182 next_alloc += trie->uniquecharcount;
3183 }
3184 state = trie->trans[ state + charid ].next;
3185 } else {
147e3846 3186 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %" IVdf, uvc );
786e8c11 3187 }
538e84ed
KW
3188 /* charid is now 0 if we dont know the char read, or
3189 * nonzero if we do */
a3621e74 3190 }
a3621e74 3191 }
3dab1dad
YO
3192 accept_state = TRIE_NODENUM( state );
3193 TRIE_HANDLE_WORD(accept_state);
a3621e74
YO
3194
3195 } /* end second pass */
3196
3dab1dad 3197 /* and now dump it out before we compress it */
2b8b4781
NC
3198 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
3199 revcharmap,
3200 next_alloc, depth+1));
a3621e74 3201
a3621e74
YO
3202 {
3203 /*
3204 * Inplace compress the table.*
3205
3206 For sparse data sets the table constructed by the trie algorithm will
3207 be mostly 0/FAIL transitions or to put it another way mostly empty.
3208 (Note that leaf nodes will not contain any transitions.)
3209
3210 This algorithm compresses the tables by eliminating most such
3211 transitions, at the cost of a modest bit of extra work during lookup:
3212
3213 - Each states[] entry contains a .base field which indicates the
3214 index in the state[] array wheres its transition data is stored.
3215
3b753521 3216 - If .base is 0 there are no valid transitions from that node.
a3621e74
YO
3217
3218 - If .base is nonzero then charid is added to it to find an entry in
3219 the trans array.
3220
3221 -If trans[states[state].base+charid].check!=state then the
3222 transition is taken to be a 0/Fail transition. Thus if there are fail
3223 transitions at the front of the node then the .base offset will point
3224 somewhere inside the previous nodes data (or maybe even into a node
3225 even earlier), but the .check field determines if the transition is
3226 valid.
3227
786e8c11 3228 XXX - wrong maybe?
a3621e74 3229 The following process inplace converts the table to the compressed
3b753521 3230 table: We first do not compress the root node 1,and mark all its
a3621e74 3231 .check pointers as 1 and set its .base pointer as 1 as well. This
3b753521
FN
3232 allows us to do a DFA construction from the compressed table later,
3233 and ensures that any .base pointers we calculate later are greater
3234 than 0.
a3621e74
YO
3235
3236 - We set 'pos' to indicate the first entry of the second node.
3237
3238 - We then iterate over the columns of the node, finding the first and
3239 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
3240 and set the .check pointers accordingly, and advance pos
3241 appropriately and repreat for the next node. Note that when we copy
3242 the next pointers we have to convert them from the original
3243 NODEIDX form to NODENUM form as the former is not valid post
3244 compression.
3245
3246 - If a node has no transitions used we mark its base as 0 and do not
3247 advance the pos pointer.
3248
3249 - If a node only has one transition we use a second pointer into the
3250 structure to fill in allocated fail transitions from other states.
3251 This pointer is independent of the main pointer and scans forward
3252 looking for null transitions that are allocated to a state. When it
3253 finds one it writes the single transition into the "hole". If the
786e8c11 3254 pointer doesnt find one the single transition is appended as normal.
a3621e74
YO
3255
3256 - Once compressed we can Renew/realloc the structures to release the
3257 excess space.
3258
3259 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
3260 specifically Fig 3.47 and the associated pseudocode.
3261
3262 demq
3263 */
a3b680e6 3264 const U32 laststate = TRIE_NODENUM( next_alloc );
a28509cc 3265 U32 state, charid;
a3621e74 3266 U32 pos = 0, zp=0;
1e2e3d02 3267 trie->statecount = laststate;
a3621e74
YO
3268
3269 for ( state = 1 ; state < laststate ; state++ ) {
3270 U8 flag = 0;
a28509cc
AL
3271 const U32 stateidx = TRIE_NODEIDX( state );
3272 const U32 o_used = trie->trans[ stateidx ].check;
3273 U32 used = trie->trans[ stateidx ].check;
a3621e74
YO
3274 trie->trans[ stateidx ].check = 0;
3275
538e84ed
KW
3276 for ( charid = 0;
3277 used && charid < trie->uniquecharcount;
3278 charid++ )
3279 {
a3621e74
YO
3280 if ( flag || trie->trans[ stateidx + charid ].next ) {
3281 if ( trie->trans[ stateidx + charid ].next ) {
3282 if (o_used == 1) {
3283 for ( ; zp < pos ; zp++ ) {
3284 if ( ! trie->trans[ zp ].next ) {
3285 break;
3286 }
3287 }
538e84ed
KW
3288 trie->states[ state ].trans.base
3289 = zp
3290 + trie->uniquecharcount
3291 - charid ;
3292 trie->trans[ zp ].next
3293 = SAFE_TRIE_NODENUM( trie->trans[ stateidx
3294 + charid ].next );
a3621e74
YO
3295 trie->trans[ zp ].check = state;
3296 if ( ++zp > pos ) pos = zp;
3297 break;
3298 }
3299 used--;
3300 }
3301 if ( !flag ) {
3302 flag = 1;
538e84ed
KW
3303 trie->states[ state ].trans.base
3304 = pos + trie->uniquecharcount - charid ;
a3621e74 3305 }
538e84ed
KW
3306 trie->trans[ pos ].next
3307 = SAFE_TRIE_NODENUM(
3308 trie->trans[ stateidx + charid ].next );
a3621e74
YO
3309 trie->trans[ pos ].check = state;
3310 pos++;
3311 }
3312 }
3313 }
cc601c31 3314 trie->lasttrans = pos + 1;
c944940b
JH
3315 trie->states = (reg_trie_state *)
3316 PerlMemShared_realloc( trie->states, laststate
3317 * sizeof(reg_trie_state) );
a3621e74 3318 DEBUG_TRIE_COMPILE_MORE_r(
147e3846 3319 Perl_re_indentf( aTHX_ "Alloc: %d Orig: %" IVdf " elements, Final:%" IVdf ". Savings of %%%5.2f\n",
cb41e5d6 3320 depth+1,
538e84ed
KW
3321 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount
3322 + 1 ),
3323 (IV)next_alloc,
3324 (IV)pos,
3325 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
a3621e74
YO
3326 );
3327
3328 } /* end table compress */
3329 }
1e2e3d02 3330 DEBUG_TRIE_COMPILE_MORE_r(
147e3846 3331 Perl_re_indentf( aTHX_ "Statecount:%" UVxf " Lasttrans:%" UVxf "\n",
cb41e5d6 3332 depth+1,
1e2e3d02
YO
3333 (UV)trie->statecount,
3334 (UV)trie->lasttrans)
3335 );
cc601c31 3336 /* resize the trans array to remove unused space */
c944940b
JH
3337 trie->trans = (reg_trie_trans *)
3338 PerlMemShared_realloc( trie->trans, trie->lasttrans
3339 * sizeof(reg_trie_trans) );
a3621e74 3340
538e84ed 3341 { /* Modify the program and insert the new TRIE node */
3dab1dad
YO
3342 U8 nodetype =(U8)(flags & 0xFF);
3343 char *str=NULL;
538e84ed 3344
07be1b83 3345#ifdef DEBUGGING
e62cc96a 3346 regnode *optimize = NULL;
7122b237
YO
3347#ifdef RE_TRACK_PATTERN_OFFSETS
3348
b57a0404
JH
3349 U32 mjd_offset = 0;
3350 U32 mjd_nodelen = 0;
7122b237
YO
3351#endif /* RE_TRACK_PATTERN_OFFSETS */
3352#endif /* DEBUGGING */
a3621e74 3353 /*
3dab1dad
YO
3354 This means we convert either the first branch or the first Exact,
3355 depending on whether the thing following (in 'last') is a branch
3356 or not and whther first is the startbranch (ie is it a sub part of
3357 the alternation or is it the whole thing.)
3b753521 3358 Assuming its a sub part we convert the EXACT otherwise we convert
3dab1dad 3359 the whole branch sequence, including the first.
a3621e74 3360 */
3dab1dad 3361 /* Find the node we are going to overwrite */
7f69552c 3362 if ( first != startbranch || OP( last ) == BRANCH ) {
07be1b83 3363 /* branch sub-chain */
3dab1dad 3364 NEXT_OFF( first ) = (U16)(last - first);
7122b237 3365#ifdef RE_TRACK_PATTERN_OFFSETS
07be1b83
YO
3366 DEBUG_r({
3367 mjd_offset= Node_Offset((convert));
3368 mjd_nodelen= Node_Length((convert));
3369 });
7122b237 3370#endif
7f69552c 3371 /* whole branch chain */
7122b237
YO
3372 }
3373#ifdef RE_TRACK_PATTERN_OFFSETS
3374 else {
7f69552c
YO
3375 DEBUG_r({
3376 const regnode *nop = NEXTOPER( convert );
3377 mjd_offset= Node_Offset((nop));
3378 mjd_nodelen= Node_Length((nop));
3379 });
07be1b83
YO
3380 }
3381 DEBUG_OPTIMISE_r(
147e3846 3382 Perl_re_indentf( aTHX_ "MJD offset:%" UVuf " MJD length:%" UVuf "\n",
cb41e5d6 3383 depth+1,
786e8c11 3384 (UV)mjd_offset, (UV)mjd_nodelen)
07be1b83 3385 );
7122b237 3386#endif
538e84ed 3387 /* But first we check to see if there is a common prefix we can
3dab1dad
YO
3388 split out as an EXACT and put in front of the TRIE node. */
3389 trie->startstate= 1;
55eed653 3390 if ( trie->bitmap && !widecharmap && !trie->jump ) {
5ee57374
YO
3391 /* we want to find the first state that has more than
3392 * one transition, if that state is not the first state
3393 * then we have a common prefix which we can remove.
3394 */
3dab1dad 3395 U32 state;
1e2e3d02 3396 for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
a3621e74 3397 U32 ofs = 0;
ee0dfd0b
YO
3398 I32 first_ofs = -1; /* keeps track of the ofs of the first
3399 transition, -1 means none */
8e11feef
RGS
3400 U32 count = 0;
3401 const U32 base = trie->states[ state ].trans.base;
a3621e74 3402
5ee57374 3403 /* does this state terminate an alternation? */
3dab1dad 3404 if ( trie->states[state].wordnum )
8e11feef 3405 count = 1;
a3621e74 3406
8e11feef 3407 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
cc601c31
YO
3408 if ( ( base + ofs >= trie->uniquecharcount ) &&
3409 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
a3621e74
YO
3410 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
3411 {
3dab1dad 3412 if ( ++count > 1 ) {
5ee57374 3413 /* we have more than one transition */
d3d91c7c
YO
3414 SV **tmp;
3415 U8 *ch;
5ee57374
YO
3416 /* if this is the first state there is no common prefix
3417 * to extract, so we can exit */
8e11feef 3418 if ( state == 1 ) break;
d3d91c7c
YO
3419 tmp = av_fetch( revcharmap, ofs, 0);
3420 ch = (U8*)SvPV_nolen_const( *tmp );
3421
5ee57374
YO
3422 /* if we are on count 2 then we need to initialize the
3423 * bitmap, and store the previous char if there was one
3424 * in it*/
3dab1dad 3425 if ( count == 2 ) {
5ee57374 3426 /* clear the bitmap */
3dab1dad
YO
3427 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
3428 DEBUG_OPTIMISE_r(
147e3846 3429 Perl_re_indentf( aTHX_ "New Start State=%" UVuf " Class: [",
cb41e5d6 3430 depth+1,
786e8c11 3431 (UV)state));
ee0dfd0b
YO
3432