2 package CharClass::Matcher;
6 use warnings FATAL => 'all';
7 no warnings 'experimental::autoderef';
9 $Data::Dumper::Useqq= 1;
10 our $hex_fmt= "0x%02X";
15 require 'regen/regen_lib.pl';
16 require 'regen/charset_translations.pl';
17 require "regen/regcharclass_multi_char_folds.pl";
21 CharClass::Matcher -- Generate C macros that match character classes efficiently
25 perl Porting/regcharclass.pl
29 Dynamically generates macros for detecting special charclasses
30 in latin-1, utf8, and codepoint forms. Macros can be set to return
31 the length (in bytes) of the matched codepoint, and/or the codepoint itself.
33 To regenerate F<regcharclass.h>, run this script from perl-root. No arguments
36 Using WHATEVER as an example the following macros can be produced, depending
37 on the input parameters (how to get each is described by internal comments at
38 the C<__DATA__> line):
42 =item C<is_WHATEVER(s,is_utf8)>
44 =item C<is_WHATEVER_safe(s,e,is_utf8)>
46 Do a lookup as appropriate based on the C<is_utf8> flag. When possible
47 comparisons involving octect<128 are done before checking the C<is_utf8>
48 flag, hopefully saving time.
50 The version without the C<_safe> suffix should be used only when the input is
51 known to be well-formed.
53 =item C<is_WHATEVER_utf8(s)>
55 =item C<is_WHATEVER_utf8_safe(s,e)>
57 Do a lookup assuming the string is encoded in (normalized) UTF8.
59 The version without the C<_safe> suffix should be used only when the input is
60 known to be well-formed.
62 =item C<is_WHATEVER_latin1(s)>
64 =item C<is_WHATEVER_latin1_safe(s,e)>
66 Do a lookup assuming the string is encoded in latin-1 (aka plan octets).
68 The version without the C<_safe> suffix should be used only when it is known
69 that C<s> contains at least one character.
71 =item C<is_WHATEVER_cp(cp)>
73 Check to see if the string matches a given codepoint (hypothetically a
74 U32). The condition is constructed as to "break out" as early as
75 possible if the codepoint is out of range of the condition.
79 (cp==X || (cp>X && (cp==Y || (cp>Y && ...))))
81 Thus if the character is X+1 only two comparisons will be done. Making
82 matching lookups slower, but non-matching faster.
84 =item C<what_len_WHATEVER_FOO(arg1, ..., len)>
86 A variant form of each of the macro types described above can be generated, in
87 which the code point is returned by the macro, and an extra parameter (in the
88 final position) is added, which is a pointer for the macro to set the byte
89 length of the returned code point.
91 These forms all have a C<what_len> prefix instead of the C<is_>, for example
92 C<what_len_WHATEVER_safe(s,e,is_utf8,len)> and
93 C<what_len_WHATEVER_utf8(s,len)>.
95 These forms should not be used I<except> on small sets of mostly widely
96 separated code points; otherwise the code generated is inefficient. For these
97 cases, it is best to use the C<is_> forms, and then find the code point with
98 C<utf8_to_uvchr_buf>(). This program can fail with a "deep recursion"
99 message on the worst of the inappropriate sets. Examine the generated macro
100 to see if it is acceptable.
102 =item C<what_WHATEVER_FOO(arg1, ...)>
104 A variant form of each of the C<is_> macro types described above can be generated, in
105 which the code point and not the length is returned by the macro. These have
106 the same caveat as L</what_len_WHATEVER_FOO(arg1, ..., len)>, plus they should
107 not be used where the set contains a NULL, as 0 is returned for two different
108 cases: a) the set doesn't include the input code point; b) the set does
109 include it, and it is a NULL.
113 The above isn't quite complete, as for specialized purposes one can get a
114 macro like C<is_WHATEVER_utf8_no_length_checks(s)>, which assumes that it is
115 already known that there is enough space to hold the character starting at
116 C<s>, but otherwise checks that it is well-formed. In other words, this is
117 intermediary in checking between C<is_WHATEVER_utf8(s)> and
118 C<is_WHATEVER_utf8_safe(s,e)>.
122 perltidy -st -bt=1 -bbt=0 -pt=0 -sbt=1 -ce -nwls== "%f"
127 Author: Yves Orton (demerphq) 2007. Maintained by Perl5 Porters.
131 No tests directly here (although the regex engine will fail tests
132 if this code is broken). Insufficient documentation and no Getopts
133 handler for using the module as a script.
137 You may distribute under the terms of either the GNU General Public
138 License or the Artistic License, as specified in the README file.
142 # Sub naming convention:
143 # __func : private subroutine, can not be called as a method
144 # _func : private method, not meant for external use
145 # func : public method.
148 #-------------------------------------------------------------------------------
150 # ($cp,$n,$l,$u)=__uni_latin($str);
152 # Return a list of arrays, each of which when interpreted correctly
153 # represent the string in some given encoding with specific conditions.
155 # $cp - list of codepoints that make up the string.
156 # $n - list of octets that make up the string if all codepoints are invariant
157 # regardless of if the string is in UTF-8 or not.
158 # $l - list of octets that make up the string in latin1 encoding if all
159 # codepoints < 256, and at least one codepoint is UTF-8 variant.
160 # $u - list of octets that make up the string in utf8 if any codepoint is
164 #-----------+----------
165 # 0 - 127 : $n (127/128 are the values for ASCII platforms)
176 my $only_has_invariants = 1;
177 my $a2n = get_a2n($charset);
178 for my $ch ( split //, $str ) {
180 $max= $cp if $max < $cp;
186 push @cp, $a2n->[$cp];
190 $only_has_invariants = ($charset =~ /ascii/i) ? $max < 128 : $max < 160;
191 if ($only_has_invariants) {
194 $l= [@cp] if $max && $max < 256;
197 for my $ch ( split //, $str ) {
198 push @u, map { ord } split //, cp_2_utfbytes(ord $ch, $charset);
202 return ( \@cp, \@cp_high, $n, $l, $u );
206 # $clean= __clean($expr);
208 # Cleanup a ternary expression, removing unnecessary parens and apply some
209 # simplifications using regexes.
218 $parens= qr/ (?> \( (?> (?: (?> [^()]+ ) | (??{ $parens }) )* ) \) ) /x;
220 ## remove redundant parens
221 1 while $expr =~ s/ \( \s* ( $parens ) \s* \) /$1/gx;
224 # repeatedly simplify conditions like
225 # ( (cond1) ? ( (cond2) ? X : Y ) : Y )
227 # ( ( (cond1) && (cond2) ) ? X : Y )
228 # Also similarly handles expressions like:
229 # : (cond1) ? ( (cond2) ? X : Y ) : Y )
230 # Note the inclusion of the close paren in ([:()]) and the open paren in ([()]) is
231 # purely to ensure we have a balanced set of parens in the expression which makes
232 # it easier to understand the pattern in an editor that understands paren's, we do
233 # not expect either of these cases to actually fire. - Yves
239 \? \s* ($parens|[^()?:\s]+?) \s*
240 : \s* ($parens|[^()?:\s]+?) \s*
244 /$1 ( $2 && $3 ) ? $4 : $5 $6/gx;
245 #$expr=~s/\(\(U8\*\)s\)\[(\d+)\]/S$1/g if length $expr > 8000;
246 #$expr=~s/\s+//g if length $expr > 8000;
248 die "Expression too long" if length $expr > 8000;
254 # $text= __macro(@args);
255 # Join args together by newlines, and then neatly add backslashes to the end
256 # of every line as expected by the C pre-processor for #define's.
260 my $str= join "\n", @_;
262 my @lines= map { s/\s+$//; s/\t/ /g; $_ } split /\n/, $str;
263 my $last= pop @lines;
264 $str= join "\n", ( map { sprintf "%-76s\\", $_ } @lines ), $last;
265 1 while $str =~ s/^(\t*) {8}/$1\t/gm;
270 # my $op=__incrdepth($op);
272 # take an 'op' hashref and add one to it and all its childrens depths.
277 return unless ref $op;
279 __incrdepth( $op->{yes} );
280 __incrdepth( $op->{no} );
284 # join two branches of an opcode together with a condition, incrementing
285 # the depth on the yes branch when we do so.
286 # returns the new root opcode of the tree.
288 my ( $cond, $yes, $no )= @_;
292 yes => __incrdepth( $yes ),
301 no => __incrdepth($no),
311 # my $obj=CLASS->new(op=>'SOMENAME',title=>'blah',txt=>[..]);
313 # Create a new CharClass::Matcher object by parsing the text in
314 # the txt array. Currently applies the following rules:
316 # Element starts with C<0x>, line is evaled the result treated as
317 # a number which is passed to chr().
319 # Element starts with C<">, line is evaled and the result treated
322 # Each string is then stored in the 'strs' subhash as a hash record
323 # made up of the results of __uni_latin1, using the keynames
324 # 'low','latin1','utf8', as well as the synthesized 'LATIN1', 'high', and
325 # 'UTF8' which hold a merge of 'low' and their lowercase equivalents.
327 # Size data is tracked per type in the 'size' subhash.
335 die "in " . __PACKAGE__ . " constructor '$_;' is a mandatory field"
341 title => $opt{title} || '',
343 foreach my $txt ( @{ $opt{txt} } ) {
345 if ( $str =~ /^[""]/ ) {
347 } elsif ($str =~ / - /x ) { # A range: Replace this element on the
348 # list with its expansion
349 my ($lower, $upper) = $str =~ / 0x (.+?) \s* - \s* 0x (.+) /x;
350 die "Format must be like '0xDEAD - 0xBEAF'; instead was '$str'" if ! defined $lower || ! defined $upper;
351 foreach my $cp (hex $lower .. hex $upper) {
352 push @{$opt{txt}}, sprintf "0x%X", $cp;
355 } elsif ($str =~ s/ ^ N (?= 0x ) //x ) {
356 # Otherwise undocumented, a leading N means is already in the
357 # native character set; don't convert.
359 } elsif ( $str =~ /^0x/ ) {
362 } elsif ( $str =~ / \s* \\p \{ ( .*? ) \} /x) {
364 use Unicode::UCD qw(prop_invlist);
366 my @invlist = prop_invlist($property, '_perl_core_internal_ok');
369 # An empty return could mean an unknown property, or merely
370 # that it is empty. Call in scalar context to differentiate
371 my $count = prop_invlist($property, '_perl_core_internal_ok');
372 die "$property not found" unless defined $count;
375 # Replace this element on the list with the property's expansion
376 for (my $i = 0; $i < @invlist; $i += 2) {
377 foreach my $cp ($invlist[$i] .. $invlist[$i+1] - 1) {
379 # prop_invlist() returns native values; add leading 'N'
381 push @{$opt{txt}}, sprintf "N0x%X", $cp;
385 } elsif ($str =~ / ^ do \s+ ( .* ) /x) {
386 die "do '$1' failed: $!$@" if ! do $1 or $@;
388 } elsif ($str =~ / ^ & \s* ( .* ) /x) { # user-furnished sub() call
389 my @results = eval "$1";
390 die "eval '$1' failed: $@" if $@;
391 push @{$opt{txt}}, @results;
394 die "Unparsable line: $txt\n";
396 my ( $cp, $cp_high, $low, $latin1, $utf8 )= __uni_latin1( $opt{charset}, $str );
397 my $UTF8= $low || $utf8;
398 my $LATIN1= $low || $latin1;
399 my $high = (scalar grep { $_ < 256 } @$cp) ? 0 : $utf8;
400 #die Dumper($txt,$cp,$low,$latin1,$utf8)
401 # if $txt=~/NEL/ or $utf8 and @$utf8>3;
403 @{ $self->{strs}{$str} }{qw( str txt low utf8 latin1 high cp cp_high UTF8 LATIN1 )}=
404 ( $str, $txt, $low, $utf8, $latin1, $high, $cp, $cp_high, $UTF8, $LATIN1 );
405 my $rec= $self->{strs}{$str};
406 foreach my $key ( qw(low utf8 latin1 high cp cp_high UTF8 LATIN1) ) {
407 $self->{size}{$key}{ 0 + @{ $self->{strs}{$str}{$key} } }++
408 if $self->{strs}{$str}{$key};
410 $self->{has_multi} ||= @$cp > 1;
411 $self->{has_ascii} ||= $latin1 && @$latin1;
412 $self->{has_low} ||= $low && @$low;
413 $self->{has_high} ||= !$low && !$latin1;
415 $self->{val_fmt}= $hex_fmt;
416 $self->{count}= 0 + keys %{ $self->{strs} };
420 # my $trie = make_trie($type,$maxlen);
422 # using the data stored in the object build a trie of a specific type,
423 # and with specific maximum depth. The trie is made up the elements of
424 # the given types array for each string in the object (assuming it is
427 # returns the trie, or undef if there was no relevant data in the object.
431 my ( $self, $type, $maxlen )= @_;
433 my $strs= $self->{strs};
435 foreach my $rec ( values %$strs ) {
436 die "panic: unknown type '$type'"
437 if !exists $rec->{$type};
438 my $dat= $rec->{$type};
440 next if $maxlen && @$dat > $maxlen;
442 foreach my $elem ( @$dat ) {
443 $node->{$elem} ||= {};
444 $node= $node->{$elem};
446 $node->{''}= $rec->{str};
448 return 0 + keys( %trie ) ? \%trie : undef;
454 # This returns a list of the positions of the bits in the input word that
460 push @positions, $position if $word & 1;
467 # my $optree= _optree()
469 # recursively convert a trie to an optree where every node represents
475 my ( $self, $trie, $test_type, $ret_type, $else, $depth )= @_;
476 return unless defined $trie;
477 if ( $self->{has_multi} and $ret_type =~ /cp|both/ ) {
478 die "Can't do 'cp' optree from multi-codepoint strings";
481 $else= 0 unless defined $else;
482 $depth= 0 unless defined $depth;
484 # if we have an empty string as a key it means we are in an
485 # accepting state and unless we can match further on should
486 # return the value of the '' key.
487 if (exists $trie->{''} ) {
488 # we can now update the "else" value, anything failing to match
489 # after this point should return the value from this.
490 if ( $ret_type eq 'cp' ) {
491 $else= $self->{strs}{ $trie->{''} }{cp}[0];
492 $else= sprintf "$self->{val_fmt}", $else if $else > 9;
493 } elsif ( $ret_type eq 'len' ) {
495 } elsif ( $ret_type eq 'both') {
496 $else= $self->{strs}{ $trie->{''} }{cp}[0];
497 $else= sprintf "$self->{val_fmt}", $else if $else > 9;
498 $else= "len=$depth, $else";
501 # extract the meaningful keys from the trie, filter out '' as
502 # it means we are an accepting state (end of sequence).
503 my @conds= sort { $a <=> $b } grep { length $_ } keys %$trie;
505 # if we haven't any keys there is no further we can match and we
506 # can return the "else" value.
507 return $else if !@conds;
509 my $test = $test_type =~ /^cp/ ? "cp" : "((U8*)s)[$depth]";
511 # First we loop over the possible keys/conditions and find out what they
512 # look like; we group conditions with the same optree together.
515 local $Data::Dumper::Sortkeys=1;
516 foreach my $cond ( @conds ) {
518 # get the optree for this child/condition
519 my $res= $self->_optree( $trie->{$cond}, $test_type, $ret_type, $else, $depth + 1 );
520 # convert it to a string with Dumper
521 my $res_code= Dumper( $res );
523 push @{$dmp_res{$res_code}{vals}}, $cond;
524 if (!$dmp_res{$res_code}{optree}) {
525 $dmp_res{$res_code}{optree}= $res;
526 push @res_order, $res_code;
530 # now that we have deduped the optrees we construct a new optree containing the merged
534 foreach my $res_code_idx (0 .. $#res_order) {
535 my $res_code= $res_order[$res_code_idx];
536 $node->{vals}= $dmp_res{$res_code}{vals};
537 $node->{test}= $test;
538 $node->{yes}= $dmp_res{$res_code}{optree};
539 $node->{depth}= $depth;
540 if ($res_code_idx < $#res_order) {
541 $node= $node->{no}= {};
551 # my $optree= optree(%opts);
553 # Convert a trie to an optree, wrapper for _optree
558 my $trie= $self->make_trie( $opt{type}, $opt{max_depth} );
559 $opt{ret_type} ||= 'len';
560 my $test_type= $opt{type} =~ /^cp/ ? 'cp' : 'depth';
561 return $self->_optree( $trie, $test_type, $opt{ret_type}, $opt{else}, 0 );
564 # my $optree= generic_optree(%opts);
566 # build a "generic" optree out of the three 'low', 'latin1', 'utf8'
567 # sets of strings, including a branch for handling the string type check.
574 $opt{ret_type} ||= 'len';
575 my $test_type= 'depth';
576 my $else= $opt{else} || 0;
578 my $latin1= $self->make_trie( 'latin1', $opt{max_depth} );
579 my $utf8= $self->make_trie( 'utf8', $opt{max_depth} );
581 $_= $self->_optree( $_, $test_type, $opt{ret_type}, $else, 0 )
585 $else= __cond_join( "( is_utf8 )", $utf8, $latin1 || $else );
586 } elsif ( $latin1 ) {
587 $else= __cond_join( "!( is_utf8 )", $latin1, $else );
589 if ($opt{type} eq 'generic') {
590 my $low= $self->make_trie( 'low', $opt{max_depth} );
592 $else= $self->_optree( $low, $test_type, $opt{ret_type}, $else, 0 );
601 # create a string length guarded optree.
607 my $type= $opt{type};
609 die "Can't do a length_optree on type 'cp', makes no sense."
612 my $else= ( $opt{else} ||= 0 );
614 my $method = $type =~ /generic/ ? 'generic_optree' : 'optree';
615 if ($method eq 'optree' && scalar keys %{$self->{size}{$type}} == 1) {
617 # Here is non-generic output (meaning that we are only generating one
618 # type), and all things that match have the same number ('size') of
619 # bytes. The length guard is simply that we have that number of
621 my @size = keys %{$self->{size}{$type}};
622 my $cond= "((e) - (s)) >= $size[0]";
623 my $optree = $self->$method(%opt);
624 $else= __cond_join( $cond, $optree, $else );
626 elsif ($self->{has_multi}) {
629 # Here, there can be a match of a multiple character string. We use
630 # the traditional method which is to have a branch for each possible
631 # size (longest first) and test for the legal values for that size.
633 %{ $self->{size}{low} || {} },
634 %{ $self->{size}{latin1} || {} },
635 %{ $self->{size}{utf8} || {} }
637 if ($method eq 'generic_optree') {
638 @size= sort { $a <=> $b } keys %sizes;
640 @size= sort { $a <=> $b } keys %{ $self->{size}{$type} };
642 for my $size ( @size ) {
643 my $optree= $self->$method( %opt, type => $type, max_depth => $size );
644 my $cond= "((e)-(s) > " . ( $size - 1 ).")";
645 $else= __cond_join( $cond, $optree, $else );
651 # Here, has more than one possible size, and only matches a single
652 # character. For non-utf8, the needed length is 1; for utf8, it is
653 # found by array lookup 'UTF8SKIP'.
655 # If want just the code points above 255, set up to look for those;
656 # otherwise assume will be looking for all non-UTF-8-invariant code
658 my $trie_type = ($type eq 'high') ? 'high' : 'utf8';
660 # If we do want more than the 0-255 range, find those, and if they
662 if ($opt{type} !~ /latin1/i && ($utf8 = $self->make_trie($trie_type, 0))) {
664 # ... get them into an optree, and set them up as the 'else' clause
665 $utf8 = $self->_optree( $utf8, 'depth', $opt{ret_type}, 0, 0 );
668 # UTF8_IS_START(*s) && ((e) - (s)) >= UTF8SKIP(s))";
669 # to avoid doing the UTF8SKIP and subsequent branches for invariants
670 # that don't match. But the current macros that get generated
671 # have only a few things that can match past this, so I (khw)
672 # don't think it is worth it. (Even better would be to use
673 # calculate_mask(keys %$utf8) instead of UTF8_IS_START, and use it
674 # if it saves a bunch. We assume that input text likely to be
676 my $cond = "LIKELY(((e) - (s)) >= UTF8SKIP(s))";
677 $else = __cond_join($cond, $utf8, $else);
679 # For 'generic', we also will want the latin1 UTF-8 variants for
680 # the case where the input isn't UTF-8.
682 if ($method eq 'generic_optree') {
683 $latin1 = $self->make_trie( 'latin1', 1);
684 $latin1= $self->_optree( $latin1, 'depth', $opt{ret_type}, 0, 0 );
687 # If we want the UTF-8 invariants, get those.
689 if ($opt{type} !~ /non_low|high/
690 && ($low= $self->make_trie( 'low', 1)))
692 $low= $self->_optree( $low, 'depth', $opt{ret_type}, 0, 0 );
694 # Expand out the UTF-8 invariants as a string so that we
695 # can use them as the conditional
696 $low = $self->_cond_as_str( $low, 0, \%opt);
698 # If there are Latin1 variants, add a test for them.
700 $else = __cond_join("(! is_utf8 )", $latin1, $else);
702 elsif ($method eq 'generic_optree') {
704 # Otherwise for 'generic' only we know that what
705 # follows must be valid for just UTF-8 strings,
706 $else->{test} = "( is_utf8 && $else->{test} )";
709 # If the invariants match, we are done; otherwise we have
710 # to go to the 'else' clause.
711 $else = __cond_join($low, 1, $else);
713 elsif ($latin1) { # Here, didn't want or didn't have invariants,
714 # but we do have latin variants
715 $else = __cond_join("(! is_utf8)", $latin1, $else);
718 # We need at least one byte available to start off the tests
719 $else = __cond_join("LIKELY((e) > (s))", $else, 0);
721 else { # Here, we don't want or there aren't any variants. A single
722 # byte available is enough.
723 my $cond= "((e) > (s))";
724 my $optree = $self->$method(%opt);
725 $else= __cond_join( $cond, $optree, $else );
732 sub calculate_mask(@) {
733 # Look at the input list of byte values. This routine returns an array of
734 # mask/base pairs to generate that list.
737 my $list_count = @list;
739 # Consider a set of byte values, A, B, C .... If we want to determine if
740 # <c> is one of them, we can write c==A || c==B || c==C .... If the
741 # values are consecutive, we can shorten that to A<=c && c<=Z, which uses
742 # far fewer branches. If only some of them are consecutive we can still
743 # save some branches by creating range tests for just those that are
744 # consecutive. _cond_as_str() does this work for looking for ranges.
746 # Another approach is to look at the bit patterns for A, B, C .... and see
747 # if they have some commonalities. That's what this function does. For
748 # example, consider a set consisting of the bytes
749 # 0xF0, 0xF1, 0xF2, and 0xF3. We could write:
750 # 0xF0 <= c && c <= 0xF4
751 # But the following mask/compare also works, and has just one test:
753 # The reason it works is that the set consists of exactly those bytes
754 # whose first 4 bits are 1, and the next two are 0. (The value of the
755 # other 2 bits is immaterial in determining if a byte is in the set or
756 # not.) The mask masks out those 2 irrelevant bits, and the comparison
757 # makes sure that the result matches all bytes which match those 6
758 # material bits exactly. In other words, the set of bytes contains
759 # exactly those whose bottom two bit positions are either 0 or 1. The
760 # same principle applies to bit positions that are not necessarily
761 # adjacent. And it can be applied to bytes that differ in 1 through all 8
762 # bit positions. In order to be a candidate for this optimization, the
763 # number of bytes in the set must be a power of 2.
765 # Consider a different example, the set 0x53, 0x54, 0x73, and 0x74. That
766 # requires 4 tests using either ranges or individual values, and even
767 # though the number in the set is a power of 2, it doesn't qualify for the
768 # mask optimization described above because the number of bits that are
769 # different is too large for that. However, the set can be expressed as
770 # two branches with masks thusly:
771 # (c & 0xDF) == 0x53 || (c & 0xDF) == 0x54
772 # a branch savings of 50%. This is done by splitting the set into two
773 # subsets each of which has 2 elements, and within each set the values
776 # This function attempts to find some way to save some branches using the
777 # mask technique. If not, it returns an empty list; if so, it
778 # returns a list consisting of
779 # [ [compare1, mask1], [compare2, mask2], ...
780 # [compare_n, undef], [compare_m, undef], ...
782 # The <mask> is undef in the above for those bytes that must be tested
785 # This function does not attempt to find the optimal set. To do so would
786 # probably require testing all possible combinations, and keeping track of
787 # the current best one.
789 # There are probably much better algorithms, but this is the one I (khw)
790 # came up with. We start with doing a bit-wise compare of every byte in
791 # the set with every other byte. The results are sorted into arrays of
792 # all those that differ by the same bit positions. These are stored in a
793 # hash with the each key being the bits they differ in. Here is the hash
794 # for the 0x53, 0x54, 0x73, 0x74 set:
822 # The set consisting of values which differ in the 4 bit positions 0, 1,
823 # 2, and 5 from some other value in the set consists of all 4 values.
824 # Likewise all 4 values differ from some other value in the 3 bit
825 # positions 0, 1, and 2; and all 4 values differ from some other value in
826 # the single bit position 5. The keys at the uppermost level in the above
827 # hash, 1, 3, and 4, give the number of bit positions that each sub-key
828 # below it has. For example, the 4 key could have as its value an array
829 # consisting of "0,1,2,5", "0,1,2,6", and "3,4,6,7", if the inputs were
830 # such. The best optimization will group the most values into a single
831 # mask. The most values will be the ones that differ in the most
832 # positions, the ones with the largest value for the topmost key. These
833 # keys, are thus just for convenience of sorting by that number, and do
834 # not have any bearing on the core of the algorithm.
836 # We start with an element from largest number of differing bits. The
837 # largest in this case is 4 bits, and there is only one situation in this
838 # set which has 4 differing bits, "0,1,2,5". We look for any subset of
839 # this set which has 16 values that differ in these 4 bits. There aren't
840 # any, because there are only 4 values in the entire set. We then look at
841 # the next possible thing, which is 3 bits differing in positions "0,1,2".
842 # We look for a subset that has 8 values that differ in these 3 bits.
843 # Again there are none. So we go to look for the next possible thing,
844 # which is a subset of 2**1 values that differ only in bit position 5. 83
845 # and 115 do, so we calculate a mask and base for those and remove them
846 # from every set. Since there is only the one set remaining, we remove
847 # them from just this one. We then look to see if there is another set of
848 # 2 values that differ in bit position 5. 84 and 116 do, so we calculate
849 # a mask and base for those and remove them from every set (again only
850 # this set remains in this example). The set is now empty, and there are
851 # no more sets to look at, so we are done.
853 if ($list_count == 256) { # All 256 is trivially masked
859 # Generate bits-differing lists for each element compared against each
861 for my $i (0 .. $list_count - 2) {
862 for my $j ($i + 1 .. $list_count - 1) {
863 my @bits_that_differ = pop_count($list[$i] ^ $list[$j]);
864 my $differ_count = @bits_that_differ;
865 my $key = join ",", @bits_that_differ;
866 push @{$hash{$differ_count}{$key}}, $list[$i] unless grep { $_ == $list[$i] } @{$hash{$differ_count}{$key}};
867 push @{$hash{$differ_count}{$key}}, $list[$j];
871 print STDERR __LINE__, ": calculate_mask() called: List of values grouped by differing bits: ", Dumper \%hash if DEBUG;
874 foreach my $count (reverse sort { $a <=> $b } keys %hash) {
875 my $need = 2 ** $count; # Need 8 values for 3 differing bits, etc
876 foreach my $bits (sort keys $hash{$count}) {
878 print STDERR __LINE__, ": For $count bit(s) difference ($bits), need $need; have ", scalar @{$hash{$count}{$bits}}, "\n" if DEBUG;
880 # Look only as long as there are at least as many elements in the
881 # subset as are needed
882 while ((my $cur_count = @{$hash{$count}{$bits}}) >= $need) {
884 print STDERR __LINE__, ": Looking at bit positions ($bits): ", Dumper $hash{$count}{$bits} if DEBUG;
886 # Start with the first element in it
887 my $try_base = $hash{$count}{$bits}[0];
888 my @subset = $try_base;
890 # If it succeeds, we return a mask and a base to compare
891 # against the masked value. That base will be the AND of
892 # every element in the subset. Initialize to the one element
894 my $compare = $try_base;
896 # We are trying to find a subset of this that has <need>
897 # elements that differ in the bit positions given by the
898 # string $bits, which is comma separated.
899 my @bits = split ",", $bits;
901 TRY: # Look through the remainder of the list for other
902 # elements that differ only by these bit positions.
904 for (my $i = 1; $i < $cur_count; $i++) {
905 my $try_this = $hash{$count}{$bits}[$i];
906 my @positions = pop_count($try_base ^ $try_this);
908 print STDERR __LINE__, ": $try_base vs $try_this: is (", join(',', @positions), ") a subset of ($bits)?" if DEBUG;;
910 foreach my $pos (@positions) {
911 unless (grep { $pos == $_ } @bits) {
912 print STDERR " No\n" if DEBUG;
913 my $remaining = $cur_count - $i - 1;
914 if ($remaining && @subset + $remaining < $need) {
915 print STDERR __LINE__, ": Can stop trying $try_base, because even if all the remaining $remaining values work, they wouldn't add up to the needed $need when combined with the existing ", scalar @subset, " ones\n" if DEBUG;
922 print STDERR " Yes\n" if DEBUG;
923 push @subset, $try_this;
925 # Add this to the mask base, in case it ultimately
927 $compare &= $try_this;
930 print STDERR __LINE__, ": subset (", join(", ", @subset), ") has ", scalar @subset, " elements; needs $need\n" if DEBUG;
932 if (@subset < $need) {
933 shift @{$hash{$count}{$bits}};
934 next; # Try with next value
939 foreach my $position (@bits) {
940 $mask |= 1 << $position;
942 $mask = ~$mask & 0xFF;
943 push @final_results, [$compare, $mask];
945 printf STDERR "%d: Got it: compare=%d=0x%X; mask=%X\n", __LINE__, $compare, $compare, $mask if DEBUG;
947 # These values are now spoken for. Remove them from future
949 foreach my $remove_count (sort keys %hash) {
950 foreach my $bits (sort keys %{$hash{$remove_count}}) {
951 foreach my $to_remove (@subset) {
952 @{$hash{$remove_count}{$bits}} = grep { $_ != $to_remove } @{$hash{$remove_count}{$bits}};
960 # Any values that remain in the list are ones that have to be tested for
963 foreach my $count (reverse sort { $a <=> $b } keys %hash) {
964 foreach my $bits (sort keys $hash{$count}) {
965 foreach my $remaining (@{$hash{$count}{$bits}}) {
967 # If we already know about this value, just ignore it.
968 next if grep { $remaining == $_ } @individuals;
970 # Otherwise it needs to be returned as something to match
972 push @final_results, [$remaining, undef];
973 push @individuals, $remaining;
978 # Sort by increasing numeric value
979 @final_results = sort { $a->[0] <=> $b->[0] } @final_results;
981 print STDERR __LINE__, ": Final return: ", Dumper \@final_results if DEBUG;
983 return @final_results;
987 # turn a list of conditions into a text expression
988 # - merges ranges of conditions, and joins the result with ||
990 my ( $self, $op, $combine, $opts_ref )= @_;
991 my $cond= $op->{vals};
992 my $test= $op->{test};
993 my $is_cp_ret = $opts_ref->{ret_type} eq "cp";
994 return "( $test )" if !defined $cond;
999 # We skip this if there are optimizations that
1000 # we can apply (below) to the individual ranges
1001 if ( ($is_cp_ret || $combine) && @ranges && ref $ranges[-1]) {
1002 if ( $ranges[-1][0] == $ranges[-1][1] ) {
1003 $ranges[-1]= $ranges[-1][0];
1004 } elsif ( $ranges[-1][0] + 1 == $ranges[-1][1] ) {
1005 $ranges[-1]= $ranges[-1][0];
1006 push @ranges, $ranges[-1] + 1;
1010 for my $condition ( @$cond ) {
1011 if ( !@ranges || $condition != $ranges[-1][1] + 1 ) {
1013 push @ranges, [ $condition, $condition ];
1020 return $self->_combine( $test, @ranges )
1027 "( $self->{val_fmt} <= $test && $test <= $self->{val_fmt} )",
1029 : sprintf( "$self->{val_fmt} == $test", $_ );
1032 return "( " . join( " || ", @ranges ) . " )";
1035 # If the input set has certain characteristics, we can optimize tests
1036 # for it. This doesn't apply if returning the code point, as we want
1037 # each element of the set individually. The code above is for this
1040 return 1 if @$cond == 256; # If all bytes match, is trivially true
1045 # See if the entire set shares optimizable characteristics, and if so,
1046 # return the optimization. We delay checking for this on sets with
1047 # just a single range, as there may be better optimizations available
1049 @masks = calculate_mask(@$cond);
1051 # Stringify the output of calculate_mask()
1054 foreach my $mask_ref (@masks) {
1055 if (defined $mask_ref->[1]) {
1056 push @return, sprintf "( ( $test & $self->{val_fmt} ) == $self->{val_fmt} )", $mask_ref->[1], $mask_ref->[0];
1058 else { # An undefined mask means to use the value as-is
1059 push @return, sprintf "$test == $self->{val_fmt}", $mask_ref->[0];
1063 # The best possible case below for specifying this set of values via
1064 # ranges is 1 branch per range. If our mask method yielded better
1065 # results, there is no sense trying something that is bound to be
1067 if (@return < @ranges) {
1068 return "( " . join( " || ", @return ) . " )";
1075 # Here, there was no entire-class optimization that was clearly better
1076 # than doing things by ranges. Look at each range.
1077 my $range_count_extra = 0;
1078 for (my $i = 0; $i < @ranges; $i++) {
1079 if (! ref $ranges[$i]) { # Trivial case: no range
1080 $ranges[$i] = sprintf "$self->{val_fmt} == $test", $ranges[$i];
1082 elsif ($ranges[$i]->[0] == $ranges[$i]->[1]) {
1083 $ranges[$i] = # Trivial case: single element range
1084 sprintf "$self->{val_fmt} == $test", $ranges[$i]->[0];
1086 elsif ($ranges[$i]->[0] == 0) {
1087 # If the range matches all 256 possible bytes, it is trivially
1089 return 1 if $ranges[0]->[1] == 0xFF; # @ranges must be 1 in
1091 $ranges[$i] = sprintf "( $test <= $self->{val_fmt} )",
1094 elsif ($ranges[$i]->[1] == 255) {
1096 # Similarly the max possible is 255, so can omit an upper bound
1097 # test if the calculated max is the max possible one.
1098 $ranges[$i] = sprintf "( $test >= $self->{val_fmt} )",
1104 # Well-formed UTF-8 continuation bytes on ascii platforms must be
1105 # in the range 0x80 .. 0xBF. If we know that the input is
1106 # well-formed (indicated by not trying to be 'safe'), we can omit
1107 # tests that verify that the input is within either of these
1108 # bounds. (No legal UTF-8 character can begin with anything in
1109 # this range, so we don't have to worry about this being a
1110 # continuation byte or not.)
1111 if ($opts_ref->{charset} =~ /ascii/i
1112 && (! $opts_ref->{safe} && ! $opts_ref->{no_length_checks})
1113 && $opts_ref->{type} =~ / ^ (?: utf8 | high ) $ /xi)
1115 my $lower_limit_is_80 = ($ranges[$i]->[0] == 0x80);
1116 my $upper_limit_is_BF = ($ranges[$i]->[1] == 0xBF);
1118 # If the range is the entire legal range, it matches any legal
1119 # byte, so we can omit both tests. (This should happen only
1120 # if the number of ranges is 1.)
1121 if ($lower_limit_is_80 && $upper_limit_is_BF) {
1124 elsif ($lower_limit_is_80) { # Just use the upper limit test
1125 $output = sprintf("( $test <= $self->{val_fmt} )",
1128 elsif ($upper_limit_is_BF) { # Just use the lower limit test
1129 $output = sprintf("( $test >= $self->{val_fmt} )",
1134 # If we didn't change to omit a test above, see if the number of
1135 # elements is a power of 2 (only a single bit in the
1136 # representation of its count will be set) and if so, it may be
1137 # that a mask/compare optimization is possible.
1139 && pop_count($ranges[$i]->[1] - $ranges[$i]->[0] + 1) == 1)
1142 push @list, $_ for ($ranges[$i]->[0] .. $ranges[$i]->[1]);
1143 my @this_masks = calculate_mask(@list);
1145 # Use the mask if there is just one for the whole range.
1146 # Otherwise there is no savings over the two branches that can
1148 if (@this_masks == 1 && defined $this_masks[0][1]) {
1149 $output = sprintf "( $test & $self->{val_fmt} ) == $self->{val_fmt}", $this_masks[0][1], $this_masks[0][0];
1153 if ($output ne "") { # Prefer any optimization
1154 $ranges[$i] = $output;
1157 # No optimization happened. We need a test that the code
1158 # point is within both bounds. But, if the bounds are
1159 # adjacent code points, it is cleaner to say
1160 # 'first == test || second == test'
1162 # 'first <= test && test <= second'
1164 $range_count_extra++; # This range requires 2 branches to
1166 if ($ranges[$i]->[0] + 1 == $ranges[$i]->[1]) {
1168 . join( " || ", ( map
1169 { sprintf "$self->{val_fmt} == $test", $_ }
1173 else { # Full bounds checking
1174 $ranges[$i] = sprintf("( $self->{val_fmt} <= $test && $test <= $self->{val_fmt} )", $ranges[$i]->[0], $ranges[$i]->[1]);
1180 # We have generated the list of bytes in two ways; one trying to use masks
1181 # to cut the number of branches down, and the other to look at individual
1182 # ranges (some of which could be cut down by using a mask for just it).
1183 # We return whichever method uses the fewest branches.
1185 . join( " || ", (@masks && @masks < @ranges + $range_count_extra)
1192 # recursively turn a list of conditions into a fast break-out condition
1193 # used by _cond_as_str() for 'cp' type macros.
1195 my ( $self, $test, @cond )= @_;
1197 my $item= shift @cond;
1199 if ( ref $item ) { # @item should be a 2-element array giving range start
1201 if ($item->[0] == 0) { # UV's are never negative, so skip "0 <= "
1202 # test which could generate a compiler warning
1203 # that test is always true
1204 $cstr= sprintf( "$test <= $self->{val_fmt}", $item->[1] );
1208 sprintf( "( $self->{val_fmt} <= $test && $test <= $self->{val_fmt} )",
1211 $gtv= sprintf "$self->{val_fmt}", $item->[1];
1213 $cstr= sprintf( "$self->{val_fmt} == $test", $item );
1214 $gtv= sprintf "$self->{val_fmt}", $item;
1217 my $combine= $self->_combine( $test, @cond );
1219 return "( $cstr || ( $gtv < $test &&\n"
1220 . $combine . " ) )";
1222 return "( $cstr || $combine )";
1230 # recursively convert an optree to text with reasonably neat formatting
1232 my ( $self, $op, $combine, $brace, $opts_ref, $def, $submacros )= @_;
1233 return 0 if ! defined $op; # The set is empty
1237 my $cond= $self->_cond_as_str( $op, $combine, $opts_ref );
1238 #no warnings 'recursion'; # This would allow really really inefficient
1239 # code to be generated. See pod
1240 my $yes= $self->_render( $op->{yes}, $combine, 1, $opts_ref, $def, $submacros );
1241 return $yes if $cond eq '1';
1243 my $no= $self->_render( $op->{no}, $combine, 0, $opts_ref, $def, $submacros );
1244 return "( $cond )" if $yes eq '1' and $no eq '0';
1245 my ( $lb, $rb )= $brace ? ( "( ", " )" ) : ( "", "" );
1246 return "$lb$cond ? $yes : $no$rb"
1247 if !ref( $op->{yes} ) && !ref( $op->{no} );
1249 my $ind= "\n" . ( $ind1 x $op->{depth} );
1251 if ( ref $op->{yes} ) {
1252 $yes= $ind . $ind1 . $yes;
1257 my $str= "$lb$cond ?$yes$ind: $no$rb";
1258 if (length $str > 6000) {
1259 push @$submacros, sprintf "#define $def\n( %s )", "_part" . (my $yes_idx= 0+@$submacros), $yes;
1260 push @$submacros, sprintf "#define $def\n( %s )", "_part" . (my $no_idx= 0+@$submacros), $no;
1261 return sprintf "%s%s ? $def : $def%s", $lb, $cond, "_part$yes_idx", "_part$no_idx", $rb;
1266 # $expr=render($op,$combine)
1268 # convert an optree to text with reasonably neat formatting. If $combine
1269 # is true then the condition is created using "fast breakouts" which
1270 # produce uglier expressions that are more efficient for common case,
1271 # longer lists such as that resulting from type 'cp' output.
1272 # Currently only used for type 'cp' macros.
1274 my ( $self, $op, $combine, $opts_ref, $def_fmt )= @_;
1277 my $macro= sprintf "#define $def_fmt\n( %s )", "", $self->_render( $op, $combine, 0, $opts_ref, $def_fmt, \@submacros );
1279 return join "\n\n", map { "/*** GENERATED CODE ***/\n" . __macro( __clean( $_ ) ) } @submacros, $macro;
1283 # make a macro of a given type.
1284 # calls into make_trie and (generic_|length_)optree as needed
1286 # type : 'cp','cp_high', 'generic','high','low','latin1','utf8','LATIN1','UTF8'
1287 # ret_type : 'cp' or 'len'
1288 # safe : don't assume is well-formed UTF-8, so don't skip any range
1289 # checks, and add length guards to macro
1290 # no_length_checks : like safe, but don't add length guards.
1292 # type defaults to 'generic', and ret_type to 'len' unless type is 'cp'
1293 # in which case it defaults to 'cp' as well.
1295 # It is illegal to do a type 'cp' macro on a pattern with multi-codepoint
1296 # sequences in it, as the generated macro will accept only a single codepoint
1299 # It is also illegal to do a non-safe macro on a pattern with multi-codepoint
1300 # sequences in it, as even if it is known to be well-formed, we need to not
1301 # run off the end of the buffer when, say, the buffer ends with the first two
1302 # characters, but three are looked at by the macro.
1304 # returns the macro.
1310 my $type= $opts{type} || 'generic';
1311 if ($self->{has_multi}) {
1312 if ($type =~ /^cp/) {
1313 die "Can't do a 'cp' on multi-codepoint character class '$self->{op}'"
1315 elsif (! $opts{safe}) {
1316 die "'safe' is required on multi-codepoint character class '$self->{op}'"
1319 my $ret_type= $opts{ret_type} || ( $opts{type} =~ /^cp/ ? 'cp' : 'len' );
1321 if ( $opts{safe} ) {
1322 $method= 'length_optree';
1323 } elsif ( $type =~ /generic/ ) {
1324 $method= 'generic_optree';
1328 my @args= $type =~ /^cp/ ? 'cp' : 's';
1329 push @args, "e" if $opts{safe};
1330 push @args, "is_utf8" if $type =~ /generic/;
1331 push @args, "len" if $ret_type eq 'both';
1332 my $pfx= $ret_type eq 'both' ? 'what_len_' :
1333 $ret_type eq 'cp' ? 'what_' : 'is_';
1334 my $ext= $type =~ /generic/ ? '' : '_' . lc( $type );
1335 $ext .= '_non_low' if $type eq 'generic_non_low';
1336 $ext .= "_safe" if $opts{safe};
1337 $ext .= "_no_length_checks" if $opts{no_length_checks};
1338 my $argstr= join ",", @args;
1339 my $def_fmt="$pfx$self->{op}$ext%s($argstr)";
1340 my $optree= $self->$method( %opts, type => $type, ret_type => $ret_type );
1341 return $self->render( $optree, ($type =~ /^cp/) ? 1 : 0, \%opts, $def_fmt );
1344 # if we aren't being used as a module (highly likely) then process
1345 # the __DATA__ below and produce macros in regcharclass.h
1346 # if an argument is provided to the script then it is assumed to
1347 # be the path of the file to output to, if the arg is '-' outputs
1351 my $path= shift @ARGV || "regcharclass.h";
1353 if ( $path eq '-' ) {
1356 $out_fh = open_new( $path );
1358 print $out_fh read_only_top( lang => 'C', by => $0,
1359 file => 'regcharclass.h', style => '*',
1360 copyright => [2007, 2011],
1362 WARNING: These macros are for internal Perl core use only, and may be
1363 changed or removed without notice.
1366 print $out_fh "\n#ifndef H_REGCHARCLASS /* Guard against nested #includes */\n#define H_REGCHARCLASS 1\n";
1368 my ( $op, $title, @txt, @types, %mods );
1372 my $charset = shift;
1374 # Skip if to compile on a different platform.
1375 return if delete $mods{only_ascii_platform} && $charset !~ /ascii/i;
1376 return if delete $mods{only_ebcdic_platform} && $charset !~ /ebcdic/i;
1378 print $out_fh "/*\n\t$op: $title\n\n";
1379 print $out_fh join "\n", ( map { "\t$_" } @txt ), "*/", "";
1380 my $obj= __PACKAGE__->new( op => $op, title => $title, txt => \@txt, charset => $charset);
1382 #die Dumper(\@types,\%mods);
1385 push @mods, 'safe' if delete $mods{safe};
1386 push @mods, 'no_length_checks' if delete $mods{no_length_checks};
1387 unshift @mods, 'fast' if delete $mods{fast} || ! @mods; # Default to 'fast'
1392 die "Unknown modifiers: ", join ", ", map { "'$_'" } sort keys %mods;
1395 foreach my $type_spec ( @types ) {
1396 my ( $type, $ret )= split /-/, $type_spec;
1398 foreach my $mod ( @mods ) {
1400 # 'safe' is irrelevant with code point macros, so skip if
1401 # there is also a 'fast', but don't skip if this is the only
1402 # way a cp macro will get generated. Below we convert 'safe'
1403 # to 'fast' in this instance
1404 next if $type =~ /^cp/
1405 && ($mod eq 'safe' || $mod eq 'no_length_checks')
1406 && grep { 'fast' =~ $_ } @mods;
1408 my $macro= $obj->make_macro(
1411 safe => $mod eq 'safe' && $type !~ /^cp/,
1412 charset => $charset,
1413 no_length_checks => $mod eq 'no_length_checks' && $type !~ /^cp/,
1415 print $out_fh $macro, "\n";
1421 foreach my $charset (get_supported_code_pages()) {
1428 print $out_fh "\n", get_conditional_compile_line_start($charset);
1429 my @data_copy = @data;
1431 s/^ \s* (?: \# .* ) ? $ //x; # squeeze out comment and blanks
1435 $doit->($charset) unless $first_time; # This starts a new
1436 # definition; do the
1439 ( $op, $title )= split /\s*:\s*/, $_, 2;
1441 } elsif ( s/^=>// ) {
1442 my ( $type, $modifier )= split /:/, $_;
1443 @types= split ' ', $type;
1445 map { $mods{$_} = 1 } split ' ', $modifier;
1451 print $out_fh get_conditional_compile_line_end();
1454 print $out_fh "\n#endif /* H_REGCHARCLASS */\n";
1457 print $out_fh "/* ex: set ro: */\n";
1459 # Some of the sources for these macros come from Unicode tables
1460 my $sources_list = "lib/unicore/mktables.lst";
1461 my @sources = ($0, qw(lib/unicore/mktables
1463 regen/regcharclass_multi_char_folds.pl
1464 regen/charset_translations.pl
1467 # Depend on mktables’ own sources. It’s a shorter list of files than
1468 # those that Unicode::UCD uses.
1469 if (! open my $mktables_list, $sources_list) {
1471 # This should force a rebuild once $sources_list exists
1472 push @sources, $sources_list;
1475 while(<$mktables_list>) {
1478 push @sources, "lib/unicore/$_" if /^[^#]/;
1482 read_only_bottom_close_and_rename($out_fh, \@sources)
1486 # The form of the input is a series of definitions to make macros for.
1487 # The first line gives the base name of the macro, followed by a colon, and
1488 # then text to be used in comments associated with the macro that are its
1489 # title or description. In all cases the first (perhaps only) parameter to
1490 # the macro is a pointer to the first byte of the code point it is to test to
1491 # see if it is in the class determined by the macro. In the case of non-UTF8,
1492 # the code point consists only of a single byte.
1494 # The second line must begin with a '=>' and be followed by the types of
1495 # macro(s) to be generated; these are specified below. A colon follows the
1496 # types, followed by the modifiers, also specified below. At least one
1497 # modifier is required.
1499 # The subsequent lines give what code points go into the class defined by the
1500 # macro. Multiple characters may be specified via a string like "\x0D\x0A",
1501 # enclosed in quotes. Otherwise the lines consist of one of:
1502 # 1) a single Unicode code point, prefaced by 0x
1503 # 2) a single range of Unicode code points separated by a minus (and
1505 # 3) a single Unicode property specified in the standard Perl form
1507 # 4) a line like 'do path'. This will do a 'do' on the file given by
1508 # 'path'. It is assumed that this does nothing but load subroutines
1509 # (See item 5 below). The reason 'require path' is not used instead is
1510 # because 'do' doesn't assume that path is in @INC.
1511 # 5) a subroutine call
1512 # &pkg::foo(arg1, ...)
1513 # where pkg::foo was loaded by a 'do' line (item 4). The subroutine
1514 # returns an array of entries of forms like items 1-3 above. This
1515 # allows more complex inputs than achievable from the other input types.
1517 # A blank line or one whose first non-blank character is '#' is a comment.
1518 # The definition of the macro is terminated by a line unlike those described.
1521 # low generate a macro whose name is 'is_BASE_low' and defines a
1522 # class that includes only ASCII-range chars. (BASE is the
1523 # input macro base name.)
1524 # latin1 generate a macro whose name is 'is_BASE_latin1' and defines a
1525 # class that includes only upper-Latin1-range chars. It is not
1526 # designed to take a UTF-8 input parameter.
1527 # high generate a macro whose name is 'is_BASE_high' and defines a
1528 # class that includes all relevant code points that are above
1529 # the Latin1 range. This is for very specialized uses only.
1530 # It is designed to take only an input UTF-8 parameter.
1531 # utf8 generate a macro whose name is 'is_BASE_utf8' and defines a
1532 # class that includes all relevant characters that aren't ASCII.
1533 # It is designed to take only an input UTF-8 parameter.
1534 # LATIN1 generate a macro whose name is 'is_BASE_latin1' and defines a
1535 # class that includes both ASCII and upper-Latin1-range chars.
1536 # It is not designed to take a UTF-8 input parameter.
1537 # UTF8 generate a macro whose name is 'is_BASE_utf8' and defines a
1538 # class that can include any code point, adding the 'low' ones
1539 # to what 'utf8' works on. It is designed to take only an input
1541 # generic generate a macro whose name is 'is_BASE". It has a 2nd,
1542 # boolean, parameter which indicates if the first one points to
1543 # a UTF-8 string or not. Thus it works in all circumstances.
1544 # generic_non_low generate a macro whose name is 'is_BASE_non_low". It has
1545 # a 2nd, boolean, parameter which indicates if the first one
1546 # points to a UTF-8 string or not. It excludes any ASCII-range
1547 # matches, but otherwise it works in all circumstances.
1548 # cp generate a macro whose name is 'is_BASE_cp' and defines a
1549 # class that returns true if the UV parameter is a member of the
1550 # class; false if not.
1551 # cp_high like cp, but it is assumed that it is known that the UV
1552 # parameter is above Latin1. The name of the generated macro is
1553 # 'is_BASE_cp_high'. This is different from high-cp, derived
1555 # A macro of the given type is generated for each type listed in the input.
1556 # The default return value is the number of octets read to generate the match.
1557 # Append "-cp" to the type to have it instead return the matched codepoint.
1558 # The macro name is changed to 'what_BASE...'. See pod for
1560 # Appending '-both" instead adds an extra parameter to the end of the argument
1561 # list, which is a pointer as to where to store the number of
1562 # bytes matched, while also returning the code point. The macro
1563 # name is changed to 'what_len_BASE...'. See pod for caveats
1566 # safe The input string is not necessarily valid UTF-8. In
1567 # particular an extra parameter (always the 2nd) to the macro is
1568 # required, which points to one beyond the end of the string.
1569 # The macro will make sure not to read off the end of the
1570 # string. In the case of non-UTF8, it makes sure that the
1571 # string has at least one byte in it. The macro name has
1572 # '_safe' appended to it.
1573 # no_length_checks The input string is not necessarily valid UTF-8, but it
1574 # is to be assumed that the length has already been checked and
1576 # fast The input string is valid UTF-8. No bounds checking is done,
1577 # and the macro can make assumptions that lead to faster
1579 # only_ascii_platform Skip this definition if the character set is for
1580 # a non-ASCII platform.
1581 # only_ebcdic_platform Skip this definition if the character set is for
1582 # a non-EBCDIC platform.
1583 # No modifier need be specified; fast is assumed for this case. If both
1584 # 'fast', and 'safe' are specified, two macros will be created for each
1587 # If run on a non-ASCII platform will automatically convert the Unicode input
1588 # to native. The documentation above is slightly wrong in this case. 'low'
1589 # actually refers to code points whose UTF-8 representation is the same as the
1590 # non-UTF-8 version (invariants); and 'latin1' refers to all the rest of the
1591 # code points less than 256.
1593 1; # in the unlikely case we are being used as a module
1596 # This is no longer used, but retained in case it is needed some day.
1597 # TRICKYFOLD: Problematic fold case letters. When adding to this list, also should add them to regcomp.c and fold_grind.t
1598 # => generic cp generic-cp generic-both :fast safe
1599 # 0x00DF # LATIN SMALL LETTER SHARP S
1600 # 0x0390 # GREEK SMALL LETTER IOTA WITH DIALYTIKA AND TONOS
1601 # 0x03B0 # GREEK SMALL LETTER UPSILON WITH DIALYTIKA AND TONOS
1602 # 0x1E9E # LATIN CAPITAL LETTER SHARP S, because maps to same as 00DF
1603 # 0x1FD3 # GREEK SMALL LETTER IOTA WITH DIALYTIKA AND OXIA; maps same as 0390
1604 # 0x1FE3 # GREEK SMALL LETTER UPSILON WITH DIALYTIKA AND OXIA; maps same as 03B0
1606 LNBREAK: Line Break: \R
1607 => generic UTF8 LATIN1 : safe
1608 "\x0D\x0A" # CRLF - Network (Windows) line ending
1611 HORIZWS: Horizontal Whitespace: \h \H
1612 => high cp_high : fast
1615 VERTWS: Vertical Whitespace: \v \V
1616 => high cp_high : fast
1619 XDIGIT: Hexadecimal digits
1620 => high cp_high : fast
1623 XPERLSPACE: \p{XPerlSpace}
1624 => high cp_high : fast
1627 REPLACEMENT: Unicode REPLACEMENT CHARACTER
1631 NONCHAR: Non character code points
1635 SURROGATE: Surrogate characters
1639 # This program was run with this enabled, and the results copied to utf8.h;
1640 # then this was commented out because it takes so long to figure out these 2
1641 # million code points. The results would not change unless utf8.h decides it
1642 # wants a maximum other than 4 bytes, or this program creates better
1643 # optimizations. Trying with 5 bytes used too much memory to calculate.
1645 # We don't generate code for invariants here because the EBCDIC form is too
1646 # complicated and would slow things down; instead the user should test for
1649 # NOTE: The number of bytes generated here must match the value in
1650 # IS_UTF8_CHAR_FAST in utf8.h
1652 #UTF8_CHAR: Matches legal UTF-8 encoded characters from 2 through 4 bytes
1653 #=> UTF8 :no_length_checks only_ascii_platform
1656 # This hasn't been commented out, but the number of bytes it works on has been
1657 # cut down to 3, so it doesn't cover the full legal Unicode range. Making it
1658 # 5 bytes would cover beyond the full range, but takes quite a bit of time and
1659 # memory to calculate. The generated table varies depending on the EBCDIC
1662 # NOTE: The number of bytes generated here must match the value in
1663 # IS_UTF8_CHAR_FAST in utf8.h
1665 UTF8_CHAR: Matches legal UTF-EBCDIC encoded characters from 2 through 3 bytes
1666 => UTF8 :no_length_checks only_ebcdic_platform
1669 QUOTEMETA: Meta-characters that \Q should quote
1673 MULTI_CHAR_FOLD: multi-char strings that are folded to by a single character
1677 ®charclass_multi_char_folds::multi_char_folds(1)
1679 MULTI_CHAR_FOLD: multi-char strings that are folded to by a single character
1682 ®charclass_multi_char_folds::multi_char_folds(0)
1685 FOLDS_TO_MULTI: characters that fold to multi-char strings
1687 \p{_Perl_Folds_To_Multi_Char}
1689 PROBLEMATIC_LOCALE_FOLD : characters whose fold is problematic under locale
1691 \p{_Perl_Problematic_Locale_Folds}
1693 PROBLEMATIC_LOCALE_FOLDEDS_START : The first folded character of folds which are problematic under locale
1695 \p{_Perl_Problematic_Locale_Foldeds_Start}
1697 PATWS: pattern white space
1698 => generic cp : safe