package Unicode::Collate::Locale; use strict; use warnings; use Carp; use base qw(Unicode::Collate); our $VERSION = '1.31'; my $PL_EXT = '.pl'; my %LocaleFile = map { ($_, $_) } qw( af ar as az be bn ca cs cu cy da dsb ee eo es et fa fi fil fo gu ha haw he hi hr hu hy ig is ja kk kl kn ko kok lkt ln lt lv mk ml mr mt nb nn nso om or pa pl ro sa se si sk sl sq sr sv ta te th tn to tr uk ur vi vo wae wo yo zh ); $LocaleFile{'default'} = ''; # aliases $LocaleFile{'bs'} = 'hr'; $LocaleFile{'bs_Cyrl'} = 'sr'; $LocaleFile{'sr_Latn'} = 'hr'; # short file names $LocaleFile{'de__phonebook'} = 'de_phone'; $LocaleFile{'de_AT_phonebook'} = 'de_at_ph'; $LocaleFile{'es__traditional'} = 'es_trad'; $LocaleFile{'fr_CA'} = 'fr_ca'; $LocaleFile{'fi__phonebook'} = 'fi_phone'; $LocaleFile{'si__dictionary'} = 'si_dict'; $LocaleFile{'sv__reformed'} = 'sv_refo'; $LocaleFile{'ug_Cyrl'} = 'ug_cyrl'; $LocaleFile{'zh__big5han'} = 'zh_big5'; $LocaleFile{'zh__gb2312han'} = 'zh_gb'; $LocaleFile{'zh__pinyin'} = 'zh_pin'; $LocaleFile{'zh__stroke'} = 'zh_strk'; $LocaleFile{'zh__zhuyin'} = 'zh_zhu'; my %TypeAlias = qw( phone phonebook phonebk phonebook dict dictionary reform reformed trad traditional big5 big5han gb2312 gb2312han ); sub _locale { my $locale = shift; if ($locale) { $locale = lc $locale; $locale =~ tr/\-\ \./_/; $locale =~ s/_([0-9a-z]+)\z/$TypeAlias{$1} ? "_$TypeAlias{$1}" : "_$1"/e; $LocaleFile{$locale} and return $locale; my @code = split /_/, $locale; my $lan = shift @code; my $scr = @code && length $code[0] == 4 ? ucfirst shift @code : ''; my $reg = @code && length $code[0] < 4 ? uc shift @code : ''; my $var = @code ? shift @code : ''; my @list; push @list, ( "${lan}_${scr}_${reg}_$var", "${lan}_${scr}__$var", # empty $scr should not be ${lan}__$var. "${lan}_${reg}_$var", # empty $reg may be ${lan}__$var. "${lan}__$var", ) if $var ne ''; push @list, ( "${lan}_${scr}_${reg}", "${lan}_${scr}", "${lan}_${reg}", ${lan}, ); for my $loc (@list) { $LocaleFile{$loc} and return $loc; } } return 'default'; } sub getlocale { return shift->{accepted_locale}; } sub locale_version { return shift->{locale_version}; } sub _fetchpl { my $accepted = shift; my $f = $LocaleFile{$accepted}; return if !$f; $f .= $PL_EXT; # allow to search @INC # use File::Spec; # my $path = File::Spec->catfile('Unicode', 'Collate', 'Locale', $f); my $path = "Unicode/Collate/Locale/$f"; my $h = do $path; croak "Unicode/Collate/Locale/$f can't be found" if !$h; return $h; } sub new { my $class = shift; my %hash = @_; $hash{accepted_locale} = _locale($hash{locale}); if (exists $hash{table}) { croak "your table can't be used with Unicode::Collate::Locale"; } my $href = _fetchpl($hash{accepted_locale}); while (my($k,$v) = each %$href) { if (!exists $hash{$k}) { $hash{$k} = $v; } elsif ($k eq 'entry') { $hash{$k} = $v.$hash{$k}; } else { croak "$k is reserved by $hash{locale}, can't be overwritten"; } } return $class->SUPER::new(%hash); } 1; __END__ =head1 NAME Unicode::Collate::Locale - Linguistic tailoring for DUCET via Unicode::Collate =head1 SYNOPSIS use Unicode::Collate::Locale; #construct $Collator = Unicode::Collate::Locale-> new(locale => $locale_name, %tailoring); #sort @sorted = $Collator->sort(@not_sorted); #compare $result = $Collator->cmp($a, $b); # returns 1, 0, or -1. B Strings in C<@not_sorted>, C<$a> and C<$b> are interpreted according to Perl's Unicode support. See L, L, L, L, L. Otherwise you can use C (cf. C) or should decode them before. =head1 DESCRIPTION This module provides linguistic tailoring for it taking advantage of C. =head2 Constructor The C method returns a collator object. A parameter list for the constructor is a hash, which can include a special key C and its value (case-insensitive) standing for a Unicode base language code (two or three-letter). For example, Cnew(locale =E 'ES')> returns a collator tailored for Spanish. C<$locale_name> may be suffixed with a Unicode script code (four-letter), a Unicode region (territory) code, a Unicode language variant code. These codes are case-insensitive, and separated with C<'_'> or C<'-'>. E.g. C for English in USA, C for Azerbaijani in the Cyrillic script, C for Spanish in Spain (Traditional). If C<$locale_name> is not available, fallback is selected in the following order: 1. language with a variant code 2. language with a script code 3. language with a region code 4. language 5. default Tailoring tags provided by C are allowed as long as they are not used for C support. Esp. the C tag is always untailorable, since it is reserved for DUCET. However C is allowed, even if it is used for C support, to add or override mappings. E.g. a collator for Spanish, which ignores diacritics and case difference (i.e. level 1), with reversed case ordering and no normalization. Unicode::Collate::Locale->new( level => 1, locale => 'es', upper_before_lower => 1, normalization => undef ) Overriding a behavior already tailored by C is disallowed if such a tailoring is passed to C. Unicode::Collate::Locale->new( locale => 'da', upper_before_lower => 0, # causes error as reserved by 'da' ) However C inherited from C allows such a tailoring that is reserved by C. Examples: new(locale => 'fr_ca')->change(backwards => undef) new(locale => 'da')->change(upper_before_lower => 0) new(locale => 'ja')->change(overrideCJK => undef) =head2 Methods C is a subclass of C and methods other than C are inherited from C. Here is a list of additional methods: =over 4 =item C<$Collator-Egetlocale> Returns a language code accepted and used actually on collation. If linguistic tailoring is not provided for a language code you passed (intensionally for some languages, or due to the incomplete implementation), this method returns a string C<'default'> meaning no special tailoring. =item C<$Collator-Elocale_version> (Since Unicode::Collate::Locale 0.87) Returns the version number (perhaps C) of the locale, as that of F. B F that a collator uses should be identified by a combination of return values from C and C. =back =head2 A list of tailorable locales locale name description -------------------------------------------------------------- af Afrikaans ar Arabic as Assamese az Azerbaijani (Azeri) be Belarusian bn Bengali bs Bosnian (tailored as Croatian) bs_Cyrl Bosnian in Cyrillic (tailored as Serbian) ca Catalan cs Czech cu Church Slavic cy Welsh da Danish de__phonebook German (umlaut as 'ae', 'oe', 'ue') de_AT_phonebook Austrian German (umlaut primary greater) dsb Lower Sorbian ee Ewe eo Esperanto es Spanish es__traditional Spanish ('ch' and 'll' as a grapheme) et Estonian fa Persian fi Finnish (v and w are primary equal) fi__phonebook Finnish (v and w as separate characters) fil Filipino fo Faroese fr_CA Canadian French gu Gujarati ha Hausa haw Hawaiian he Hebrew hi Hindi hr Croatian hu Hungarian hy Armenian ig Igbo is Icelandic ja Japanese [1] kk Kazakh kl Kalaallisut kn Kannada ko Korean [2] kok Konkani lkt Lakota ln Lingala lt Lithuanian lv Latvian mk Macedonian ml Malayalam mr Marathi mt Maltese nb Norwegian Bokmal nn Norwegian Nynorsk nso Northern Sotho om Oromo or Oriya pa Punjabi pl Polish ro Romanian sa Sanskrit se Northern Sami si Sinhala si__dictionary Sinhala (U+0DA5 = U+0DA2,0DCA,0DA4) sk Slovak sl Slovenian sq Albanian sr Serbian sr_Latn Serbian in Latin (tailored as Croatian) sv Swedish (v and w are primary equal) sv__reformed Swedish (v and w as separate characters) ta Tamil te Telugu th Thai tn Tswana to Tonga tr Turkish ug_Cyrl Uyghur in Cyrillic uk Ukrainian ur Urdu vi Vietnamese vo Volapu"k wae Walser wo Wolof yo Yoruba zh Chinese zh__big5han Chinese (ideographs: big5 order) zh__gb2312han Chinese (ideographs: GB-2312 order) zh__pinyin Chinese (ideographs: pinyin order) [3] zh__stroke Chinese (ideographs: stroke order) [3] zh__zhuyin Chinese (ideographs: zhuyin order) [3] -------------------------------------------------------------- Locales according to the default UCA rules include am (Amharic) without C<[reorder Ethi]>, bg (Bulgarian) without C<[reorder Cyrl]>, chr (Cherokee) without C<[reorder Cher]>, de (German), en (English), fr (French), ga (Irish), id (Indonesian), it (Italian), ka (Georgian) without C<[reorder Geor]>, mn (Mongolian) without C<[reorder Cyrl Mong]>, ms (Malay), nl (Dutch), pt (Portuguese), ru (Russian) without C<[reorder Cyrl]>, sw (Swahili), zu (Zulu). B [1] ja: Ideographs are sorted in JIS X 0208 order. Fullwidth and halfwidth forms are identical to their regular form. The difference between hiragana and katakana is at the 4th level, the comparison also requires C<(variable =E 'Non-ignorable')>, and then C has no effect. [2] ko: Plenty of ideographs are sorted by their reading. Such an ideograph is primary (level 1) equal to, and secondary (level 2) greater than, the corresponding hangul syllable. [3] zh__pinyin, zh__stroke and zh__zhuyin: implemented alt='short', where a smaller number of ideographs are tailored. =head2 A list of variant codes and their aliases variant code alias ------------------------------------------ dictionary dict phonebook phone phonebk reformed reform traditional trad ------------------------------------------ big5han big5 gb2312han gb2312 pinyin stroke zhuyin ------------------------------------------ Note: 'pinyin' is Han in Latin, 'zhuyin' is Han in Bopomofo. =head1 INSTALL Installation of C requires F, F, F and F. On building, C doesn't require any of F, F, and F. Tests for C are named F. =head1 CAVEAT =over 4 =item Tailoring is not maximum Even if a certain letter is tailored, its equivalent would not always tailored as well as it. For example, even though W is tailored, fullwidth W (C), W with acute (C), etc. are not tailored. The result may depend on whether source strings are normalized or not, and whether decomposed or composed. Thus C<(normalization =E undef)> is less preferred. =item Collation reordering is not supported The order of any groups including scripts is not changed. =back =head2 Reference locale based CLDR or other reference -------------------------------------------------------------------- af 30 = 1.8.1 ar 30 = 28 ("compat" wo [reorder Arab]) = 1.9.0 as 30 = 28 (without [reorder Beng..]) = 23 az 30 = 24 ("standard" wo [reorder Latn Cyrl]) be 30 = 28 (without [reorder Cyrl]) bn 30 = 28 ("standard" wo [reorder Beng..]) = 2.0.1 bs 30 = 28 (type="standard": [import hr]) bs_Cyrl 30 = 28 (type="standard": [import sr]) ca 30 = 23 (alt="proposed" type="standard") cs 30 = 1.8.1 (type="standard") cu 34 = 30 (without [reorder Cyrl]) cy 30 = 1.8.1 da 22.1 = 1.8.1 (type="standard") de__phonebook 30 = 2.0 (type="phonebook") de_AT_phonebook 30 = 27 (type="phonebook") dsb 30 = 26 ee 30 = 21 eo 30 = 1.8.1 es 30 = 1.9.0 (type="standard") es__traditional 30 = 1.8.1 (type="traditional") et 30 = 26 fa 22.1 = 1.8.1 fi 22.1 = 1.8.1 (type="standard" alt="proposed") fi__phonebook 22.1 = 1.8.1 (type="phonebook") fil 30 = 1.9.0 (type="standard") = 1.8.1 fo 22.1 = 1.8.1 (alt="proposed" type="standard") fr_CA 30 = 1.9.0 gu 30 = 28 ("standard" wo [reorder Gujr..]) = 1.9.0 ha 30 = 1.9.0 haw 30 = 24 he 30 = 28 (without [reorder Hebr]) = 23 hi 30 = 28 (without [reorder Deva..]) = 1.9.0 hr 30 = 28 ("standard" wo [reorder Latn Cyrl]) = 1.9.0 hu 22.1 = 1.8.1 (alt="proposed" type="standard") hy 30 = 28 (without [reorder Armn]) = 1.8.1 ig 30 = 1.8.1 is 22.1 = 1.8.1 (type="standard") ja 22.1 = 1.8.1 (type="standard") kk 30 = 28 (without [reorder Cyrl]) kl 22.1 = 1.8.1 (type="standard") kn 30 = 28 ("standard" wo [reorder Knda..]) = 1.9.0 ko 22.1 = 1.8.1 (type="standard") kok 30 = 28 (without [reorder Deva..]) = 1.8.1 lkt 30 = 25 ln 30 = 2.0 (type="standard") = 1.8.1 lt 22.1 = 1.9.0 lv 22.1 = 1.9.0 (type="standard") = 1.8.1 mk 30 = 28 (without [reorder Cyrl]) ml 22.1 = 1.9.0 mr 30 = 28 (without [reorder Deva..]) = 1.8.1 mt 22.1 = 1.9.0 nb 22.1 = 2.0 (type="standard") nn 22.1 = 2.0 (type="standard") nso [*] 26 = 1.8.1 om 22.1 = 1.8.1 or 30 = 28 (without [reorder Orya..]) = 1.9.0 pa 22.1 = 1.8.1 pl 30 = 1.8.1 ro 30 = 1.9.0 (type="standard") sa [*] 1.9.1 = 1.8.1 (type="standard" alt="proposed") se 22.1 = 1.8.1 (type="standard") si 30 = 28 ("standard" wo [reorder Sinh..]) = 1.9.0 si__dictionary 30 = 28 ("dictionary" wo [reorder Sinh..]) = 1.9.0 sk 22.1 = 1.9.0 (type="standard") sl 22.1 = 1.8.1 (type="standard" alt="proposed") sq 22.1 = 1.8.1 (alt="proposed" type="standard") sr 30 = 28 (without [reorder Cyrl]) sr_Latn 30 = 28 (type="standard": [import hr]) sv 22.1 = 1.9.0 (type="standard") sv__reformed 22.1 = 1.8.1 (type="reformed") ta 22.1 = 1.9.0 te 30 = 28 (without [reorder Telu..]) = 1.9.0 th 22.1 = 22 tn [*] 26 = 1.8.1 to 22.1 = 22 tr 22.1 = 1.8.1 (type="standard") uk 30 = 28 (without [reorder Cyrl]) ug_Cyrl https://en.wikipedia.org/wiki/Uyghur_Cyrillic_alphabet ur 22.1 = 1.9.0 vi 22.1 = 1.8.1 vo 30 = 25 wae 30 = 2.0 wo [*] 1.9.1 = 1.8.1 yo 30 = 1.8.1 zh 22.1 = 1.8.1 (type="standard") zh__big5han 22.1 = 1.8.1 (type="big5han") zh__gb2312han 22.1 = 1.8.1 (type="gb2312han") zh__pinyin 22.1 = 2.0 (type='pinyin' alt='short') zh__stroke 22.1 = 1.9.1 (type='stroke' alt='short') zh__zhuyin 22.1 = 22 (type='zhuyin' alt='short') -------------------------------------------------------------------- [*] http://www.unicode.org/repos/cldr/tags/latest/seed/collation/ =head1 AUTHOR The Unicode::Collate::Locale module for perl was written by SADAHIRO Tomoyuki, . This module is Copyright(C) 2004-2020, SADAHIRO Tomoyuki. Japan. All rights reserved. This module is free software; you can redistribute it and/or modify it under the same terms as Perl itself. =head1 SEE ALSO =over 4 =item Unicode Collation Algorithm - UTS #10 L =item The Default Unicode Collation Element Table (DUCET) L =item Unicode Locale Data Markup Language (LDML) - UTS #35 L =item CLDR - Unicode Common Locale Data Repository L =item L =item L =back =cut