Don't use C comments; replace with C++ comments

X-SVN-Rev: 1438
This commit is contained in:
Alan Liu 2000-05-23 16:47:48 +00:00
parent aca0c53ac2
commit ed4a6972fc
3 changed files with 67 additions and 75 deletions

View File

@ -5,8 +5,8 @@
*******************************************************************************
*
* $Source: /xsrl/Nsvn/icu/icu4j/src/com/ibm/text/resources/Attic/TransliterationRule_Fullwidth_Halfwidth.java,v $
* $Date: 2000/03/10 04:07:29 $
* $Revision: 1.2 $
* $Date: 2000/05/23 16:47:48 $
* $Revision: 1.3 $
*
*****************************************************************************************
*/
@ -22,8 +22,7 @@ public class TransliterationRule_Fullwidth_Halfwidth extends ListResourceBundle
return new Object[][] {
{ "Rule", ""
/* Mechanically generated from Unicode Character Database
*/
// Mechanically generated from Unicode Character Database
// multicharacter

View File

@ -5,8 +5,8 @@
*******************************************************************************
*
* $Source: /xsrl/Nsvn/icu/icu4j/src/com/ibm/text/resources/Attic/TransliterationRule_Latin_Cyrillic.java,v $
* $Date: 2000/04/26 21:34:16 $
* $Revision: 1.4 $
* $Date: 2000/05/23 16:47:48 $
* $Revision: 1.5 $
*
*****************************************************************************************
*/
@ -22,12 +22,11 @@ public class TransliterationRule_Latin_Cyrillic extends ListResourceBundle {
return new Object[][] {
{ "Rule", ""
/* This class is designed to be a general Latin-Cyrillic
transliteration. The standard Russian transliterations
are generally used for the letters from Russian,
with additional Cyrillic characters given consistent
mappings.
*/
// This class is designed to be a general Latin-Cyrillic
// transliteration. The standard Russian transliterations
// are generally used for the letters from Russian,
// with additional Cyrillic characters given consistent
// mappings.
+ "$S_hacek=\u0160;"
+ "$s_hacek=\u0161;"
@ -126,12 +125,10 @@ public class TransliterationRule_Latin_Cyrillic extends ListResourceBundle {
+ "$letter=[[:Lu:][:Ll:]];"
+ "$lower=[[:Ll:]];"
/*
Modified to combine display transliterator and typing transliterator.
The display mapping uses accents for the "soft" vowels.
It does not, although it could, use characters like \u0161 instead of digraphs
like sh.
*/
// Modified to combine display transliterator and typing transliterator.
// The display mapping uses accents for the "soft" vowels.
// It does not, although it could, use characters like \u0161 instead of digraphs
// like sh.
// #############################################
// Special titlecase forms, not duplicated

View File

@ -5,8 +5,8 @@
*******************************************************************************
*
* $Source: /xsrl/Nsvn/icu/icu4j/src/com/ibm/text/resources/Attic/TransliterationRule_Latin_Kana.java,v $
* $Date: 2000/04/26 21:22:25 $
* $Revision: 1.5 $
* $Date: 2000/05/23 16:47:48 $
* $Revision: 1.6 $
*
*****************************************************************************************
*/
@ -24,53 +24,53 @@ public class TransliterationRule_Latin_Kana extends ListResourceBundle {
// Uppercase Latin to katakana
{ "Rule", ""
//# $Revision: 1.5 $
//# $Revision: 1.6 $
// Transliteration rules for Japanese Hiragana and Katakana to
// romaji
// lower case roman generates hiragana.
// upper case roman generates katakana.
// Uses modified Hepburn. Small changes to make unambiguous.
// Kunrei-shiki: Hepburn/MHepburn
/*
si: shi
si ~ya: sha
si ~yu: shu
si ~yo: sho
zi: ji
zi ~ya: ja
zi ~yu: ju
zi ~yo: jo
ti: chi
ti ~ya: cha
ti ~yu: chu
ti ~yu: cho
tu: tsu
di: ji/dji
du: zu/dzu
hu: fu
// for foreign words
se ~i si
si ~e she
ze ~i zi
zi ~e je
te ~i ti
ti ~e che
te ~u tu
de ~i di
de ~u du
de ~i di
he ~u: hu
hu ~a fa
hu ~i fi
hu ~e he
hu ~o ho
// Most small forms are generated, but if necessary
// explicit small forms are given with ~a, ~ya, etc.
*/
//| si: shi
//| si ~ya: sha
//| si ~yu: shu
//| si ~yo: sho
//| zi: ji
//| zi ~ya: ja
//| zi ~yu: ju
//| zi ~yo: jo
//| ti: chi
//| ti ~ya: cha
//| ti ~yu: chu
//| ti ~yu: cho
//| tu: tsu
//| di: ji/dji
//| du: zu/dzu
//| hu: fu
//| // for foreign words
//| se ~i si
//| si ~e she
//|
//| ze ~i zi
//| zi ~e je
//|
//| te ~i ti
//| ti ~e che
//| te ~u tu
//|
//| de ~i di
//| de ~u du
//| de ~i di
//|
//| he ~u: hu
//| hu ~a fa
//| hu ~i fi
//| hu ~e he
//| hu ~o ho
//| // Most small forms are generated, but if necessary
//| // explicit small forms are given with ~a, ~ya, etc.
//#######################################
// Definitions of variables to be substituted
//#######################################
@ -548,13 +548,11 @@ public class TransliterationRule_Latin_Kana extends ListResourceBundle {
// ALL ROMAJI MAPS TO KANA
// ########################################
/*
+ "SH>$SI;"
+ "TS>$TU;"
+ "CH>$TI;"
+ "DJ>$DI;"
+ "DZ>$DU;"
*/
//| + "SH>$SI;"
//| + "TS>$TU;"
//| + "CH>$TI;"
//| + "DJ>$DI;"
//| + "DZ>$DU;"
// THE FOLLOWING ARE NOT REALLY NECESSARY, BUT PRODUCE
// SLIGHTLY MORE NATURAL RESULTS.
@ -1025,13 +1023,11 @@ public class TransliterationRule_Latin_Kana extends ListResourceBundle {
// all romaji maps to kana
// ########################################
/*
+ "sh>$si;"
+ "ts>$tu;"
+ "ch>$ti;"
+ "dj>$di;"
+ "dz>$du;"
*/
//| + "sh>$si;"
//| + "ts>$tu;"
//| + "ch>$ti;"
//| + "dj>$di;"
//| + "dz>$du;"
// the following are not really necessary, but produce
// slightly more natural results.