summaryrefslogtreecommitdiff
path: root/Build/source/libs/icu/icu-src/source/layout/KhmerReordering.h
diff options
context:
space:
mode:
Diffstat (limited to 'Build/source/libs/icu/icu-src/source/layout/KhmerReordering.h')
-rw-r--r--Build/source/libs/icu/icu-src/source/layout/KhmerReordering.h132
1 files changed, 0 insertions, 132 deletions
diff --git a/Build/source/libs/icu/icu-src/source/layout/KhmerReordering.h b/Build/source/libs/icu/icu-src/source/layout/KhmerReordering.h
deleted file mode 100644
index eecd12dff63..00000000000
--- a/Build/source/libs/icu/icu-src/source/layout/KhmerReordering.h
+++ /dev/null
@@ -1,132 +0,0 @@
-/*
- * (C) Copyright IBM Corp. 1998-2005 - All Rights Reserved
- *
- * This file is a modification of the ICU file IndicReordering.h
- * by Jens Herden and Javier Sola for Khmer language
- *
- */
-
-#ifndef __KHMERREORDERING_H
-#define __KHMERREORDERING_H
-
-/**
- * \file
- * \internal
- */
-
-#include "LETypes.h"
-#include "OpenTypeTables.h"
-
-U_NAMESPACE_BEGIN
-
-class LEGlyphStorage;
-
-// Vocabulary
-// Base -> A consonant or an independent vowel in its full (not subscript) form. It is the
-// center of the syllable, it can be souranded by coeng (subscript) consonants, vowels,
-// split vowels, signs... but there is only one base in a syllable, it has to be coded as
-// the first character of the syllable.
-// split vowel --> vowel that has two parts placed separately (e.g. Before and after the consonant).
-// Khmer language has five of them. Khmer split vowels either have one part before the
-// base and one after the base or they have a part before the base and a part above the base.
-// The first part of all Khmer split vowels is the same character, identical to
-// the glyph of Khmer dependent vowel SRA EI
-// coeng --> modifier used in Khmer to construct coeng (subscript) consonants
-// Differently than indian languages, the coeng modifies the consonant that follows it,
-// not the one preceding it Each consonant has two forms, the base form and the subscript form
-// the base form is the normal one (using the consonants code-point), the subscript form is
-// displayed when the combination coeng + consonant is encountered.
-// Consonant of type 1 -> A consonant which has subscript for that only occupies space under a base consonant
-// Consonant of type 2.-> Its subscript form occupies space under and before the base (only one, RO)
-// Consonant of Type 3 -> Its subscript form occupies space under and after the base (KHO, CHHO, THHO, BA, YO, SA)
-// Consonant shifter -> Khmer has to series of consonants. The same dependent vowel has different sounds
-// if it is attached to a consonant of the first series or a consonant of the second series
-// Most consonants have an equivalent in the other series, but some of theme exist only in
-// one series (for example SA). If we want to use the consonant SA with a vowel sound that
-// can only be done with a vowel sound that corresponds to a vowel accompanying a consonant
-// of the other series, then we need to use a consonant shifter: TRIISAP or MUSIKATOAN
-// x17C9 y x17CA. TRIISAP changes a first series consonant to second series sound and
-// MUSIKATOAN a second series consonant to have a first series vowel sound.
-// Consonant shifter are both normally supercript marks, but, when they are followed by a
-// superscript, they change shape and take the form of subscript dependent vowel SRA U.
-// If they are in the same syllable as a coeng consonant, Unicode 3.0 says that they
-// should be typed before the coeng. Unicode 4.0 breaks the standard and says that it should
-// be placed after the coeng consonant.
-// Dependent vowel -> In khmer dependent vowels can be placed above, below, before or after the base
-// Each vowel has its own position. Only one vowel per syllable is allowed.
-// Signs -> Khmer has above signs and post signs. Only one above sign and/or one post sign are
-// Allowed in a syllable.
-//
-//
-
-struct KhmerClassTable // This list must include all types of components that can be used inside a syllable
-{
- enum CharClassValues // order is important here! This order must be the same that is found in each horizontal
- // line in the statetable for Khmer (file KhmerReordering.cpp).
- {
- CC_RESERVED = 0,
- CC_CONSONANT = 1, // consonant of type 1 or independent vowel
- CC_CONSONANT2 = 2, // Consonant of type 2
- CC_CONSONANT3 = 3, // Consonant of type 3
- CC_ZERO_WIDTH_NJ_MARK = 4, // Zero Width non joiner character (0x200C)
- CC_CONSONANT_SHIFTER = 5,
- CC_ROBAT = 6, // Khmer special diacritic accent -treated differently in state table
- CC_COENG = 7, // Subscript consonant combining character
- CC_DEPENDENT_VOWEL = 8,
- CC_SIGN_ABOVE = 9,
- CC_SIGN_AFTER = 10,
- CC_ZERO_WIDTH_J_MARK = 11, // Zero width joiner character
- CC_COUNT = 12 // This is the number of character classes
- };
-
- enum CharClassFlags
- {
- CF_CLASS_MASK = 0x0000FFFF,
-
- CF_CONSONANT = 0x01000000, // flag to speed up comparing
- CF_SPLIT_VOWEL = 0x02000000, // flag for a split vowel -> the first part is added in front of the syllable
- CF_DOTTED_CIRCLE = 0x04000000, // add a dotted circle if a character with this flag is the first in a syllable
- CF_COENG = 0x08000000, // flag to speed up comparing
- CF_SHIFTER = 0x10000000, // flag to speed up comparing
- CF_ABOVE_VOWEL = 0x20000000, // flag to speed up comparing
-
- // position flags
- CF_POS_BEFORE = 0x00080000,
- CF_POS_BELOW = 0x00040000,
- CF_POS_ABOVE = 0x00020000,
- CF_POS_AFTER = 0x00010000,
- CF_POS_MASK = 0x000f0000
- };
-
- typedef le_uint32 CharClass;
-
- typedef le_int32 ScriptFlags;
-
- LEUnicode firstChar; // for Khmer this will become x1780
- LEUnicode lastChar; // and this x17DF
- const CharClass *classTable;
-
- CharClass getCharClass(LEUnicode ch) const;
-
- static const KhmerClassTable *getKhmerClassTable();
-};
-
-
-class KhmerReordering /* not : public UObject because all methods are static */ {
-public:
- static le_int32 reorder(const LEUnicode *theChars, le_int32 charCount, le_int32 scriptCode,
- LEUnicode *outChars, LEGlyphStorage &glyphStorage);
-
- static const FeatureMap *getFeatureMap(le_int32 &count);
-
-private:
- // do not instantiate
- KhmerReordering();
-
- static le_int32 findSyllable(const KhmerClassTable *classTable, const LEUnicode *chars, le_int32 prev, le_int32 charCount);
-
-};
-
-
-U_NAMESPACE_END
-#endif