home
***
CD-ROM
|
disk
|
FTP
|
other
***
search
/
Columbia Kermit
/
kermit.zip
/
old
/
ckermit70
/
ckcxla.h
< prev
next >
Wrap
C/C++ Source or Header
|
2020-01-01
|
10KB
|
333 lines
/*
File CKCXLA.H
System-independent character-set translation header file for C-Kermit.
*/
/*
Author: Frank da Cruz <fdc@columbia.edu>,
The Kermit Project - Columbia University, New York City.
Copyright (C) 1985, 2000,
Trustees of Columbia University in the City of New York.
All rights reserved. See the C-Kermit COPYING.TXT file or the
copyright text in the ckcmai.c module for disclaimer and permissions.
*/
/*
NOTE:
ISO 204 is Latin-1 + Euro.
ISO 205 is Latin-4 + Euro.
ISO 206 is Latin-7 + Euro.
*/
#ifndef CKCXLA_H /* Guard against multiple inclusion */
#define CKCXLA_H
#ifndef KANJI /* Systems supporting Kanji */
#ifdef OS2
#define KANJI
#endif /* OS2 */
#endif /* KANJI */
#ifdef NOKANJI /* Except if NOKANJI is defined. */
#ifdef KANJI
#undef KANJI
#endif /* KANJI */
#endif /* NOKANJI */
#ifndef NOUNICODE
#ifndef UNICODE /* Unicode support */
#ifdef OS2ORUNIX /* Only for K95, UNIX, VMS,... */
#define UNICODE
#else
#ifdef VMS
#define UNICODE
#endif /* VMS */
#endif /* OS2ORUNIX */
#endif /* UNICODE */
#endif /* NOUNICODE */
#define XLA_NONE 0 /* Translation types - none */
#define XLA_BYTE 1 /* Byte-for-byte */
#define XLA_JAPAN 2 /* Japanese */
#define XLA_UNICODE 3 /* Unicode */
#ifndef UNIORKANJI /* Unicode OR Kanji */
#ifdef UNICODE /* i.e. some support for */
#define UNIORKANJI /* multibyte character sets */
#endif /* UNICODE */
#ifdef KANJI
#define UNIORKANJI
#endif /* KANJI */
#endif /* UNIORKANJI */
/*
Disable all support for all classes of character sets
if NOCSETS is defined.
*/
#ifdef NOCSETS
#ifdef CKOUNI
#undef CKOUNI
#endif /* CKOUNI */
#ifdef KANJI
#undef KANJI
#endif /* KANJI */
#ifdef CYRILLIC
#undef CYRILLIC
#endif /* CYRILLIC */
#ifdef LATIN2
#undef LATIN2
#endif /* LATIN2 */
#ifdef HEBREW
#undef HEBREW
#endif /* HEBREW */
#ifdef UNICODE
#undef UNICODE
#endif /* UNICODE */
#else /* Not NOCSETS - Rest of this file... */
#ifdef NOUNICODE /* Unicode */
#ifdef UNICODE
#undef UNICODE
#endif /* UNICODE */
#endif /* NOUNICODE */
#ifdef UNICODE
#ifdef OS2
#ifndef CKOUNI
#define CKOUNI /* Special Unicode features for K95 */
#endif /* CKOUNI */
#endif /* OS2 */
#endif /* UNICODE */
#ifndef OS2
#ifdef CKOUNI
#undef CKOUNI
#endif /* CKOUNI */
#endif /* OS2 */
#ifndef NOLATIN2 /* If they didn't say "no Latin-2" */
#ifndef LATIN2 /* Then if LATIN2 isn't already */
#define LATIN2 /* defined, define it. */
#endif /* LATIN2 */
#endif /* NOLATIN2 */
#ifdef NOCYRILLIC /* (spelling variant...) */
#ifndef NOCYRIL
#define NOCYRIL
#endif /* NOCYRIL */
#endif /* NOCYRILLIC */
#ifndef NOCYRIL /* If they didn't say "no Cyrillic" */
#ifndef CYRILLIC /* Then if CYRILLIC isn't already */
#define CYRILLIC /* defined, define it. */
#endif /* CYRILLIC */
#endif /* NOCYRIL */
#ifndef NOHEBREW /* If they didn't say "no Hebrew" */
#ifndef HEBREW /* Then if HEBREW isn't already */
#define HEBREW /* defined, define it. */
#endif /* HEBREW */
#endif /* NOHEBREW */
#ifndef NOGREEK /* If not no Greek */
#ifndef GREEK /* then if GREEK isn't already */
#define GREEK /* defined, define it. */
#endif /* GREEK */
#endif /* NOGREEK */
#ifndef NOKANJI /* If not no Kanji */
#ifndef KANJI /* then if KANJI isn't already */
#define KANJI /* defined, define it. */
#endif /* KANJI */
#endif /* NOKANJI */
/* File ckcxla.h -- Character-set-related definitions, system independent */
/* Codes for Kermit Transfer Syntax Level (obsolete) */
#define TS_L0 0 /* Level 0 (Transparent) */
#define TS_L1 1 /* Level 1 (one standard character set) */
#define TS_L2 2 /* Level 2 (multiple character sets in same file) */
#define UNK 63 /* Symbol to use for unknown character (63 = ?) */
/*
Codes for the base alphabet of a given character set.
These are assigned in roughly ISO 8859 order.
(Each is assumed to include ASCII/Roman.)
*/
#define AL_UNIV 0 /* Universal (like ISO 10646) */
#define AL_ROMAN 1 /* Roman (Latin) alphabet */
#define AL_CYRIL 2 /* Cyrillic alphabet */
#define AL_ARABIC 3 /* Arabic */
#define AL_GREEK 4 /* Greek */
#define AL_HEBREW 5 /* Hebrew */
#define AL_KANA 6 /* Japanese Katakana */
#define AL_JAPAN 7 /* Japanese Katakana+Kanji ideograms */
#define AL_HAN 8 /* Chinese/Japanese/Korean ideograms */
#define AL_INDIA 9 /* Indian scripts (ISCII) */
#define AL_VIET 10 /* Vietnamese (VISCII) */
/* Add more here... */
#define AL_UNK 999 /* Unknown (transparent) */
/* Codes for languages */
/*
NOTE: It would perhaps be better to use ISO 639-1988 2-letter "Codes for
Representation of Names of Languages" here, shown in the comments below.
*/
#define L_ASCII 0 /* EN ASCII, English */
#define L_USASCII 0 /* EN ASCII, English */
#define L_DUTCH 1 /* NL Dutch */
#define L_FINNISH 2 /* FI Finnish */
#define L_FRENCH 3 /* FR French */
#define L_GERMAN 4 /* DE German */
#define L_HUNGARIAN 5 /* HU Hungarian */
#define L_ITALIAN 6 /* IT Italian */
#define L_NORWEGIAN 7 /* NO Norwegian */
#define L_PORTUGUESE 8 /* PT Portuguese */
#define L_SPANISH 9 /* ES Spanish */
#define L_SWEDISH 10 /* SV Swedish */
#define L_SWISS 11 /* RM Swiss (Rhaeto-Romance) */
#define L_DANISH 12 /* DA Danish */
#define L_ICELANDIC 13 /* IS Icelandic */
#define L_RUSSIAN 14 /* RU Russian */
#define L_JAPANESE 15 /* JA Japanese */
#define L_HEBREW 16 /* IW Hebrew */
#define L_GREEK 17 /* Greek */
#define MAXLANG 17 /* Number of languages */
/*
File character-sets are defined in the system-specific ck?xla.h file,
except for the following ones, which must be available to all versions:
*/
#define FC_TRANSP 254 /* Transparent */
#define FC_UNDEF 255 /* Undefined */
/*
Designators for Kermit's transfer character sets. These are all standard
sets, or based on them. Symbols must be unique in the first 8 characters,
because some C preprocessors have this limit.
*/
/* LIST1 */
#define TC_TRANSP 0 /* Transparent, no character translation */
#define TC_USASCII 1 /* ISO 646 IRV / US 7-bit ASCII */
#define TC_1LATIN 2 /* ISO 8859-1, Latin Alphabet 1 */
#define TC_2LATIN 3 /* ISO 8859-2, Latin Alphabet 2 */
#define TC_CYRILL 4 /* ISO 8859-5, Latin/Cyrillic */
#define TC_JEUC 5 /* Japanese EUC = JIS 0201+0202+0208 */
#define TC_HEBREW 6 /* ISO 8859-8, Latin/Hebrew */
#define TC_GREEK 7 /* ISO 8859-7, Latin/Greek */
#define TC_9LATIN 8 /* ISO 8859-15 Latin Alphabet 9 (with Euro) */
#define TC_UCS2 9 /* ISO 10646 / Unicode UCS-2 */
#define TC_UTF8 10 /* ISO 10646 / Unicode UTF-8 */
#define MAXTCSETS 10 /* Highest Transfer Character Set Number */
#ifdef COMMENT
/*
Not used and probably won't be due to ISO-10646 / Unicode.
*/
#define TC_3LATIN 11 /* ISO 8859-3, Latin-3 */
#define TC_4LATIN 12 /* ISO 8859-4, Latin-4 */
#define TC_5LATIN 13 /* ISO 8859-9, Latin-5 */
#define TC_ARABIC 14 /* ISO-8859-6, Latin/Arabic */
#define TC_JIS208 15 /* Japanese JIS X 0208 multibyte set */
#define TC_CHINES 16 /* Chinese Standard GB 2312-80 */
#define TC_KOREAN 17 /* Korean KS C 5601-1987 */
#define TC_ISCII 18 /* Indian standard code for ii... */
#define TC_VSCII 19 /* Vietnam standard code for ii... */
/* etc... */
#endif /* COMMENT */
/* Structure for character-set information */
struct csinfo {
char *name; /* Descriptive name of character set */
int size; /* Size (e.g. 128, 256, 16384) */
int code; /* Like TC_1LATIN, etc. */
char *designator; /* Designator, like I2/100 = Latin-1 */
int alphabet; /* Base alphabet */
char *keyword; /* Keyword for this character-set */
};
/* Structure for language information */
struct langinfo {
int id; /* Language ID code (L_whatever) */
int fc; /* File character set to use */
int tc; /* Transfer character set to use */
char *description; /* Description of language */
};
/* Now take in the system-specific definitions */
#ifdef UNIX
#include "ckuxla.h"
#endif /* UNIX */
#ifdef OSK /* OS-9 */
#include "ckuxla.h"
#endif /* OS-9 */
#ifdef VMS /* VAX/VMS */
#include "ckuxla.h"
#endif /* VMS */
#ifdef GEMDOS /* Atari ST */
#include "ckuxla.h"
#endif /* GEMDOS */
#ifdef MAC /* Macintosh */
#include "ckmxla.h"
#endif /* MAC */
#ifdef OS2 /* OS/2 */
#include "ckuxla.h" /* Uses big UNIX version */
#endif /* OS2 */
#ifdef AMIGA /* Commodore Amiga */
#include "ckuxla.h"
#endif /* AMIGA */
#ifdef datageneral /* Data General MV AOS/VS */
#include "ckuxla.h"
#endif /* datageneral */
#ifdef STRATUS /* Stratus Computer, Inc. VOS */
#include "ckuxla.h"
#endif /* STRATUS */
#ifdef UNICODE
#include "ckcuni.h" /* Unicode */
#endif /* UNICODE */
#ifdef KANJI
#define UNKSJIS 0x817f
_PROTOTYP(USHORT eu_to_sj, (USHORT) ); /* EUC-JP to Shift-JIS */
_PROTOTYP(USHORT sj_to_eu, (USHORT) ); /* Shift-JIS to EUC-JP */
_PROTOTYP( int xkanjf, (void) );
_PROTOTYP( int xkanji, (int, int (*)(char)) );
_PROTOTYP( int xkanjz, (int (*)(char) ) );
_PROTOTYP( int zkanjf, (void) );
_PROTOTYP( int zkanji, (int (*)(void)) ); /* Kanji function prototypes */
_PROTOTYP( int zkanjz, (void) );
_PROTOTYP(VOID j7init, ( void ) ); /* Initialize JIS-7 parser */
_PROTOTYP(int getj7, ( void ) ); /* Get next JIS-7 character */
#endif /* KANJI */
#ifndef MAC
#ifndef NOLOCAL
_PROTOTYP( int cs_size, (int) );
_PROTOTYP( int cs_is_std, (int) );
_PROTOTYP( int cs_is_nrc, (int) );
#endif /* NOLOCAL */
#endif /* MAC */
_PROTOTYP(VOID setxlatype, (int, int));
#endif /* NOCSETS */
#endif /* CKCXLA_H */
/* End of ckcxla.h */