/* -*- Mode: C; c-basic-offset:4 ; -*- */ /* * Copyright (c) 2004-2006 The Trustees of Indiana University and Indiana * University Research and Technology * Corporation. All rights reserved. * Copyright (c) 2004-2006 The University of Tennessee and The University * of Tennessee Research Foundation. All rights * reserved. * Copyright (c) 2004-2006 High Performance Computing Center Stuttgart, * University of Stuttgart. All rights reserved. * Copyright (c) 2004-2006 The Regents of the University of California. * All rights reserved. * Copyright (c) 2008 Cisco Systems, Inc. All rights reserved. * $COPYRIGHT$ * * Additional copyrights may follow * * $HEADER$ */ #ifndef OPAL_ARCH_H_HAS_BEEN_INCLUDED #define OPAL_ARCH_H_HAS_BEEN_INCLUDED #include "opal_config.h" #include #include /*************************************************** ** This file tries to classify the most relevant ** platforms regarding their data representation. ** Three aspects are important: ** - byte ordering (little or big endian) ** - integer representation ** - floating point representation. ** In addition, don't forget about the C/Fortran problems. ** *****************************************************/ /***************************************************************** ** Part 1: Integer representation. ** ** The following data types are considered relevant: ** ** short ** int ** long ** long long ** integer (fortran) ** ** The fortran integer is dismissed here, since there is no ** platform known to me, were fortran and C-integer do not match ** ** The following abbriviations are introduced: ** ** a) il32 (int long are 32 bits) (e.g. IA32 LINUX, SGI n32, SUN) ** ** short: 16 (else it would appear in the name) ** int: 32 ** long: 32 ** long long: 64 ** ** b) il64 ( int long are 64 bits) (e.g. Cray T3E ) ** short: 32 ** int: 64 ** long: 64 ** long long: 64 ** ** c) l64 (long are 64 bits) (e.g. SGI 64 IRIX, NEC SX5) ** ** short: 16 ** int: 32 ** long: 64 ** long long: 64 ** ***********************************************************************/ /********************************************************************* ** Part 2: Floating point representation ** ** The following datatypes are considered relevant ** ** float ** double ** long double ** real ** double precision ** ** Unfortunatly, here we have to take care, whether float and real, ** respectively double and double precision do match... ** ** a) fr32 (float and real are 32 bits) (e.g. SGI n32 and 64, SUN, NEC SX5,...) ** float: 32 ** double: 64 ** long double: 128 ** real: 32 ** double prec.:64 ** ** a1) fr32ld96 (float and real 32, long double 96) (e.g. IA32 LINUX gcc/icc) ** see a), except long double is 96 ** ** a2) fr32ld64 (e.g. IBM ) ** see a), except long double is 64 ** ** b) cray ( e.g. Cray T3E) ** float: 32 ** double: 64 ** long double: 64 ** real: 64 ** double prec.:64 ** ** ** Problem: long double is really treated differently on every machine. Therefore, ** we are storing besides the length of the long double also the length of the mantisee, ** and the number of *relevant* bits in the exponent. Here are the values: ** ** Architecture sizeof(long double) mantisee relevant bits for exp. ** ** SGIn32/64: 128 107 10 ** SUN(sparc): 128 113 14 ** IA64: 128 64 14 ** IA32: 96 64 14 ** Alpha: 128 113 14 ** 64 53 10 (gcc) ** IBM: 64 53 10 ** (128 106 10) (special flags required). ** SX5: 128 105 22 ** ** We will not implement all of these routiens, but we consider them ** now when defining the header-settings ** ***********************************************************************/ /******************************************************************** ** ** Classification of machines: ** ** IA32 LINUX: il32, fr32ld96, little endian ** SUN: il32, fr32, big endian ** SGI n32: il32, fr32, big endian ** SGI 64: l64, fr32, big endian ** NEC SX5: l64, fr32 big endian ** Cray T3E: il64, cray, big endian ** Cray X1: i32(+), fr32, big endian ** IBM: il32, fr32ld64, big endian ** ALPHA: l64, fr32, little endian ** ITANIUM: l64, fr32, little endian ** ** ** + sizeof ( long long ) not known ** ? alpha supports both, big and little endian ***********************************************************************/ /* Current conclusions: ** we need at the moment three settings: ** - big/little endian ? ** - is long 32 or 64 bits ? ** - is long double 64, 96 or 128 bits ? ** - no. of rel. bits in the exponent of a long double ( 10 or 14 ) ** - no. of bits of the mantiss of a long double ( 53, 64, 105, 106, 107, 113 ) ** ** To store this in a 32 bit integer, we use the following definition: ** ** 1 2 3 4 ** 12345678 12345678 12345678 12345678 ** ** 1. Byte: ** bits 1 & 2: 00 (header) (to recognize the correct end) ** bits 3 & 4: encoding: 00 = little, 01 = big ** bits 5 & 6: reserved for later use. currently set to 00 ** bits 7 & 8: reserved for later use. currently set to 00 ** 2. Byte: ** bits 1 & 2: length of long: 00 = 32, 01 = 64 ** bits 3 & 4: lenght of long long (not used currently, set to 00). ** bits 5 & 6: length of C/C++ bool (00 = 8, 01 = 16, 10 = 32) ** bits 7 & 8: length of Fortran Logical (00 = 8, 01 = 16, 10 = 32) ** 3. Byte: ** bits 1 & 2: length of long double: 00=64, 01=96,10 = 128 ** bits 3 & 4: no. of rel. bits in the exponent: 00 = 10, 01 = 14) ** bits 5 - 7: no. of bits of mantisse ( 000 = 53, 001 = 64, 010 = 105, ** 011 = 106, 100 = 107,101 = 113 ) ** bit 8: intel or sparc representation of mantisse (0 = sparc, ** 1 = intel ) ** 4. Byte: ** bits 1 & 2: 11 (header) (to recognize the correct end) ** bits 3 & 4: reserved for later use. currently set to 11 ** bits 5 & 6: reserved for later use. currently set to 11 ** bits 7 & 8: reserved for later use. currently set to 11 */ /* These masks implement the specification above above */ #define OPAL_ARCH_HEADERMASK 0x03000000 /* set the fields for the header */ #define OPAL_ARCH_HEADERMASK2 0x00000003 /* other end, needed for checks */ #define OPAL_ARCH_UNUSEDMASK 0xfc000000 /* mark the unused fields */ /* BYTE 1 */ #define OPAL_ARCH_ISBIGENDIAN 0x00000008 /* BYTE 2 */ #define OPAL_ARCH_LONGISxx 0x0000c000 /* mask for sizeof long */ #define OPAL_ARCH_LONGIS64 0x00001000 #define OPAL_ARCH_LONGLONGISxx 0x00003000 /* mask for sizeof long long */ #define OPAL_ARCH_BOOLISxx 0x00000c00 /* mask for sizeof bool */ #define OPAL_ARCH_BOOLIS8 0x00000000 /* bool is 8 bits */ #define OPAL_ARCH_BOOLIS16 0x00000400 /* bool is 16 bits */ #define OPAL_ARCH_BOOLIS32 0x00000800 /* bool is 32 bits */ #define OPAL_ARCH_LOGICALISxx 0x00000300 /* mask for sizeof Fortran logical */ #define OPAL_ARCH_LOGICALIS8 0x00000000 /* logical is 8 bits */ #define OPAL_ARCH_LOGICALIS16 0x00000100 /* logical is 16 bits */ #define OPAL_ARCH_LOGICALIS32 0x00000200 /* logical is 32 bits */ /* BYTE 3 */ #define OPAL_ARCH_LONGDOUBLEIS96 0x00020000 #define OPAL_ARCH_LONGDOUBLEIS128 0x00010000 #define OPAL_ARCH_LDEXPSIZEIS15 0x00080000 #define OPAL_ARCH_LDMANTDIGIS64 0x00400000 #define OPAL_ARCH_LDMANTDIGIS105 0x00200000 #define OPAL_ARCH_LDMANTDIGIS106 0x00600000 #define OPAL_ARCH_LDMANTDIGIS107 0x00100000 #define OPAL_ARCH_LDMANTDIGIS113 0x00500000 #define OPAL_ARCH_LDISINTEL 0x00800000 BEGIN_C_DECLS OPAL_DECLSPEC int32_t opal_arch_compute_local_id( uint32_t *var); OPAL_DECLSPEC int32_t opal_arch_checkmask ( uint32_t *var, uint32_t mask ); static inline int32_t opal_arch_isbigendian ( void ) { const uint32_t value = 0x12345678; const char *ptr = (char*)&value; int x = 0; /* if( sizeof(int) == 8 ) x = 4; */ if( ptr[x] == 0x12) return 1; /* big endian, true */ if( ptr[x] == 0x78 ) return 0; /* little endian, false */ assert( 0 ); /* unknown architecture not little nor big endian */ return -1; } /* we must find which representation of long double is used * intel or sparc. Both of them represent the long doubles using a close to * IEEE representation (seeeeeee..emmm...m) where the mantissa look like * 1.????. For the intel representaion the 1 is explicit, and for the sparc * the first one is implicit. If we take the number 2.0 the exponent is 1 * and the mantissa is 1.0 (the sign of course should be 0). So if we check * for the first one in the binary representation of the number, we will * find the bit from the exponent, so the next one should be the begining * of the mantissa. If it's 1 then we have an intel representaion, if not * we have a sparc one. QED */ static inline int32_t opal_arch_ldisintel( void ) { long double ld = 2.0; int i, j; uint32_t* pui = (uint32_t*)(void*)&ld; j = LDBL_MANT_DIG / 32; i = (LDBL_MANT_DIG % 32) - 1; if( opal_arch_isbigendian() ) { /* big endian */ j = (sizeof(long double) / sizeof(unsigned int)) - j; if( i < 0 ) { i = 31; j = j+1; } } else { if( i < 0 ) { i = 31; j = j-1; } } return (pui[j] & (1 << i) ? 1 : 0); } static inline void opal_arch_setmask ( uint32_t *var, uint32_t mask) { *var |= mask; } END_C_DECLS #endif /* OPAL_ARCH_H_HAS_BEEN_INCLUDED */