| 1 |
/* Manipulating the FPU control word. -*- coding: utf-8 -*- |
| 2 |
Copyright (C) 2007-2026 Free Software Foundation, Inc. |
| 3 |
Written by Bruno Haible <[email protected]>, 2007. |
| 4 |
|
| 5 |
This file is free software: you can redistribute it and/or modify |
| 6 |
it under the terms of the GNU Lesser General Public License as |
| 7 |
published by the Free Software Foundation; either version 2.1 of the |
| 8 |
License, or (at your option) any later version. |
| 9 |
|
| 10 |
This file is distributed in the hope that it will be useful, |
| 11 |
but WITHOUT ANY WARRANTY; without even the implied warranty of |
| 12 |
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the |
| 13 |
GNU Lesser General Public License for more details. |
| 14 |
|
| 15 |
You should have received a copy of the GNU Lesser General Public License |
| 16 |
along with this program. If not, see <https://www.gnu.org/licenses/>. */ |
| 17 |
|
| 18 |
#ifndef _FPUCW_H |
| 19 |
#define _FPUCW_H |
| 20 |
|
| 21 |
/* The i386 floating point hardware (the 387 compatible FPU, not the modern |
| 22 |
SSE/SSE2 hardware) has a controllable rounding precision. It is specified |
| 23 |
through the 'PC' bits in the FPU control word ('fctrl' register). (See |
| 24 |
the GNU libc i386 <fpu_control.h> header for details.) |
| 25 |
|
| 26 |
On some platforms, such as Linux or Solaris, the default precision setting |
| 27 |
is set to "extended precision". This means that 'long double' instructions |
| 28 |
operate correctly, but 'double' computations often produce slightly |
| 29 |
different results as on strictly IEEE 754 conforming systems. |
| 30 |
|
| 31 |
On some platforms, such as NetBSD, the default precision is set to |
| 32 |
"double precision". This means that 'long double' instructions will operate |
| 33 |
only as 'double', i.e. lead to wrong results. Similarly on FreeBSD 6.4, at |
| 34 |
least for the division of 'long double' numbers. |
| 35 |
|
| 36 |
The FPU control word is under control of the application, i.e. it is |
| 37 |
not required to be set either way by the ABI. (In fact, the i386 ABI |
| 38 |
https://www.linux-mips.org/pub/linux/mips/doc/ABI/abi386-4.pdf page 3-12 = page 38 |
| 39 |
is not clear about it. But in any case, gcc treats the control word |
| 40 |
like a "preserved" register: it emits code that assumes that the control |
| 41 |
word is preserved across calls, and it restores the control word at the |
| 42 |
end of functions that modify it.) |
| 43 |
|
| 44 |
See Vincent Lefèvre's page https://www.vinc17.net/research/extended.en.html |
| 45 |
for a good explanation. |
| 46 |
See https://web.archive.org/web/20060905133417/http://www.uwsg.iu.edu/hypermail/linux/kernel/0103.0/0453.html |
| 47 |
some argumentation which setting should be the default. */ |
| 48 |
|
| 49 |
/* This header file provides the following facilities: |
| 50 |
fpucw_t integral type holding the value of 'fctrl' |
| 51 |
FPU_PC_MASK bit mask denoting the precision control |
| 52 |
FPU_PC_DOUBLE precision control for 53 bits mantissa |
| 53 |
FPU_PC_EXTENDED precision control for 64 bits mantissa |
| 54 |
GET_FPUCW () yields the current FPU control word |
| 55 |
SET_FPUCW (word) sets the FPU control word |
| 56 |
DECL_LONG_DOUBLE_ROUNDING variable declaration for |
| 57 |
BEGIN/END_LONG_DOUBLE_ROUNDING |
| 58 |
BEGIN_LONG_DOUBLE_ROUNDING () starts a sequence of instructions with |
| 59 |
'long double' safe operation precision |
| 60 |
END_LONG_DOUBLE_ROUNDING () ends a sequence of instructions with |
| 61 |
'long double' safe operation precision |
| 62 |
*/ |
| 63 |
|
| 64 |
#ifdef __cplusplus |
| 65 |
extern "C" { |
| 66 |
#endif |
| 67 |
|
| 68 |
|
| 69 |
/* Inline assembler like this works only with GNU C and clang. */ |
| 70 |
#if (defined __i386__ || defined __x86_64__) && (defined __GNUC__ || defined __clang__) |
| 71 |
|
| 72 |
typedef unsigned short fpucw_t; /* glibc calls this fpu_control_t */ |
| 73 |
|
| 74 |
# define FPU_PC_MASK 0x0300 |
| 75 |
# define FPU_PC_DOUBLE 0x200 /* glibc calls this _FPU_DOUBLE */ |
| 76 |
# define FPU_PC_EXTENDED 0x300 /* glibc calls this _FPU_EXTENDED */ |
| 77 |
|
| 78 |
# define GET_FPUCW() __extension__ \ |
| 79 |
({ fpucw_t _cw; \ |
| 80 |
__asm__ __volatile__ ("fnstcw %0" : "=m" (*&_cw)); \ |
| 81 |
_cw; \ |
| 82 |
}) |
| 83 |
# define SET_FPUCW(word) __extension__ \ |
| 84 |
(void)({ fpucw_t _ncw = (word); \ |
| 85 |
__asm__ __volatile__ ("fldcw %0" : : "m" (*&_ncw)); \ |
| 86 |
}) |
| 87 |
|
| 88 |
# define DECL_LONG_DOUBLE_ROUNDING \ |
| 89 |
fpucw_t oldcw; |
| 90 |
# define BEGIN_LONG_DOUBLE_ROUNDING() \ |
| 91 |
(void)(oldcw = GET_FPUCW (), \ |
| 92 |
SET_FPUCW ((oldcw & ~FPU_PC_MASK) | FPU_PC_EXTENDED)) |
| 93 |
# define END_LONG_DOUBLE_ROUNDING() \ |
| 94 |
SET_FPUCW (oldcw) |
| 95 |
|
| 96 |
#else |
| 97 |
|
| 98 |
typedef unsigned int fpucw_t; |
| 99 |
|
| 100 |
# define FPU_PC_MASK 0 |
| 101 |
# define FPU_PC_DOUBLE 0 |
| 102 |
# define FPU_PC_EXTENDED 0 |
| 103 |
|
| 104 |
# define GET_FPUCW() 0 |
| 105 |
# define SET_FPUCW(word) (void)(word) |
| 106 |
|
| 107 |
# define DECL_LONG_DOUBLE_ROUNDING |
| 108 |
# define BEGIN_LONG_DOUBLE_ROUNDING() |
| 109 |
# define END_LONG_DOUBLE_ROUNDING() |
| 110 |
|
| 111 |
#endif |
| 112 |
|
| 113 |
|
| 114 |
#ifdef __cplusplus |
| 115 |
} |
| 116 |
#endif |
| 117 |
|
| 118 |
#endif /* _FPUCW_H */ |