Subject: Re: Math emulator
To: Andre Skarzynski <abs@itu3.sun.ac.za>
From: J.T. Conklin <jconklin@netcom.com>
List: port-i386
Date: 03/27/1995 11:48:01
> Due to my present bad financial state, I do not have a hardware FPU
> and thus have to rely on the math emulator in the kernel. Unfortantly,
> I find it slow and buggy as compared to the GPL Math emulator found in
> FreeBSD (I was using version 1.1.5.1) before going to NetBSD-1.0A.
Part of the reason almost nothing has been done about the FPU emulator
is that you need: 1) someone who knows a lot about FP math; 2) someone
who hacks well enough to write the emulator code; 3) and someone who
doesn't have a FPU. Rarely are all three attributes found in the same
person.
Recently some of us have exchanged mail about what we could/should do
about the FP emulator. One suggestion was that the software FP code
in sparc/fpu be generalized for 96 and 80 bit FP formats, and the
machine dependent code for the i386 (i387 emulator), m68k (m68881
emulator) and the sparc be modified to call routines provided by
a common core. This would be a lot of work.
This approach would be probably slower than the current i387 emulator.
> Before I have a crack at fixing things or breaking them as is more
> likley the case :-), has someone ever tried improving the the math
> emulator in the intel port of NetBSD? Would looking at the Math
> emulator for MACH (Object code from Intel, no source) be worth a
> try?
Even if the Intel code had no redistribution restrictions, I have
doubts that the core team would want to incorporate object only code
into the Kernel. And this approach doesn't help the m68881 emulator
effort at all.
However, I'm willing to take a look. I looked on the Intel ftp site
and couldn't find anything like this. Could you forward me a copy?
--jtc