Subject: Re: accuracy of "long double"
To: None <M.Drochner@fz-juelich.de>
From: Vincent <10.50@free.fr>
List: port-amd64
Date: 08/31/2007 08:59:29
Also Matthias Drochner schrieb:

> Technically, the "long double" type is not worth much on i386,
> at least with newer processors which speed up "double" arithmetics
> by SSE2+.
> There is a need however to keep the "long double" type as-is:
> the ELF ABI spec.

Well, I'm quite unhappy with this. I suspect, though I can obviously not 
prove it by simple intuition, that some math tools I use at least 
occasionally (Octave, Maxima, some other simulation tools) rely on <long 
double> != <double>, at least partly because they have been written to 
be run on Linux. Beyond this, on the piece of code I write, I was using 
<long double> to absorb roundoff error in cumulative sums (like, e.g., 
to compute exact <double> roots with a Newton-like algorithm).

So, I'll try to put the x86 coprocessor in "Linux mode" by default. If I 
understand my NetBSD correctly, this has to be done somewhere in 
arch/i386/isa/npx.c, but can someone tell me exactly in what function 
this shall be done?

Thanks
Vincent