On Wed, 19 Feb 2003, gabriel rosenkoetter wrote:

> On Wed, Feb 19, 2003 at 10:11:25PM +1100, Terence Tan wrote:
> > I'm also using Bruce ONeel's distribution. It's basically a standard
> > NetBSD distribution except the userland is compiled as -msoft-float (to
> > bypass the fpu). No simulation needed.
> >
> > I'm working off a stock Performa 475, and everything is working
> > swimmingly well. Obviously everything has to be re-built with
> > - -msoft-float, but I've set up pkgsrc... got apache, perl, postfix, and
> > other good things compiled and working. Nary a segmentation fault in
> > sight.

> How's the performance relative to an 030 using NetBSD's fpu
> emulation? (My guess would be pretty good, provided gcc can optimize
> m68k decently these days; I can dig out an LC III to do some
> benchmarking on the 030 side, if no one else has one handy.)

It's probably much better, because you avoid the overhead of the traps.

> Is there a reliable way we can detect an LC040? Do we really just
> want to do this instead of bothering to emulate fpu all the time for
> mac68k in an fpu's absence, or are there negative consequences (say,
> for math applications)?

locore *does* check the processor capabilites, so FPU_EMULATE is just
harmless dead weight on fpu'd machines.

People have reported that GENERIC kernels don't work on fpu-less
machines. I think the multi-processor case is fixed, but there's still
an issue with FPSP vs FPU_EMULATE (I suspect). However, no one who's
reported problems has ever stayed interested long enough to say there
is or there isn't, understandably -- it requires enormous patience to
get anything done with a machine that slow.

In any case, I think we should stop imposing --soft-float on everyone
(-m68020-40 should be good enough for all fpu'd machines), and have
the release mechanism build a 'NOFPU' kernel with --soft-float and
without M68040.

Frederick