Re: floating point speed

[Ignatios Souvatzis <ignatios%cs.uni-bonn.de@localhost>]

There are two factors hurting performance with NetBSD:

a) the c and math lbrary are compiled without -m68060. You should
recompile them with this flag, too, for maximum performance on the 060
(but this will hurt slower machines, thats why it isn't in the normal
distribution).

b) at the early time the -1.2 release was feature frozen, some wierd
effect made it undesirable to have branch prediction on (gzip became a
lot slower on my first tests, and I didn't find the problem in
time). This hurts normal programs a bit (10 to 40 %), and hurts the
emulation of unimplemented instructions a lot, as it needs lots of
conditional branches. E.g., the 32x32->64bit multiplication is done,
on a 68060/50 machine, at a rate of about 111000 /second with
NetBSD-1.2, and of 150000/second, with NetBSD-current.

c) Even _if_ you recompile libm/libc with -m68060, some assembler
stuff would need to be rewritten (e.g. the "ldexp()" function).

I could well believe that the sum of all of this can be a factor of 2
for some programs.

Regards,
        Ignatios Souvatzis