On Fri, Jun 14, 2002 at 12:02:20PM -0400, Brian Hechinger wrote:

> On Wed, Jun 12, 2002 at 03:00:31PM -0400, Charles Shannon Hendrix wrote:
> > 
> > According to the consensus in multiple discussions of this matter,
> > the turbo and micro do respond well to -mcpu=supersparc.  
> 
> which confuses that crap outta me. ;)

It would help if Sun had some official word on their site.  If they do,
I haven't found it yet.

The SS5 is actually a newer machine than the SS20, in which the CPU is
actually called a supersparc.

One person suggested that in some bencharks he did, using ultrasparc
tuning, but not instructions (of course) was beneficial to some non-Ultra
CPUs.  I suppose this is possible, depending on the instruction mix.

> so what would be a good set of test to run to check the difference in speed?
> my sparc5 is kinda busy at the moment, but i may be able to dig up another on
> for testing.  i'd like to be able to slap some conclusive evidence in the
> NetBSD/sparc FAQ if at all possible.

I have a completely free SS5/170.  It's got a NetBSD 1.5.2 build on it.

I suppose a decent benchmark suite would be a way to get started.
Build it with no optimization, the v8 instruction set, the v8 optimized
with supersparc ordering, etc.

openssl benchmarks should be good, but I don't believe that supersparc
helped substantially there, I cannot remember now.  We can certainly
test that theory.

For the most part, the greatest benefit is in using multiply and divide
instructions that gcc doesn't use by default (because the v7 doesn't
have them).




-- 
UNIX/Perl/C/Pizza__________________________________shannon@widomaker.com