[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: ATHLON performance tips

Hello guys,

I had thought of the same thing, so I think I will go with this solution.

My reason is 3DNow! I want to make all load and add instructions 4 bytes
long by using short adresses, and make all mul instructions 8 bytes long
by using long adresses, and then hope for good alignment. It was just an
idea I got, I don't know if it will be effective.



On Wed, 30 May 2001, Julian Ruhe wrote:

> Ups,
> I could kick myself for being so stupid! Of course, Paul's solution
> is an elegant solution! That happens when one writes without thinking...
> Thank you Paul.
> But why, Peter, do you want to expand your instructions artificially?
> Regards
> Julian
> paul.meehan@amd.com wrote:
> > 
> > Peter,
> > 
> > You could bias the base address by subtracting a constant, and then use the
> > offsets to add the constant back in. For example, if 132 is subtracted from
> > the base address, then all the offsets will be bigger than 128. If the base
> > address is known at compile time, then a biased base address could be
> > computed at compile time. However, if the base address is passed in as an
> > argument of a function, then biasing the base address would have to be done
> > at run time, incurring the overhead of a sub instruction.
> > 
> > Regards,
> > Paul