Re: Submission of pa8000 changes?

["Jerry Quinn" <Jerry dot Quinn dot jquinn at nt dot com>]

Jeffrey A Law wrote:
> 
> Actually, I doubt this is going to help scheduling anyway.
> 
> Most of the papers I've read on the PA8000 discourage latency scheduling for
> everything except a few FP ops where they encourage an inaccurate description
> of latencies.
> 
> Did you benchmark after your changes?  If so, what did you benchmark and what
> were the results of those benchmarks.

Yeah.  I got no improvement at all.  I was more doing it to provide a
hook for those with more knowledge to work on it.  What I typically find
is that double performance is acceptable but single float is
significantly worse.

> I think you'll find that altering some of the code generation choices wil give
> you much better results.  For example, on the PA8000 it is wise to avoid insns
> which have multiple outputs (like fmpyadd and fmpysub).  I also suspect that
> implementing fused multiply-add type sequences will be a big win.  My sources
> indicate that traditional latency scheduling is almost worthless on this chip.

How do you go about making gcc avoid the fmpyadd and fmpysub
instructions for just the 8000?  If we wanted to include the fused
multiply-add instructions, it appears that all the output would have to
be converted to pa2.0 opcodes, which includes revising gas.  Does this
sound about right?
 
-- 
Jerry Quinn                             Tel: (514) 761-8737
jquinn@nortel.ca                        Fax: (514) 761-8505
Speech Recognition Research