This is the mail archive of the
gcc@gcc.gnu.org
mailing list for the GCC project.
Re: FOUND! PA2.0 float performance problem
- To: Mike Stump <mrs at wrs dot com>
- Subject: Re: FOUND! PA2.0 float performance problem
- From: "Jerry Quinn" <jquinn at nortelnetworks dot com>
- Date: Tue, 23 Feb 1999 17:12:09 -0500
- CC: egcs at egcs dot cygnus dot com
- Organization: Nortel Technology
- References: <199902231916.LAA01800@kankakee.wrs.com>
Mike Stump wrote:
> > 2) represent this type of dependency to egcs?
>
> Change
>
> #define REG_ALLOC_ORDER \
[snip]
> and retest, and let us know how it goes. I suspect the compiler will
> always get better code with the exception of possibly much worse
> prologues and epilogues. Anyway, I suspect this isn't the right fix,
> or at best is only part of it, but for you, it may be better than
> wating for the right fix. At least part of the right fix is to expose
> the conflicts to the scheduler.
This fixes the simple benchmark as expected.
I'm just wondering - would it be possible to fake gcc into believing
that any instruction that writes to a single precision float reg must be
followed by a dummy operation that has that reg as input and its other
half as the output? Then, the dependency between writing to one half
and not being able to use the other half could be fit into the existing
framework.
--
Jerry Quinn Tel: (514) 761-8737
jquinn@nortelnetworks.com Fax: (514) 761-8505
Speech Recognition Research