This is the mail archive of the gcc-patches@gcc.gnu.org mailing list for the GCC project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: Patch: MOVE_RATIO on x86 prescott


On 8/27/05, Dale Johannesen <dalej@apple.com> wrote:
> On Aug 27, 2005, at 1:11 AM, Richard Guenther wrote:
> > On 8/27/05, Dale Johannesen <dalej@apple.com> wrote:
> >> Feedback from users led us to conclude this is set too low.
> >> The test here tries copying arrays of size 9 to 16 words with
> >> mainline (which uses 'rep movsl') and a modified copy with
> >> the patch below (which uses loads and stores); you'll see
> >> loads and stores are faster in all cases, by 2 to 1 or more.
> 
> >> This patch increases MOVE_RATIO to 17 for Prescott, with
> >> the effect that the 'rep; movsl' sequence is never used.
> >> A SPEC run shows +1/2% on int (which is within the margin
> >> of error, really), break-even on FP.  OK?
> >
> > I don't think this is appropriate for stage3
> 
> Fair enough.  It's not a regression or a bug, I guess.
> 
> > - noting that MOVE_RATIO
> > will f.i. affect inlining and sra decisions.
> 
> I know.  The SPEC results convince me it's a good idea overall,
> considering all that.  Do you think additional testing is needed?
> If so, what?

I suppose MOVE_RATIO is not changed for -Os?  In any way,
comparing compile time and code size of SPEC and a CSiBE
run with -O2 would be useful.

Richard.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]