On Wed, Mar 01, 2006 at 02:19:47PM -0800, Mark Mitchell wrote:
> H. J. Lu wrote:
>
> > Here are diffs of SPEC CPU 2K between before and after with gcc 4.1
> > using "-O2 -ffast-math" on Nocona:
>
> Steven's right; this is clearly a new feature. HJ's also right; we've
> made exceptions before.
>
> Like Steven, I would like to see what the difference is between
> -mtune=generic and -mtune=nocona. If there is not much difference, then
Here are diffs of "-O2 -mtune=nocona -ffast-math" vs
"-O2 -mtune=generic -ffast-math" on Nocona:
164.gzip 1.30%
175.vpr -0.29%
176.gcc -4.04%
181.mcf 0.74%
186.crafty 1.65%
197.parser -0.10%
252.eon -0.38%
253.perlbmk 5.25%
254.gap 0.93%
255.vortex 1.54%
256.bzip2 0.72%
300.twolf 9.70%
SPECint_base2000 1.31%
168.wupwise 2.49%
171.swim -0.55%
172.mgrid 6.46%
173.applu 0.65%
177.mesa 2.56%
178.galgel 0.14%
179.art 1.79%
183.equake 0.79%
187.facerec -0.07%
188.ammp 4.86%
189.lucas 1.21%
191.fma3d 1.08%
200.sixtrack 0.00%
301.apsi 1.21%
SPECfp_base2000 1.62%
I'd like to see the default -O2 generate decent code for Nocona. It
is independent of improviong -mtune=nocona for Nocona.
> that's an argument against a backport; just tell users to add
> -mtune=nocona. It's up to architecture maintainers to get defaults set
> well for their systems in advance of releases.
>
> I'd also like to hear from our IA32 maintainers.
My IA32 numbers show that -mtune=generic gives the better SPEC CPU 2K
-O2 numbers on Dothan, Yonah, Northwood and Yonah than the previous
-O2 option.
>
> 2. IA32 is one of, if not the, most popular GCC architectures. On the
> one hand, yes, that argues for making it perform well there; on the
> other, it also argues very strongly for stability. Effectively changing
> the default optimization behavior on a release branch substantially
> invalidates previous validation done on that branch.
>
The last thing I want to see is the instability on IA32. I can't
guarantee it won't happen. But I will do my best to fix it if the
backported patch is the cause.
H.J.
