[PATCH] [AArch64] Improve and enable the SeparateConstOffsetFromGEP for AArch64 backend.

[Jingyue Wu]

SPEC2006 benchmarks are single-threaded, but this single thread can be
scheduled by the operating system on different CPUs (cores, hyperthreads,
etc.) in different time slices. I used to encounter fluctuation too and
fixed that using the taskset command. Not sure if that can help you too.

Also, I don't think x86 matters because this pass is only enabled in
AArch64 and NVPTX for now. I do feel a little worried on the 1%-2%
regression you observed on AArch64. It would be great if you can get a more
stable result.

Jingyue

On Tue Nov 18 2014 at 11:01:24 PM Hao Liu <Hao.Liu at arm.com> wrote:

> >>! In D5864#58, @jingyue wrote:
> > Did you try pinning the benchmark on one CPU? Scheduling can
> significantly
> > disturb the benchmark results. Also, if your CPU has any boosting feature
> > (similar to Intel's turbo boost), try disabling that. One of our folks,
> > Mark, found that can make benchmarking sensitive too.
> >
> > Jingyue
>
>
>
> >>! In D5864#58, @jingyue wrote:
> > Did you try pinning the benchmark on one CPU? Scheduling can
> significantly
> > disturb the benchmark results. Also, if your CPU has any boosting feature
> > (similar to Intel's turbo boost), try disabling that. One of our folks,
> > Mark, found that can make benchmarking sensitive too.
> >
> > Jingyue
>
> I think SPEC2006 benchmarks always run on one CPU, so I didn't pin to one
> CPU. For the boosting feature, I tested on an old x86 machine, so I'm not
> sure about the features. Anyway, I'm not familiar with x86 performance
> tests, so I think it's better to let other experts to have a try.
>
> Thanks,
> -Hao
>
> http://reviews.llvm.org/D5864
>
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20141119/3729ab86/attachment.html>