SLP/Loop vectorizer pass ordering

[James Molloy]

Hi Hal,

I ran some numbers on Cortex-A57 (AArch64 mode) with the SLP and Loop
vectorizers swapped and with your noalias metadata enabled. I see no change
outside of noise. I ran SPEC(2000+2006) and EEMBC - LNT is currently
unusable on our hardware due to lack of "perf" - it's just too noisy :(

Can someone try a test on x86? If that shows the same, we should be in a
good position to swap the ordering as soon as you enable noalias metadata
by default.

Cheers,

James


On 26 July 2014 02:41, Chad Rosier <mcrosier at codeaurora.org> wrote:

> FWIW, I tend to agree with James and Tobias; the loop vectorizer should
> run first.
>
> >> Also, James, did you happen to run some performance test with the new
> > order?
> >
> > Not yet. I've got one queued to run over the weekend.
>
> James,
> If you send me a patch, I can provide some numbers as well.
>
> > Cheers,
> >
> > James
>
>  Chad
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20140806/d13696ab/attachment.html>