On Tue, Oct 14, 2014 at 07:28:45PM -0400, Vladimir Makarov wrote:
> On 2014-10-14 4:17 PM, Peter Bergner wrote:
> >On Fri, 2014-10-10 at 11:02 -0400, Vladimir Makarov wrote:
> >>Here is a new rematerialization sub-pass of LRA.
> >
> >When Mike and I build with this patch along with the patch that
> >enables LRA by default on powerpc64*-linux (attached below), we're
> >seeing the following error message. I'm not sure how your patch
> >can cause this error, but it does go away if we remove your patch
> >and build again.
>
> Peter, thanks for checking the patch and reporting this. I had
> several wrong code generation problems with rematerialization on x86
> and arm. I've solved them before posting the patch but I did not
> check ppc64.
>
> As a lot of people started to try the patch, several problems were
> reported. I'll address them and do some patch modifications. Now I
> think that I'll commit the patch into the trunk not earlier the next
> week. And I'll check with ppc64 too to be sure that we have no
> wrong code generation problems on this target too.
I started a power7 spec run, and so far I've gotten through SpecINT on trunk,
trunk + LRA, and trunk + LRA + Vlad's patches (including rematerialization).
There are 3 benchmarks that fail with my usual options with the new LRA
patches: gcc, gobmk, h264ref.
My usual options are:
-fpeel-loops -funroll-loops -ftree-vectorize -fvect-cost-model -mlra
-ffast-math -O3 -mcpu=power7
Right now, on the remaining benchmarks, there is a very slight performance loss
for mcf with lra and remat, but MCF is one of those benchmarks that seems to
vary on factors other than compiler options (location of loops in memory,
etc.), but other than that there is no real difference in using the current
reload, the current LRA, and the new patches.
--
Michael Meissner, IBM
IBM, M/S 2506R, 550 King Street, Littleton, MA 01460-6245, USA
email: meiss...@linux.vnet.ibm.com, phone: +1 (978) 899-4797
