[patch][graphite] Fix code generated by Graphite for vectorization

Sebastian Pop sebpop@gmail.com
Wed Mar 17 15:27:00 GMT 2010


On Wed, Mar 17, 2010 at 10:08, Richard Guenther <rguenther@suse.de> wrote:
> On Wed, 17 Mar 2010, Sebastian Pop wrote:
>
>> On Wed, Mar 17, 2010 at 05:05, Richard Guenther <rguenther@suse.de> wrote:
>> > On Tue, 16 Mar 2010, Sebastian Pop wrote:
>> >
>> >> Hi,
>> >>
>> >> The two patches 0002 and 0003 are needed to improve the vectorization
>> >> of the code generated by Graphite.  0001 is a first step towards
>> >> pretty printing the SCoPs under the scoplib format.  I bootstrapped
>> >> and tested the attached patches in the Graphite branch, Java disabled,
>> >> as the Java libs are not building for now.  Ok for trunk?
>> >
>> > Are you sure it works if the induction variable is used after the
>> > loop?  Also I think this is stage1 material.
>>
>>   create_iv (build_int_cst_type (type, 0), build_int_cst (type, 1), NULL_TREE,
>>            loop, &gsi, bump_in_latch, &var_before, NULL);
>>
>> Only the var_before is taken from the IV creation, so the ssa_name
>> after bump is not used.
>
> Ok.
>
>> Also note that the only change of this patch is for Graphite, the
>> autopar usage of canonicalize_loop_ivs remains unchanged: autopar
>> relies on the bump to be in the latch to be able to eliminate it.
>
> Yes, I've seen that.  0002 and 0003 are ok for trunk then, please
> leave 0001 for 4.6.

Ok, I will commit 0001 to the Graphite branch and will leave it for 4.6.

I will first commit 0002 and 0003 to the Graphite branch for further
testing.  If everything is alright with the SPEC, I will commit to trunk.

Sebastian



More information about the Gcc-patches mailing list