Re: [PATCH] Fix PR45948: add ssa defs from builtin partitions to the last partition.

[Richard Guenther <richard dot guenther at gmail dot com>]

On Mon, Dec 13, 2010 at 6:05 PM, Sebastian Pop <sebpop@gmail.com> wrote:
> On Mon, Dec 13, 2010 at 10:46, Richard Guenther
> <richard.guenther@gmail.com> wrote:
>> I think the idea is ok, but as loop distribution already has to track
>> use-def scalar dependences can't we simply add all loop-closed
>> PHI nodes (args?) to the last partition? ?Because it is those that
>> we need to preserve (I think). ?Your patch might effectively do
>> that, but the extra code looks odd to me, as the loop distribution
>> machinery would already need to know most of this, no?
>
> Note that this patch does use the same code as the loop distribution
> machinery: rdg_flag_vertex_and_dependent
>
> Adding all the close phi nodes to the last partition would duplicate a
> lot of the scalar computations that are already computed in other
> partitions: for example,
>
> for (i = 0, j = 0; i < n; i++, j+=2)
> {
> ?A[i] = 0;
> ?B[j] = 0;
> }
>
> use (i, j);
>
> in this code we would have a close phi node for both i and j, and it
> is possible to distribute the memset zero on array A so the
> computation of i should be duplicated on the last partition, but the
> loop on j will still remain as B[j] cannot be generated with a memset
> zero, and so if you add the close phi of j to the last partition you
> would duplicate the scalar computation of j.
>
> So we have to detect all the partitions to find out which scalar
> computations were not code generated in these partitions.

Hm, that's true.  Your patch is ok then, as is.

Thanks,
Richard.

> Sebastian
>