[PATCH] D15552: TTI: Allow targets to override getScalarizationOverhead

Matt Arsenault via llvm-commits llvm-commits at lists.llvm.org
Wed Mar 23 11:22:53 PDT 2016


arsenm abandoned this revision.
arsenm added a comment.

In http://reviews.llvm.org/D15552#381685, @hfinkel wrote:

> In http://reviews.llvm.org/D15552#379215, @arsenm wrote:
>
> > In http://reviews.llvm.org/D15552#375044, @hfinkel wrote:
> >
> > > If the vectors don't exist, shouldn't the insert/extract costs be zero?
> >
> >
> > The extract cost is zero. I'm not sure a vector insert should be considered as 0, because the only time this would really happen in the ISA is if packing a vector for a store
>
>
> I don't understand what this means. What are you packing into? Is this packing in any way associated with the vector insertion IR instruction? But if you don't actually have vectors, then why is this cost not just part of the store cost?


That would make more sense. I'll go with the other plan of making insertelement free


http://reviews.llvm.org/D15552





More information about the llvm-commits mailing list