[PATCH] D49642: AMDGPU: Rework extract-lowbits test

[Jan Vesely via Phabricator via llvm-commits]

jvesely abandoned this revision.
jvesely added a comment.

In https://reviews.llvm.org/D49642#1172185, @arsenm wrote:

> In https://reviews.llvm.org/D49642#1171179, @jvesely wrote:
>
> > In https://reviews.llvm.org/D49642#1171132, @arsenm wrote:
> >
> > > I'd rather stop trying to share tests with r600 at all. I would like to split out most of the shared tests as-i
> >
> >
> > Any reason for that? Both bfe instructions use the same patterns so it'd be just a copy paste.
>
>
> A lot of tests have too many run lines as is, and adding more for r600 increases the mess. In this case you are actually changing the tested content. The original used VGPR inputs for everything, and this changes everything to be SGPR inputs. Both would be useful as separate tests, but we don't try particular hard to match scalar BFEs currently. Also, I want to stop artificially sharing some of the intrinsics.


Fair enough https://reviews.llvm.org/D49641 has been updated to include a copy of the file with EG/CM checks.
This sounds like chasing fools gold. Most RUN lines are added for new generations of gpu (gfx9, ...) and more for HSA, two lines for cayman and cypress barely make a difference.
The bit extract patterns should be recognized irrespective of whether it can be scalarized or not.


Repository:
  rL LLVM

https://reviews.llvm.org/D49642