[PATCH] D64497: [AMDGPU] Allow abs/neg source modifiers on v_cndmask_b32

[Matt Arsenault via Phabricator via llvm-commits]

arsenm added a comment.

In D64497#1578483 <https://reviews.llvm.org/D64497#1578483>, @arsenm wrote:

> In D64497#1578475 <https://reviews.llvm.org/D64497#1578475>, @foad wrote:
>
> > In D64497#1578444 <https://reviews.llvm.org/D64497#1578444>, @arsenm wrote:
> >
> > > We still need more work to handle the integer cases in the complex pattern.
> >
> >
> > Which pattern?
>
>
> VOP3Mods only looks for FNEG and FABS, although the equivalent bit operations should also work
>
> > 
> > 
> >> Plus the combine that tries to avoid source modifiers on select should be updated
> > 
> > Where is that? I looked but couldn't find anything like this.
>
> Look at performFNegCombine, performFAbsCombine, fnegFoldsIntoOp and hasSourceMods


You might also want to look at picking up https://github.com/arsenm/llvm/commit/77ac293d3510093e6db3f2825c8bef859bac4dcf
IIRC this had some improvements in shader-db, but also some regressions. It might be better with select using mods


Repository:
  rL LLVM

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D64497/new/

https://reviews.llvm.org/D64497