[PATCH] D135609: [AMDGPU] Fix True16 patterns for cmp on GFX11

[Joe Nash via Phabricator via llvm-commits]

Joe_Nash marked an inline comment as not done.
Joe_Nash added inline comments.


================
Comment at: llvm/test/CodeGen/AMDGPU/v_cmp_gfx11.ll:16
+entry:
+  %0 = tail call i64 @llvm.amdgcn.icmp.i64.i16(i16 0, i16 0, i32 32)
+  %cmp0 = icmp eq i64 %0, 0
----------------
Joe_Nash wrote:
> arsenm wrote:
> > arsenm wrote:
> > > Should also test llvm.amdgcn.fcmp. I'd assume that's missing coverage too if this was broken 
> > Plus ballot 
> Is ballot well defined for 16 bit destinations? I think ballot should work over the waveSize, so only 32 or 64 bit returns are valid.
> For an instrinsic like this
> 
> ```
> declare i16 @llvm.amdgcn.ballot.i16(i1)
> ```
> I would expect a problem here
> AMDGPUInstructionSelector.cpp:1226 
> AMDGPUInstructionSelector::selectBallot
> 
> ```
>   if (Size != STI.getWavefrontSize())
>     return false;
> ```
On second thought, I think you probably meant a test of a 16 bit compare feeding into ballot. See test https://reviews.llvm.org/D135782


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D135609/new/

https://reviews.llvm.org/D135609