[PATCH] D117253: [AMDGPU] Select VGPR versions of MFMA if possible

[Stanislav Mekhanoshin via Phabricator via llvm-commits]

rampitec added inline comments.


================
Comment at: llvm/lib/Target/AMDGPU/VOP3PInstructions.td:448
+class AgprMAIFrag<SDPatternOperator Op> :
+  MAIFrag<Op, [{ return MF->getInfo<SIMachineFunctionInfo>()->mayNeedAGPRs(); }]>;
+
----------------
arsenm wrote:
> rampitec wrote:
> > arsenm wrote:
> > > The point of checking it here is you don't need to inspect the original IR. You can check if there are any AGPR uses in the machine function
> > This is selection, there MF is not yet selected. Plus it needs a check for calls.
> We know if there are calls at selection time already, and have already lowered all asm and should have seen the AGPRs already
In fact we don't. As discussed offline we only know it for current bb. This test will fail:

```
declare <32 x float> @llvm.amdgcn.mfma.f32.32x32x1f32(float, float, <32 x float>, i32, i32, i32)

define amdgpu_kernel void @test_mfma_f32_32x32x1f32_call(<32 x float> addrspace(1)* %arg) #0 {
bb1:
  %in.1 = load <32 x float>, <32 x float> addrspace(1)* %arg
  %mai.1 = tail call <32 x float> @llvm.amdgcn.mfma.f32.32x32x1f32(float 1.0, float 2.0, <32 x float> %in.1, i32 1, i32 2, i32 3)
  store <32 x float> %mai.1, <32 x float> addrspace(1)* %arg
  br label %bb2

bb2:
  call void @foo()
  ret void
}

declare void @foo()

attributes #0 = { "amdgpu-flat-work-group-size"="1,256" "amdgpu-waves-per-eu"="2" }
```


CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D117253/new/

https://reviews.llvm.org/D117253