[llvm] [AMDGPU] Recognise bitmask operations as srcmods (PR #149110)

[via llvm-commits]

================
@@ -3036,6 +3036,36 @@ bool AMDGPUDAGToDAGISel::SelectVOP3ModsImpl(SDValue In, SDValue &Src,
     Src = Src.getOperand(0);
   }
 
+  // Convert various sign-bit masks to src mods. Currently disabled for 16-bit
----------------
LU-JOHN wrote:

Do these changes need a guard to ensure that Src is a floating-point type?

https://github.com/llvm/llvm-project/pull/149110