[llvm] [AMDGPU][SDAG] Legalise v2i32 or/xor/and instructions to make use of 64-bit wide instructions (PR #140694)

[Matt Arsenault via llvm-commits]

================
@@ -430,6 +430,14 @@ SITargetLowering::SITargetLowering(const TargetMachine &TM,
     setOperationAction(ISD::VECTOR_SHUFFLE, {MVT::v2i32, MVT::v2f32}, Legal);
   }
 
+  setOperationAction({ISD::AND, ISD::OR, ISD::XOR}, MVT::v2i32, Legal);
+  // Prevent SELECT from being implemented with the above bitwise ops and
+  // instead use cndmask.
+  setOperationAction(ISD::SELECT, MVT::v2i32, Custom);
----------------
arsenm wrote:

Probably should fix whatever combiner logic is doing this. In the future we probably should make a similar change for select to better make use of s_cselect_b64 

https://github.com/llvm/llvm-project/pull/140694