[llvm] AMDGPU: Make v2f64 -> v2f16 conversion Legal only when unsafe fast math is set (PR #134738)

[Changpeng Fang via llvm-commits]

================
@@ -6893,6 +6893,19 @@ SDValue SITargetLowering::lowerFP_ROUND(SDValue Op, SelectionDAG &DAG) const {
     return Op;
 
   EVT DstVT = Op.getValueType();
+
+  if (DstVT == MVT::v2f16) {
----------------
changpeng wrote:

> Same comments as the globalisel path. The vectoriness shouldn't matter, we don't need to directly consume the operation

Okay, we can do as https://github.com/llvm/llvm-project/pull/135659, which essentially scalarize v2f16.  WE may do instcombine to geenrate the vector cvt when the target support. Thanks.  

https://github.com/llvm/llvm-project/pull/134738