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

[Changpeng Fang via llvm-commits]

================
@@ -915,7 +915,7 @@ SITargetLowering::SITargetLowering(const TargetMachine &TM,
     setOperationAction(ISD::BUILD_VECTOR, MVT::v2bf16, Legal);
   }
 
-  if (Subtarget->hasCvtPkF16F32Inst()) {
+  if (Subtarget->hasCvtPkF16F32Inst() && TM.Options.UnsafeFPMath) {
     setOperationAction(ISD::FP_ROUND, MVT::v2f16, Legal);
----------------
changpeng wrote:

I am still learning. I think even though we specify Custom lowering but actually doing nothing, like:
if (SrcVT. != MVT::f64)
    return Op;  // **f32 is legal for fp_round**
This is essentially "Legal", which means either there is a native instruction, or a pattern to match to native instructions.  The compiler will throw an error if there is nothing to select.

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