[llvm] [X86] Lower bf16->f32/f64 fpext without a GPR round-trip (PR #218359)

[Phoebe Wang via llvm-commits]

================
@@ -697,6 +697,8 @@ X86TargetLowering::X86TargetLowering(const X86TargetMachine &TM,
     setOperationAction(ISD::FP_ROUND, MVT::f16, Custom);
     setOperationAction(ISD::FP_EXTEND, MVT::f32, Custom);
     setOperationAction(ISD::FP_EXTEND, MVT::f64, Custom);
+    setOperationAction(ISD::BF16_TO_FP, MVT::f32, Custom);
+    setOperationAction(ISD::BF16_TO_FP, MVT::f64, Custom);
----------------
phoebewang wrote:

Sorry, line 323 is a mistake. The code is in SSE2 block already. I actually mean you should move out the such F16 related code. Maybe line 742.

https://github.com/llvm/llvm-project/pull/218359