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

[via llvm-commits]

================
@@ -462,6 +462,11 @@ X86TargetLowering::X86TargetLowering(const X86TargetMachine &TM,
     setOperationAction(ISD::BF16_TO_FP, VT, Expand);
----------------
tfzee wrote:

Can you expand what you mean with don't care? Since changing it it will crash compiler on tests like  bfloat-calling-conv-no-sse2.ll which explicitly check for the non SSE2 case.

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