https://github.com/arsenm commented: How does using isfpclass avoid scalarization here? I think it's somewhat preferably to use the named operations here, they are subtly different since they canonicalize the input unlike is.fpclass https://github.com/llvm/llvm-project/pull/124145