[llvm] [X86] Fix definition of VCVTNE2PS2BF16, make it SchedWriteCvtPS2BF (PR #177792)

[Simon Pilgrim via llvm-commits]

================
@@ -646,6 +646,10 @@ def SchedWriteCvtPD2PS
  : X86SchedWriteWidths<WriteCvtSD2SS, WriteCvtPD2PS,
                        WriteCvtPD2PSY, WriteCvtPD2PSZ>;
 
+def SchedWriteCvtPS2BF
+ : X86SchedWriteWidths<WriteCvtSS2I, WriteCvtPS2I,
+                       WriteCvtPS2IY, WriteCvtPS2IZ>;
----------------
RKSimon wrote:

Getting SchedWriteCvtPS2BF in place is a great first step, but BF16 instructions are pretty slow on all targets that support them at the moment so I'd recommend keeping the slower PD2PS classes for now to make this NFC:
```
def SchedWriteCvtPS2BF
 : X86SchedWriteWidths<WriteCvtSD2SS, WriteCvtPD2PS,
                       WriteCvtPD2PSY, WriteCvtPD2PSZ>;
```
A later patch would investigate what options we have for better classes - then general aim is to reduce the number of InstRW overrides the scheduler models have to maintain.

https://github.com/llvm/llvm-project/pull/177792