[llvm] [AMDGPU] Add custom lowering of llvm.convert.from.arbitrary.fp for FP8 (PR #194144)

[Matt Arsenault via llvm-commits]

================
@@ -985,6 +986,18 @@ SITargetLowering::SITargetLowering(const TargetMachine &TM,
     setOperationAction({ISD::FEXP2, ISD::FLOG2, ISD::FSQRT}, MVT::bf16, Legal);
   }
 
+  if (Subtarget->hasOCPFP8ConversionInsts()) {
+    setOperationAction(ISD::CONVERT_FROM_ARBITRARY_FP,
+                       {MVT::f32, MVT::v2f32, MVT::v4f32}, Custom);
+    setOperationAction(ISD::CONVERT_FROM_ARBITRARY_FP, {MVT::v2i8, MVT::v4i8},
----------------
arsenm wrote:

Splitting the v4 cases should probably be handled by the generic legalizer 

https://github.com/llvm/llvm-project/pull/194144