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

[Matt Arsenault via llvm-commits]

================
@@ -765,14 +780,15 @@ def V_CVT_F16_F8_True16_Profile : VOP3_Profile_True16<V_CVT_F16_F8_Profile>;
 def V_CVT_F16_F8_Fake16_Profile : VOP3_Profile_Fake16<V_CVT_F16_F8_Profile>;
 }
 
-let SubtargetPredicate = isGFX11Plus, OtherPredicates = [HasFP8ConversionInsts],
+let SubtargetPredicate = isGFX11Plus,
+    OtherPredicates = [HasFP8ConversionInsts, HasOCPFP8ConversionInsts],
     mayRaiseFPException = 0, SchedRW = [WriteFloatCvt] in {
   let SubtargetPredicate = isGFX11PlusNot12_50 in
-    defm V_CVT_F32_FP8_OP_SEL    : VOP1Inst<"v_cvt_f32_fp8_op_sel", VOPProfile_Base_CVT_F_F8_ByteSel<f32>>;
+    defm V_CVT_F32_FP8_OP_SEL    : VOP1Inst<"v_cvt_f32_fp8_op_sel", VOPProfile_Base_CVT_F_F8_ByteSel<f32>, cvt_from_fp8>;
   let SubtargetPredicate = isGFX125xOnly in
-    defm V_CVT_F32_FP8_gfx1250   : VOP1Inst<"v_cvt_f32_fp8_gfx1250", VOPProfile_Base_CVT_F_F8_ByteSel<f32, 1>>;
+    defm V_CVT_F32_FP8_gfx1250   : VOP1Inst<"v_cvt_f32_fp8_gfx1250", VOPProfile_Base_CVT_F_F8_ByteSel<f32, 1>, cvt_from_fp8>;
 
-  defm V_CVT_F32_BF8_OP_SEL    : VOP1Inst<"v_cvt_f32_bf8_op_sel", VOPProfile_Base_CVT_F_F8_ByteSel<f32>>;
+  defm V_CVT_F32_BF8_OP_SEL    : VOP1Inst<"v_cvt_f32_bf8_op_sel", VOPProfile_Base_CVT_F_F8_ByteSel<f32>, cvt_from_bf8>;
----------------
arsenm wrote:

The _OP_SEL should be lowercased 

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