[llvm] AMDGPU: Fix incorrectly selecting fp8/bf8 conversion intrinsics (PR #107291)

[Stanislav Mekhanoshin via llvm-commits]

================
@@ -599,8 +599,8 @@ class Cvt_F32_F8_Pat<SDPatternOperator node, int index,
     (inst_sdwa 0, $src, 0, 0, index)
 >;
 
-let SubtargetPredicate = isGFX9Only in {
-let OtherPredicates = [HasCvtFP8VOP1Bug] in {
+let SubtargetPredicate = HasFP8ConversionInsts in {
----------------
rampitec wrote:

> But is the bug really present on gfx12?

No, the bug is that instruction requires SWDA even for the first byte. GFX12 does not have SDWA and hence does not have the bug.

https://github.com/llvm/llvm-project/pull/107291