I suspect that combining generics with subtyping might always lead you into the land on undecidable type-inference. After all, pure genericity (System F) is already undecidable.