While support for explicit preemption in the llvm IR would use something like this, it can be a long time before such support is added. Before that happens this is dead code and should probably be deleted.
I tend to agree to remove it, since it isn't being used at all yet. But I think some llvm-lto2 tests need to be updated (that's the only way this could get set now). E.g. test/ThinLTO/X86/internalize.ll