We can detect when bytecode is not well formed and fall back on interpretation, which can handle these nasty cases. This is more of a problem for AVM1 than AVM2.