Any reason why the overflow / zero division error cannot be handled with a thread-local context? The current behavior could then remain the default for backwards ...