Several years ago Next (or someone at Next) implemented a compiler pass that sets the x86 coprocessor precision to agree with the user program declarations. If you could track that down it might be the basis of a useful improvement.