64-bit floating-point arithmetic, so conversion routines for these
formats cannot be generated by the macros
macro-make-representation->double and macro-make-double->representation
in the sample implementation.
Well, it's good to have the prospect of a standardized format, and it's understandable that killing 14 NaNs (8 8 mantissa values * 2 signs - 2 values for infinities) doesn't seem like the Right Thing when you only have 256 to play with. Of course when the mantissa is bigger the wastage is even worse.