See #335
Our float handling generally doesn't work without IEEE 754 and I am not aware of any real architecture that doesn't support it => consider restricting support to such platforms.
Main practical implication would be that we could bring back signalling NaNs with only minor workarounds for legacy systems that use the on bit