[adgjlsfhk1]

it is for finite values, but because IEEE did some dumb things it isn't specified to be for NaN values (and on several architectures, isn't).