New floating-point conversion routines