Big and little endian ARM both use 64bit doubles, so use the more general __arm__ rather than __ARMEL__.