Use <cmath>'s functions over fdlibm's for performance reasons. No significant precision loss when doing this.