The focus on superoptimization is laughable, and puts me off on even considering the other opinions presented here. Superoptimization is such a joke in so many cases that it's basically undeployable: if it takes a days of a srever's compute to save a microsecond, you better be planning to run that program 8.64e+10 times...