At the risk of being pedantic, I want to point out that it's not always so clear-cut. Take sorting, for example. Everybody knows that quicksort is quick; it says so right there in the name. It runs in O(n lg n) time with small constant factors. But for small arrays, or arrays which are mostly sorted already, you can do better by using insertion sort, which runs in O(n^2) time. Somewhere in the bowels of glibc is a highly-optimized quicksort which does exactly this: it drops down to insertion sort for subarrays of size 6 or less.
And don't even get me started on the effects of cache locality. The point is, sometimes intuition and analysis of algorithms can only go so far; at some point you've got to either test it or just call it "good enough" and think about something else.
