A reader who's known me for a while just sent me a note asking me to say something about premature optimization. The reason that I mention that he's known me for a while (in fact, someone who used to be a coworker) is because this is a topic that I'm prone to rant about in person, but I've never mentioned on the blog. So he knows that it's something that I've got a strong opinion about. Basically, he's having trouble dealing with an annoying coworker doing this, and he wants a rant-by-proxy. I'm OK with that. :-).
When you're writing a new piece of software, particularly on a modern computer, one of the unintuitive things that frequently happens is that the performance of your system is really quite different from what you'd expect.And even when everything is as expected, most people don't have a particularly good sense of tradeoffs - if I make this thing faster, what effect will it have on that? and if it does really improve the performance, how much will it improve it, and at what cost? If you can't answer that - and answer it precisely, with supporting evidence - then you have no business optimizing.
So when you sit down to write some new code, what you should really do is write code that's algorithmically efficient - that is, you pick an algorithm that has good performance in asymptotic time - and implement it in a straightforward way.
But what many people do - in fact, what pretty much all of us do at least some of the time - is try to be clever. We try to find opportunities to change the code to make it faster. Doing that before you understand where the computer actually spends its time when it's running the program is what we call premature optimization.