I've been sitting in a lot of code reviews and code inspections where somebody suggests a change to optimize something. The response from the author is almost always; but that's premature optimization! End of argument... But I almost never hear somebody doing the opposite. I came to think about this again a few days ago when it happened again and just a few hours later I read this post about the subject.
Premature optimization is not black or white. Some things we know are bad for performance (like memory allocation and copies in a high performance C program) and nobody questions if you optimize that without actual proof it's a problem. As the post above describes it's all about paying a tax up front that is big enough to avoid a large fine later. I don't think the idea behind the quote about premature optimization is about making no optimizations. It's about making a reasonable amount of optimizations. Kind of the same kind of trade-off you have to do when it comes to writing tests. Is it worth adding a test for code that you know will pass just because to document the expected behavior? Sometimes yes, but not always as this post points out.
There are a lot of optimization kinds in an application, and each kind has a favorite place.
- Conception, for example is the good place to start optimizing the processes (yes, that's an optimization, just a "business process optimization").
- In daily development, optimization is oriented towards the developped modules (it's what I would call "local optimizations").
- In integration test phases, optimization will focus on the most time consuming activities.
With today technologies, optimizations most often relate to algorithms, thus they can be integrated painlessly during all development cycle.
And "Continuous optimization" is also closer to agile development (exactly as "release often, release soon", we could say "optimize often, optimize soon").
So yes, 10 years ago, premature optimization was a bad thing (it implied people rewriting code parts in assembly language instead of C), but we must stop listening to dinosaurs :)
Maybe you should read this one: