Sorry, but this is just wrong.
I can guarantee you that BigO is very often used not "in theory", but in practice, to find real-world performance problems. And not having "situational awareness" of certain commonly occurring complexity profiles can be a significant source for performance headaches and technical debt.
One trivial-seeming, but frequently occurring example: not knowing when to use hash maps.
In fact, many people solve performance problems (including both those where order-of-magnitude performance really is the most important factor, and various other kinds) not by using a profiler but by, you know, understanding the code and thinking about it. Being as profilers, while they can tell you a lot about certain kinds of performance issues, are still generally quite limited in what they can tell you.
Is it the silver bullet people seem to think it is? No.
Of course not, and I've never heard anyone saying that it was, either.