For a minor performance decrease you in return get profiling data which allows you to identify areas where you can improve performance. For the most expensive lines of code it's important that you can optimize on production traffic.
This results in a net saving of money over time.
But you are running code all the time, not fixing performance code all the time.
You are also updating the code all of the time. So it makes sense to be profiling all of the new changes.
I imagine the develper may be, but not even then not 'all' the time.
If 1% saves you millions you want to recompile the code anyway to get use of optimizing for your current hardware, not generic distro target. Then you can turn on all flags you want.