While that paper makes a generally true statement, trying to do the most profitable optimizations in the most profitable way is what dead-ended gcc for many many years :)

At some point, you have to step back from what you are doing and actually design and engineer optimizations coherently, not just "do this little thing and that little thing"