Why can't you just leave it as-is?
I mean is 1% improvement (for cpu intensive workload) really worth
Instead of messing arround with stuff like that, I guess a lot of code
would benefit of beeing build with profile driven optimizations, which
often yields a 5-15% improvement without sacrifycing anything.
On amd64 it would even enable the auto-vectorizer (if enabled) to
vectorize only parts which count, without bloating code unescessary.
However that would be _real_ work, instead of just changing switches
and discussing it forth and back ;)