Rethinking optimization for size

I might be wrong, but what we really need is not just a new compiler flag that says "optimize for speed knowing that code size heavily affects speed".

What would be nice to have would be profile-driven optimizations.

You compile your kernel, profile it during a "normal workload" (the definition of which is really tricky), and then, maybe, optimize functions differently.
For instance, for very hot ones you could afford loop unrolling, and for colder ones you'd keep size as small as possible.

My 2c,
_ Massi