This seems like an ad-hoc collection of hand-written hacks. What I've wondered about is the feasibility of a public domain collection of superoptimized (
http://en.wikipedia.org/wiki/Superoptimization) code sequences for particular tasks for particular architectures. The repository itself would need to have the actual machines running to do timing, but the code sequence search can be done in a distributed fashion like folding@home (on the assumption that shorter sequences will generally result in shorter timings).