It is worth noting that the profiler is fully open-source: https://github.com/plasma-umass/coz. There is even a package for Arch Linux!

Man, I love the idea, even though a lot of the code I want to test is tied to hardware that can't be just slowed down.

They show several examples where speeding up a line is predicted to slow down execution, but ~actually optimizing it had the opposite effect, so kudos to them for including bizarre results too. I hope they take this farther; seems promising.

Very interesting work. Would love to see something like this for the jvm.