Evolving Faster Haskell Programs (now with LLVM)

How is this relevant?

From the article:

"Computing fitness

The programs we want to improve have to report their fitness. To do this I’ll use criterion’s timing function, wrapping up the code we actually want to optimize in some measurement functions. This will just report total wall clock running time for the code under analysis."

They only care about speed. And if code size turns out to have an impact on speed, it won't be ignored by this metric.

The article does ignore it, and doesn't claim to do otherwise. However, it would be trivial to modify their algorithms to improve code size.

Just solve

    argmin running_time(prog) + alpha * code_size(prog)

instead of

    argmin running_time(prog)

Here, alpha is a parameter representing how much you care about code size. I.e., if alpha = 1 sec / 1mb, then you will accept code which is 0.99 seconds slower if it shaves 1mb off the code size, but not code which is 1.01 seconds slower.