@regehr I'd heard of using program search to find computational kernels but I hadn't heard it called "superoptimization" before. kind of makes me want to try a trick I used to do basically genetic programming with bayesian optimization that might make this scale to larger programs. basically you assume that over fixed-length "snippets" the distribution of performance (eg change in runtime with the snippet added vs not) is zipfian and you to thompson sampling but you only store the top few percentile explicitly (in a heap) and you use a count-min sketch for the scores of all the other snippets. I got that to work for generating programs in a toy virtual machine (that's constructed so programs running in it can't crash) but in principle you could do the same thing with code for a real machine