Automatic Generation of Peephole Superoptimizers

Abstract Peephole optimizers are typically constructed using human-writtenpattern matching rules, an approach that requires expertise and time, as well as being less than systematic at exploiting all oppor-tunities for optimization. We explore fully automatic construction of peephole optimizers using brute force superoptimization. Whilethe optimizations discovered by our automatic system may be less general than human-written counterparts, our approach has the po-tential to automatically learn a database of thousands to millions of optimizations, in contrast to the hundreds found in current peep-hole optimizers. We show experimentally that our optimizer is able to exploit performance opportunities not found by existing com-pilers; in particular, we show speedups from 1.7 to a factor of 10 on some compute intensive kernels over a conventional optimizingcompiler