Souper, a LLVM-based superoptimization framework, has seen adoption in both academic research and industry projects. Given an LLVM IR as an input, Souper tries to generate peephole patterns by synthesizing semantically equivalent but shorter instruction sequences. However, as a platform-independent framework, it lacks a model of the actual cost of an instruction sequence on the target machine. This leads to missing optimizations opportunities or generation of peephole patterns which degrade performance. In this talk, we’re going to demonstrate how Souper can benefit from target machine information. Then, we will explore some possible approaches to providing Souper with target machine info to steer the superoptimizer to find more patterns with improvements than regressions. This will enable Souper to be used in a more automated way and reduce the manual intervention required.
