Title:Precondition Inference for Peephole Optimizations in LLVM

Abstract:Peephole optimizations are a common source of compiler bugs. Compiler developers typically transform an incorrect peephole optimization into a valid one by strengthening the precondition. This process is challenging and tedious. This paper proposes PInfer, a data-driven approach for inferring preconditions for peephole optimizations expressed in Alive. PInfer generates positive and negatives examples for an optimization, enumerates predicates on-demand, and learns a set of predicates that separate the positive and negative examples. PInfer repeats this process until it finds a precondition that ensures the validity of the optimization. PInfer reports both the weakest precondition and a set of succinct partial preconditions to the developer. The PInfer prototype successfully generates either the partial precondition or the weakest precondition for 164 out of 174 peephole optimizations in the Alive suite. It also generates preconditions that are weaker than LLVM's precondition for 73 optimizations. We also demonstrate the applicability of this technique to generalize 54 concrete expression directed acyclic graphs generated by an LLVM IR-based super optimizer.