Harnessing heap analysis for the synthesis of superoptimized bytecode

Abstract

Superoptimization is a type of program synthesis technique that, given an original loop-free sequence of instructions, synthesizes an alternative semantically-equivalent sequence that is optimal wrt the considered objective function. Working on loop-free sequences restricts the kind of achievable optimizations to the local scope of the considered sequences. This article harnesses a global heap analysis for the synthesis of superoptimized loop-free sequences of bytecode. The global heap analysis will allow us to infer useless write heap accesses, aliasing and non-aliasing properties, and calling-contexts for the sequences. Exploiting this information on an existing superoptimizer for Ethereum bytecode has required novel extensions: (1) developing a finer-grained heap analysis able to infer heap properties that can boost superoptimization, (2) adapting several components of the superoptimizer to leverage the heap properties, and (3) extending the superoptimization algorithm to work in a context-sensitive way. Our experimental results on more than 200,000 sequences show that harnessing heap analysis for superoptimization not only improves the quality of the optimization but it can even reduce the optimization time.