Automated program repair with canonical constraints

Abstract

Automated program repair (APR) seeks to improve the speed and decrease the cost of repairing software bugs. Existing APR approaches use unit tests or constraint solving to find and validate program patches. We propose Canonical Search And Repair (CSAR), a program repair technique based on semantic search which uses a canonical form of the path conditions to characterize buggy and patch code and allows for easy storage and retrieval of software patches, without the need for expensive constraint solving. CSAR uses string metrics over the canonical forms to cheaply measure semantic distance between patches and buggy code and uses a classifier to identify situations in which test suite executions are unnecessary-and to provide a finer-grained means of differentiating between potential patches. We evaluate CSAR on the IntroClass benchmark, and show that CSAR finds more correct patches (96% increase) than previous semantic search approaches, and more correct patches (34% increase) than other previous state-of-the-art in program repair.