ACDC: Altering Control Dependence Chains for Automated Patch Generation

TL;DR

ACDC is an automated patch generation method that identifies minimal control dependence chains causing failures, then creates classifiers to selectively negate predicates, effectively repairing faulty Java programs with high accuracy.

Contribution

It introduces a novel approach combining control dependence analysis with predicate classifiers to generate patches that fix failures in Java programs.

Findings

Generated 56 full patches and 46 partial patches.

Achieved an average classifier accuracy of 84%.

Effectively repaired defects in Java benchmarks.

Abstract

Once a failure is observed, the primary concern of the developer is to identify what caused it in order to repair the code that induced the incorrect behavior. Until a permanent repair is afforded, code repair patches are invaluable. The aim of this work is to devise an automated patch generation technique that proceeds as follows: Step1) It identifies a set of failure-causing control dependence chains that are minimal in terms of number and length. Step2) It identifies a set of predicates within the chains along with associated execution instances, such that negating the predicates at the given instances would exhibit correct behavior. Step3) For each candidate predicate, it creates a classifier that dictates when the predicate should be negated to yield correct program behavior. Step4) Prior to each candidate predicate, the faulty program is injected with a call to its corresponding classifier passing it the program state and getting a return value predictively indicating whether to negate the predicate or not. The role of the classifiers is to ensure that: 1) the predicates are not negated during passing runs; and 2) the predicates are negated at the appropriate instances within failing runs. We implemented our patch generation approach for the Java platform and evaluated our toolset using 148 defects from the Introclass and Siemens benchmarks. The toolset identified 56 full patches and another 46 partial patches, and the classification accuracy averaged 84%.