Automatically Repairing Programs Using Both Tests and Bug Reports
share
The success of automated program repair (APR) depends significantly on its ability to localize the defects it is repairing. For fault localization (FL), APR tools typically use either spectrum-based (SBFL) techniques that use test executions or information-retrieval-based (IRFL) techniques that use bug reports. These two approaches often complement each other, patching different defects. No existing repair tool uses both SBFL and IRFL. We develop RAFL (Rank-Aggregation-Based Fault Localization), a novel FL approach that combines multiple FL techniques. We also develop Blues, a new IRFL technique that uses bug reports, and an unsupervised approach to localize defects. On a dataset of 818 real-world defects, SBIR (combined SBFL and Blues) consistently localizes more bugs and ranks buggy statements higher than the two underlying techniques. For example, SBIR correctly identifies a buggy statement as the most suspicious for 18.1 extend SimFix, a state-of-the-art APR tool, to use SBIR, SBFL, and Blues. SimFix using SBIR patches 112 out of the 818 defects; 110 when using SBFL, and 55 when using Blues. The 112 patched defects include 55 defects patched exclusively using SBFL, 7 patched exclusively using IRFL, 47 patched using both SBFL and IRFL and 3 new defects. SimFix using Blues significantly outperforms iFixR, the state-of-the-art IRFL-based APR tool. Overall, SimFix using our FL techniques patches ten defects no prior tools could patch. By evaluating on a benchmark of 818 defects, 442 previously unused in APR evaluations, we find that prior evaluations on the overused Defects4J benchmark have led to overly generous findings. Our paper is the first to (1) use combined FL for APR, (2) apply a more rigorous methodology for measuring patch correctness, and (3) evaluate on the new, substantially larger version of Defects4J.
READ FULL TEXT