Automatic software fault localization using generic program invariants

Abstract

Despite extensive testing in the development phase, residual defects can be a great threat to dependability in the operational phase. This paper studies the utility of low-cost, generic invariants ("screeners") in their capacity of error detectors within a spectrum-based fault localization (SFL) approach aimed to diagnose program defects in the operational phase. The screeners considered are simple bit-mask and range invariants that screen every load/store and function argument/return program point. Their generic nature allows them to be automatically instrumented without any programmer-effort, while training is straightforward given the test cases available in the development phase. Experiments based on the Siemens program set demonstrate diagnostic performance that is similar to the traditional, development-time application of SFL based on the program pass/fail information known before-hand. This diagnostic performance is currently attained at an average 14% screener execution time overhead, but this overhead can be reduced at limited performance penalty.