Customized Program Mutation: Inferring Mutant Utility from Program Context

ABSTRACT The purpose of selective mutation strategies is to reduce the inherent redundancy of full mutation analysis and hence obtain most of its bene t for a fraction of the cost. Unfortunately, recent research has shown that there is no xed selective mutation strategy that is e ective across a broad range of programs. In other words, for any given mutation operator, the utility (i.e., usefulness) of a mutant produced by that operator varies greatly across programs. Hence, selective mutation, as currently de ned, is a dead end despite the fact that existing mutation systems are known to produce highly redundant mutants. This paper explores a novel path out of this conundrum. Speci cally, it hypothesizes that mutant utility, in terms of equivalency, triviality, and dominance, can be predicted by incorporating context information from the program in which the mutant is embedded. This paper rst explains the intuition behind this hypothesis with a motivational example and then tests the hypothesis by evaluating the predictive power of a series of program-context models and machine learning classi ers. The results show that it is important to consider mutants at the mutation operator level, and not just at the mutation operator group level. Further, the results show that context extracted from a program’s abstract syntax tree can indeed strongly predict mutant utility. The cross-validation results additionally show that mutant utility can be predicted within and across programs, with an area under the ROC curve above 0.8.