Exploring design decisions for mutation testing

Software testing is by far the most popular technique used in industry for quality assurance. One key challenge of software testing is how to evaluate the quality of test suites in terms of their bug-finding capability. A test suite with a large number of tests, or that achieves a high statement or branch coverage, does not necessarily have a high bug-finding capability. Mutation testing is widely used in research to evaluate the quality of test suites, and it is often considered the most powerful approach for this purpose. Mutation testing proceeds in two steps. The first step is mutant generation. A mutant is a modified version of the original program obtained by applying a mutation operator. A mutation operator is a program transformation that introduces a small syntactic change to the original program. The second step of mutation testing is to run the test suite and determine which mutants are killed, i.e., which mutants lead to tests having a different output when run on them compared against running on the original program. Mutation testing produces a measure of quality of the test suite called mutation score. The mutation score of a given test suite is the percentage of mutants killed by that test suite out of the total number of generated mutants. In this dissertation, we explore three design decisions related to mutation testing and provide recommendations to researchers in those regards. First, we look into mutation operators. To provide insights about how to improve the test suites, mutation testing requires both high quality and diverse mutation operators that lead to different program behaviors. We propose the use of approximate transformations as mutation operators. Approximate transformations were introduced in the emerging area of approximate computing for changing program semantics to trade the accuracy of results for improved energy efficiency or performance. We compared three approximate transformations with a set of conventional mutation operators from the literature, on nine open-source Java subjects. The results showed that approximate transformations change program behavior differently from conventional mutation operators. Our analysis uncovered code patterns in which approximate mutants survived (i.e., were not killed) and showed the practical value of approximate transformations both for understanding code amenable to approximations and for discovering bad tests. We submitted 11 pull requests to fix bad tests. Seven have already been integrated by the developers. Second, we explore the effect of compiler optimizations on mutation testing. Multiple mutation testing tools were developed that perform mutation at different levels. More recently mutation testing has been performed at the level of compiler intermediate representation (IR), e.g., for the LLVM IR and Java bytecode/IR. Compiler optimizations are automatic program transformations applied at the IR level with the goal of improving a measure of program performance, while preserving program semantics. Applying mutations at the IR level means that mutation testing becomes more susceptible to the effects of compiler optimizations. We investigate a new perspective on mutation testing: evaluating how standard compiler optimizations affect the cost and results of mutation testing performed at the IR level. Our study targets LLVM, a popular compiler infrastructure that supports multiple source and target languages. Our evaluation on 16 Coreutils programs discovers several interesting relations between the numbers of mutants (including the numbers on equivalent and duplicated mutants) and mutation scores on unoptimized and optimized programs. Third, we perform an empirical study to compare mutation testing at the source (SRC) and IR levels. Applying mutation at different levels offers different advantages and disadvantages, and the relation between mutants at the different levels is not clear. In our study, we compare mutation testing at the SRC and IR levels, specifically in the C programming language and the LLVM compiler IR. To make the comparison fair, we develop two mutation tools that implement conceptually the same operators at both levels. We also employ automated techniques to account for equivalent and duplicated mutants, and to determine hard-tokill mutants. We carry out our study on 16 programs from the Coreutils library, using a total of 948 tests. Our results show interesting characteristics that can help researchers better understand the relationship between mutation testing at both levels. Overall, we find mutation testing to be better at the SRC level than at the IR level: the SRC level produces much fewer (non-equivalent) mutants and is thus less expensive, but the SRC level still generates a similar number of hard-to-kill mutants