Quantifying the complexity of dataflow testing

It is common belief that dataflow testing criteria are harder to satisfy than statement and branch coverage. As motivations, several researchers indicate the difficulty of finding test suites that exercise many dataflow relations and the increased impact of infeasible program paths on the maximum coverage rates that can be indeed obtained. Yet, although some examples are given in research papers, we lack data on the validity of these hypotheses. This paper presents an experiment with a large sample of object oriented classes and provides solid empirical evidence that dataflow coverage rates are steadily lower than statement and branch coverage rates, and that the uncovered dataflow elements do not generally depend on the feasibility of single statements.