Deep imperative mutations have less impact

Abstract Information theory and entropy loss predict deeper more hierarchical software will be more robust. Suggesting silent errors and equivalent mutations will be more common in deeper code, highly structured code will be hard to test, so explaining best practise preference for unit testing of small methods rather than system wide analysis. Using the genetic improvement (GI) tool MAGPIE , we measure the impact of source code mutations and how this varies with execution depth in two diverse multi-level nested software. gem5 is a million line single threaded state-of-the-art C++ discrete time VLSI circuit simulator, whilst PARSEC VIPS is a non-deterministic parallel computing multi-threaded image processing benchmark written in C. More than 28–53% of mutants compile and generate identical results to the original program. We observe 12% and 16% Failed Disruption Propagation (FDP). Excluding internal errors, exceptions and asserts, here most faults below about 30 nested function levels which are E xecuted and I nfect data or divert control flow are not P ropagated to the output, i.e. these deep PIE changes have no visible external effect. Suggesting automatic software engineering on highly structured code will be hard.