Avoiding the Familiar to Speed Up Test Case Reduction

Abstract

Delta Debugging is a longstanding approach to automated test case reduction. It divides an input into chunks and attempts to remove them to produce a smaller input. When a chunk is successfully removed, all chunks are revisited, as they may become removable from the smaller input. When no chunk can be removed, the chunks are subdivided and the process continues recursively. In the worst case, this revisiting behavior has an O(n^2) running time. We explore the possibility that good test case reduction can be achieved without revisiting, yielding an O(n) algorithm. We identify three independent conditions that can make this reasonable in practice and validate the hypothesis on a suite of user-reported and fuzzer-generated test cases. Results show that on a suite of large fuzzer-generated test cases for compilers, our O(n) approach yields reduced test cases with similar size, while decreasing the reduction time by 65% on average.