Mining Control Flow Abnormality for Logic Error Isolation

Analyzing the executions of a buggy program is essentially a data mining process: Tracing the data generated during program executions may disclose important patterns and outliers that could eventually reveal the location of software errors. In this paper, we investigate program logic errors, which rarely incur memory access violations but generate incorrect outputs. We show that through mining program control flow abnormality, we could isolate many logic errors without knowing the program semantics. In order to detect the control abnormality, we propose a hypothesis testing-like approach that statistically contrasts the evaluation probability of condition statements between correct and incorrect executions. Based on this contrast, we develop two algorithms that effectively rank functions with respect to their likelihood of containing the hidden error. We evaluated these two algorithms on a set of standard test programs, and the result clearly indicates their effectiveness.