Abstract

When program verification tools fail to verify a program, either the program is buggy or the report is a false alarm. In this situation, the burden is on the user to manually classify the report, but this task is time-consuming, error-prone, and does not utilize facts already proven by the analysis. We present a new technique for assisting users in classifying error reports. Our technique computes small, relevant queries presented to a user that capture exactly the information the analysis is missing to either discharge or validate the error. Our insight is that identifying these missing facts is an instance of the abductive inference problem in logic, and we present a new algorithm for computing the smallest and most general abductions in this setting. We perform the first user study to rigorously evaluate the accuracy and effort involved in manual classification of error reports. Our study demonstrates that our new technique is very useful for improving both the speed and accuracy of error report classification.