Learning Strategies Using Boolean Program Metrics to Verify Industrial Code

Verification tools and techniques are known to possess strengths and weaknesses with respect to different program syntax and semantics. Thus in practice, a sequence of verification techniques is often custom built to verify a class of similar programs. Such a sequence of techniques is called a strategy. So far, verification strategies have been created manually or through machine learning methods. Manual methods of strategy creation are expensive. They create few strategies from which a suitable one is selected for a given program based on its class. The program's class is identified by observing the status of a few manually defined boolean program features. On the other hand, machine learning methods rely on a relatively large set of complex features such as program construct counts, ratios of construct counts or program graphs. In this paper we utilize a machine learning approach to create strategies. This approach combines the strengths of both previously known methods. It uses boolean program features with machine learning to predict verification strategies. Further, we introduce novel program features termed as relative boolean metrics that are boolean abstractions of ratios of construct counts. We implement the novel methods in a tool, extensively evaluate it on a large set of diverse academic benchmarks, and use it to verify four industrial applications. On an average our tool leads the state of the art manual and machine learning-based strategy prediction methods by 11% in terms of the number of properties it successfully verified.