Random model sampling: making Craig interpolation work when it should not

One of the most serious problems when doing program analyses is dealing with function calls. While function inlining is the traditional approach to this problem, it nonetheless suffers from the increase in analysis complexity due to the state space explosion. Craig interpolation has been successfully used in recent years in the context of bounded model checking to do function summarization which allows one to replace the complete function body with its succinct summary and, therefore, reduce the complexity. Unfortunately this technique can be applied only to a pair of unsatisfiable formulae.
In this work-in-progress paper we present an approach to function summarization based on Craig interpolation that overcomes its limitation by using random model sampling. It captures interesting input/output relations, strengthening satisfiable formulae into unsatisfiable ones and thus allowing the use of Craig interpolation. Preliminary experiments show the applicability of this approach; in our future work we plan to do a full evaluation on real-world examples.