A failed proof can yield a useful test

Abstract

AbstractA successful automated program proof is, in software verification, the ultimate triumph. In practice, however, the road to such success is paved with many failed proof attempts. Unlike a failed test, which provides concrete evidence of an actual bug in the program, a failed proof leaves the programmer in the dark. Can we instead learn something useful from it? The work reported here takes advantage of the rich information that some automatic provers internally collect about the program when attempting a proof. If the proof fails, the Proof2Test tool presented in this article uses the counterexample generated by the prover (specifically, the SMT solver underlying the Boogie tool used in the AutoProof system to perform correctness proofs of contract‐equipped Eiffel programs) to produce a failed test, which provides the programmer with immediately exploitable information to correct the program. The discussion presents Proof2Test and the application of the ideas and tool to a collection of representative examples.