Non-monotonic Refinement of Control Abstraction for Concurrent Programs

Verification based on abstraction refinement is a successful technique for checking program properties. Conventional abstraction refinement schemes increase precision of the abstraction monotonically, and therefore cannot recover from overly precise refinement decisions. This problem is exacerbated in the context of multi-threaded programs, where keeping track of all control locations in concurrent threads is the inevitably discovered abstraction and is prohibitively expensive. In contrast to the conventional (partition refinement-based) approaches, non-monotonic abstraction refinement schemes rely on re-partitioning and have promising potential for avoiding excess of precision. In this paper, we propose a non-monotonic refinement scheme for the control abstraction (of concurrent programs). Our approach employs a constraint solver to discover re-partitioning at each refinement step. An experimental evaluation of our non-monotonic control abstraction refinement on a collection of multi-threaded verification benchmarks indicates its effectiveness in practice.