ADCL: Acceleration Driven Clause Learning for Constrained Horn Clauses (2303.01827v3)

Abstract: Constrained Horn Clauses (CHCs) are often used in automated program verification. Thus, techniques for (dis-)proving satisfiability of CHCs are a very active field of research. On the other hand, acceleration techniques for computing formulas that characterize the N-fold closure of loops have successfully been used for static program analysis. We show how to use acceleration to avoid repeated derivations with recursive CHCs in resolution proofs, which reduces the length of the proofs drastically. This idea gives rise to a novel calculus for (dis)proving satisfiability of CHCs, called Acceleration Driven Clause Learning (ADCL). We implemented this new calculus in our tool LoAT and evaluate it empirically in comparison to other state-of-the-art tools.