Turbo-charging Lemmas on demand with don't care reasoning

Lemmas on demand is an abstraction/refinement technique for procedures deciding Satisfiability Modulo Theories (SMT), which iteratively refines full candidate models of the formula abstraction until convergence. In this paper, we introduce a dual propagation-based technique for optimizing lemmas on demand by extracting partial candidate models via don't care reasoning on full candidate models. Further, we compare our approach to a justification-based approach similar to techniques employed in the context of model checking. We implemented both optimizations in our SMT solver Boolector and provide an extensive experimental evaluation, which shows that by enhancing lemmas on demand with don't care reasoning, the number of lemmas generated, and consequently the solver runtime, is reduced considerably.