Synthesis of domain specific CNF encoders for bit-vector solvers

Abstract

SMT solvers are at the heart of a number of software engineering tools. These SMT solvers use a SAT solver as the back-end and convert the high-level constraints given by the user down to low-level boolean formulas that can be efficiently mapped to CNF clauses and fed into a SAT solver. Current SMT solvers are designed to be general purpose solvers that are suited to a wide range of problems. However, SAT solvers are very non-deterministic and hence, it is difficult to optimize a general purpose solver across all different problems. In this thesis, we propose a system that can automatically generate parts of SMT solvers in a way that is tailored to particular problem domains. In particular, we target the translation from high-level constraints to CNF clauses which is one of the crucial parts of all SMT solvers. We achieve this goal by using a combination of program synthesis and machine learning techniques. We use a program synthesis tool called Sketch to generate optimal encoding rules for this translation and then use auto-tuning to only select the subset of these encodings that actually improve the performance for a particular class of problems. Using this technique, the thesis shows that we can improve upon the basic encoding strategy used by CVC4 (a state of the art SMT solver). We can automatically generate variants of the solver tailored to different domains of problems represented in the bit-vector benchmark suite from the SMT competition 2015.