Unit propagation with stable watches

Unit propagation is the hottest path in CDCL SAT solvers, therefore the related data-structures, algorithms and implementation details are well studied and highly optimized. State-of-the-art implementations are based on reduced occurrence tracking with two watched literals per clause and one blocking literal per watcher in order to further reduce the number of clause accesses. In this paper, we show that using runtime statistics for watched literal selection can improve the performance of state-of-the-art SAT solvers. We present a method for efficiently keeping track of spans during which literals are satisfied and using this statistic to improve watcher selection. An implementation of our method in the SAT solver CaDiCaL can solve more instances of the SAT Competition 2019 and 2020 benchmark sets and is specifically strong on satisfiable cryptographic instances.