Property-Driven Fence Insertion Using Reorder Bounded Model Checking

Modern architectures provide weaker memory consistency guarantees than sequential consistency. These weaker guarantees allow programs to exhibit behaviours where the program statements appear to have executed out of program order. Fortunately, modern architectures provide memory barriers (fences) to enforce the program order between a pair of statements if needed. Due to the intricate semantics of weak memory models, the placement of fences is challenging even for experienced programmers. Too few fences lead to bugs whereas overuse of fences results in performance degradation. This motivates automated placement of fences. Tools that restore sequential consistency in the program may insert more fences than necessary for the program to be correct. Therefore, we propose a property-driven technique that introduces reorderbounded exploration to identify the smallest number of program locations for fence placement. We implemented our technique on top of CBMC; however, in principle, our technique is generic enough to be used with any model checker. For a parametric program that increases in size, our technique not only solves more instances but also yields a 17x speedup for the largest instance solved by an earlier approach. We report experimental results on relevant benchmarks and compare with earlier approaches.