Inter-procedural data-flow analyses are slow. We parallelize these predicate propagation fixpoint algorithms efficiently on a GPU. Our approach is (mostly) synchronization free even though the processed graphs in gene...
详细信息
ISBN:
(纸本)9781450362771
Inter-procedural data-flow analyses are slow. We parallelize these predicate propagation fixpoint algorithms efficiently on a GPU. Our approach is (mostly) synchronization free even though the processed graphs in general are cyclic and have nodes with fan-in and fan-out degrees above 1. We detect and fix any data races that occur while propagating predicates in a SIMD fashion. Additionally, we solve the parallel termination problem by means of heuristics. The GPU-codes of five data-flow analyses are up to 89.8 times faster than their sequential LLVM variants. Offloading the analyses to the GPU saves up to 26.5% of the total compilation time.
暂无评论