TSO-Atomicity: Efficient TSO Enforcement for Aggressive Program Optimization

Strong memory consistency models like Sequential Consistency (SC) and Total Store Ordering (TSO) impose tremendous memory consistency constraints on program optimizations. Existing techniques leverage HW atomicity supports to enable aggressive optimizations while enforcing sequential consistency. By grouping the optimized code into a region and committing the whole region atomically, sequential consistency is preserved. But for optimization in TSO, the sequential consistency enforced by HW atomicity supports imposes stronger consistency than necessary, resulting in in inefficient region commit operations. In this paper, we introduce the HW support for a novel TSO-atomicity, which provides weaker consistency than atomicity for efficient region commit, but still can enforce TSO for aggressive optimizations. We show in our experiments that, dynamic binary optimizations without HW atomicity supports can improve binary program performance by only 6.7% and 8.1% in frame-level (around 20 instructions per frame) and region-level binary optimization (around 70 instructions per region) respectively. With HW atomicity supports, region-level binary optimization can improve the performance further by 7.8%. However, frame-level binary optimization actually losses performance by 2.7% due to the overheads associated with the commit for atomicity. HW supports for TSOatomicity can reduce the overheads at region commit and achieve a total 12% and 20.4% performance gain for frame-level and region-level binary optimization respectively.