Date of this Version
Proc. ACM Program. Lang., Vol. 7, No. OOPSLA2, Article 283. Publication date: October 2023. https://doi.org/10.1145/3622859
Interprocedural alias analyses often sacrifice precision for scalability. Thus, modern compilers such as GCC and LLVM implement more scalable but less precise intraprocedural alias analyses. This compromise makes the compilers miss out on potential optimization opportunities, affecting the performance of the application. Modern compilers implement loop-versioning with dynamic checks for pointer disambiguation to enable the missed optimizations. Polyhedral access range analysis and symbolic range analysis enable O(1) range checks for non-overlapping of memory accesses inside loops. However, these approaches work only for the loops in which the loop bounds are loop invariants. To address this limitation, researchers proposed a technique that requires O(log n) memory accesses for pointer disambiguation. Others improved the performance of dynamic checks to single memory access by constraining the object size and alignment. However, the former approach incurs noticeable overhead due to its dynamic checks, whereas the latter has a noticeable allocator overhead. Thus, scalability remains a challenge.
In this work, we present a tool, Rapid, that further reduces the overheads of the allocator and dynamic checks proposed in the existing approaches. The key idea is to identify objects that need disambiguation checks using a profiler and allocate them in different regions, which are disjoint memory areas. The disambiguation checks simply compare the regions corresponding to the objects. The regions are aligned such that the top 32 bits in the addresses of any two objects allocated in different regions are always different. As a consequence, the dynamic checks do not require any memory access to ensure that the objects belong to different regions, making them efficient.
Rapid achieved a maximum performance benefit of around 52.94% for Polybench and 1.88% for CPU SPEC 2017 benchmarks. The maximum CPU overhead of our allocator is 0.57% with a geometric mean of -0.2% for CPU SPEC 2017 benchmarks. Due to the low overhead of the allocator and dynamic checks, Rapid could improve the performance of 12 out of 16 CPU SPEC 2017 benchmarks. In contrast, a state-of-the-art approach used in the comparison could improve only five CPU SPEC 2017 benchmarks.