Efficient pointer management of stack data for software managed multicores

Scratchpad-memory (SPM) based memory hierarchy is a promising alternative to cache-based memory hierarchies, due to the difficulty in scaling caches to processors with high core count. However, explicit data management in software is required on SPM-based memory hierarchies. This paper focuses on optimizing the stack data management on SPM-based multicore processors, as memory accesses to call stack present in most applications. While previous works have developed techniques to enable correct stack pointer management, they have not optimized it. As a result, existing techniques still incur high overhead. This paper proposes an automated compiler-based scheme for efficient pointer management. Our experiments on MiBench benchmarks demonstrate that our scheme almost completely eliminates pointer management overhead. As a result, as compared to the state-of-the-art approach, our approach reduces the average execution time of the benchmarks by 52%. Furthermore, with our approach, the performance of stack management on SPM becomes better than hardware caches on average even with conservative estimates.