On a Technique for Transparently Empowering Classical Compiler Optimizations on Multithreaded Code

摘要: A large body of data-flow analyses exists for analyzing and optimizing sequential code. Unfortunately, much it cannot be directly applied on parallel code, reasons correctness. This article presents a technique to automatically, aggressively, yet safely apply sequentially-sound transformations, without change, shared-memory programs. The is founded the notion program references being “siloed” certain control-flow paths. Intuitively, siloed are free interference from other threads within confines such Data-flow transformations can, in general, unblocked references.The solution has been implemented widely used compiler. Results benchmarks SPLASH-2 show that performance improvements up 41p possible, with an average improvement 6p across all tested programs over thread counts.