The Effect of Compiler Optimizations on Available Parallelism in Scalar Programs.
Scott A. Mahlke, Nancy J. Warter, William Y. Chen, Pohua P. Chang, Wen‐mei Hwu · 1991
In this paper we analyze the effect of compiler optimizations on fine grain parallelism in scalar programs. We characterize three levels of optimization: classical, superscalar, and multiprocessor. We show that classical optimizations not only improve a program's efficiency but also its parallelism. Superscalar optimizations further improve the parallelism for moderately parallel machines. For highly parallel machines, however, they actually constrain available parallelism. The multiprocessor optimizations we consider are memory renaming and data migration. Introduction Compiler optimizations are designed to reduce a program's execution time. Traditionally, these optimizations are customized for a given machine model. Classical optimizations are designed to improve the program's efficiency for a machine model which has one thread of execution and can issue one instruction per cycle. Superscalar optimizations are designed for a machine model with a single thread of execution and a lim...