Integrating Scalar Optimization and Parallelization

Compiling programs to use parallelism and memory hierarchy efficiently requires both parallelizing (high-level) transformations and traditional scalar (low-level) optimizations. Our compiler has only one intermediate language — SUIF (Stanford University Intermediate Form) — for both parallelization and scalar optimizations. Having only one intermediate language offers two advantages: no duplication of functions between the two levels, and availability of high-level information at the low-level. This paper shows how SUIF integrates the two levels of abstraction, and how the SUIF compiler is organized to take advantage of this integration.