Systematic Detection of Parallelism in Ordinary Programs

摘要

This dissertation discusses a general model for compilers that take imperativecode written for sequential machines (ordinary code) and detect the parallelism in that code that is compatible with the semantics of the underlying programming language. This model is based on the idea of separating the concerns of parallelism detection and parallelism exploitation. This separation is made possible by having the detection component provide an explicit representation of the parallelism available in the original code. This explicitly parallel representation is based on a formalization of the notion of permissible execution sequences for a given mass of code. The model discussed here prescribes the structure of the parallelism detector. This structure depends on (1) recognizing a hierarchical structure on a graph representation of the program, and (2) separately encoding parallelization conditions and effects. Opportunities for parallelization can then be discovered by traversing the hierarchical structure from the bottom up. During this traversal, progressively larger parts of the program are compared against the independently encoded conditions, and transformed when the conditions are satisfied. The hierarchy guarantees that the results of transforming a piece of a program propagate in time to affect the possible parallelization of larger pieces. Although some of the algorithms used have exponential worst cases for general graphs, their observed behavior on real flow graphs is no worse than quadratic on the size of the original program.