Language Extensions in Support of Compiler Parallelization

In this paper, we propose an approach to automatic compiler parallelization based on language extensions that is applicable to a broader range of program structures and application domains than in past work. As a complement to ongoing work on high productivity languages for explicit parallelism, the basic idea in this paper is to make sequential languages more amenable to compiler parallelization by adding enforceable declarations and annotations. Specifically, we propose the addition of annotations and declarations related to multidimensional arrays, points, regions, array views, parameter intents, array and object privatization, pure methods, absence of exceptions, and gather/reduce computations. In many cases, these extensions are also motivated by best practices in software engineering, and can also contribute to performance improvements in sequential code. A detailed case study of the Java Grande Forum benchmark suite illustrates the obstacles to compiler parallelization in current object-oriented languages, and shows that the extensions proposed in this paper can be effective in enabling compiler parallelization. The results in this paper motivate future work on building an automatically parallelizing compiler for the language extensions proposed in this paper.