Generating fast code from concurrent program dependence graphs

Reviewer: R. Clayton

Representing a program as a graph recognizes program concurrency as independent subgraphs. Complicating recognition is the need to preserve the program's shared-data semantics, and the desire to minimize concurrency overhead. Annotating the graph with data dependencies, and reconfiguring the graph, takes care of the first complication. A number of approaches address the second complication. One approach is to preschedule one, or a small number of, possible concurrent executions. Prescheduling trades the responsiveness of concurrent execution for minimal concurrency overhead. Given the cost of concurrency on stock systems, this tradeoff is almost always successful. This paper presents a set of annotation, reconfiguration, and prescheduling algorithms for concurrent program dependence graphs (cPDGs). The work described is relative to Esterel, a concurrent programming language, but cPDGs are general enough to represent concurrency in many programming languages, although some necessary assumptions, particularly with respect to acyclic cPDG structure, may cause some problems in the more general case. Some algorithms, because they are concerned with minimization, are NP-complete, and are heuristics; the remaining algorithms are exact. Data presented in the paper shows that the algorithms provide the expected speed-ups, in some cases impressively so, when compared to other approaches for compiling Esterel. The costs of the algorithms, in terms of compilation time or compiler complexity, are not discussed. The paper is short and clear; algorithms are described textually, and in pseudocode. Knowledge of concurrent compilation techniques is not necessary; knowledge of more advanced compilation techniques is helpful. Online Computing Reviews Service