Extending the Applicability and Improving the Performance of Runtime Parallelization

When static analysis of a sequential loop fails to yield reliable information on its dependence structure, a parallelizing compiler is left with three alternatives: it can take the conservative option of emitting code for a sequential execution; it can optimistically emit code to speculatively execute the loop as a DOALL [6, 7]; or it can emit inspector-executor code to determine the actual dependence structure at runtime and to respect it in a parallel execution [8, 9]. The rst approach is certain to yield a slow execution. The second approach gives very good results when the loop can in fact be executed as a DOALL, but is of no help otherwise. In this paper we concentrate on the nal approach, runtime parallelization through the inspectorexecutor method. We have two goals in this work. The rst is to expand the class of loop to which the approach may be applied by removing restrictions on the loop dependence structures that it can handle. To achieve this goal, we introduce new forms of the inspector and executor that together remove all restrictions on the loop dependence structure. Thus, we show how to parallelize a class of loop that previously would have compelled the compiler to emit sequential code. Our second goal is to improve the performance of the inspector-executor approach through specializations applicable when static analysis yields some (weak) information about the array indexing functions used in assignments. We validate our work through a set of examples designed to illustrate the fundamental performance tradeo s characterizing the specialized implementations, using results taken from executions on 32 processors of a KSR1.