Speculative Run-Time Parallelization of Loopsy

Current parallelizing compilers cannot identify a signiicant fraction of fully parallel loops because they have complex or statically insuuciently deened access patterns. Since fully parallel loops arise frequently in practice, we have developed methods to speculatively execute loops concurrently. These methods can be applied to any loop, even if the iteration space of the loop is unknown, as in WHILE Loops or DO Loops with conditional exits. To verify the validity of our speculation, we have devised a fully parallel run-time technique for detecting the presence of cross-iteration data dependences in loops. This technique can also be used to eliminate some memory-related dependences by dynamically privatizing scalars and arrays. We outline a cost/performance analysis that can be performed to decide when the methods should be used. Our conclusion is that they should almost always be applied { because, as we show, the expected speedup for fully parallel loops is signiicant, and the cost of a failed speculation (a not fully parallel loop), is minimal. We present experimental results on loops from the PERFECT Benchmarks which substantiate our conclusion that these techniques can yield signiicant speedups. representative of the positions or policies of the Army or the Government.