Compiler Optimizations in Reform Prolog: Experiments on the KSR-1 Multiprocessor

We describe the compiler analyses of Reform Prolog and evaluate their eeectiveness in eliminating suspension and locking on a range of benchmarks. The results of the analysis may also be used to extract non-strict independent and-parallelism. We nd that 90% of the predicate arguments are ground or local, and that 95% of the predicate arguments do not require suspension code. Hence, very few suspension operations need to be generated to maintain sequential semantics. The compiler can also remove unnecessary locking of local data by locking only updates to shared data; however, even though locking writes are reduced to 52% of the unoptimized number for our benchmark set, this has little eeect on execution times. We nd that the ineeectiveness of locking elimination is due to the relative rarity of locking writes, and the execution model of Reform Prolog, which results in few invalidations of shared cache lines when such writes occur. The benchmarks are evaluated on a cache-coherent KSR-1 multiprocessor with physically distributed memory, using up to 48 processors. Speedups scale from previous results on smaller, bus-based multiprocessors, and previous low parallelization overheads are retained.