Online dynamic dependence analysis for speculative polyhedral parallelization

摘要: We present a dynamic dependence analyzer whose goal is to compute dependences from instrumented execution samples of loop nests. The resulting information serves as prediction the behavior during remaining iterations and can be used select apply speculatively optimizing parallelizing polyhedral transformation target sequential nest. Thus, parallel lock-free version generated which should not induce any rollback if correct. computes distance vectors linear functions interpolating memory addresses accessed by each instruction, values some scalars. Phases showing changing are detected thanks adjustment instrumentation frequency. The takes part whole framework dedicated speculative parallelization nests has been implemented with extensions LLVM compiler an x86-64 runtime system.