Data-Dependency Graph Transformations for Instruction Scheduling

Instruction scheduling is one of the most important code optimizations because of its role in increasing processor utilization. Instruction scheduling reorders instructions to minimize pipeline stalls while preserving program semantics. For realistic processor models, this problem is NP-hard, and numerous heuristic and optimal (enumerative) techniques have been proposed. Heuristics can sometimes produce high-quality schedules, but do not guarantee optimal or near-optimal results. Optimal techniques can guarantee an optimal schedule, but scheduling time may be long or no optimal schedule may be found within a bounded time. This dissertation presents a set of novel graph transformations which addresses these deficiencies. The transformations can be applied as a pre-processing step to any instruction scheduling technique. Heuristically scheduling the transformed problems significantly reduces the expected execution time of the schedule for hard problems, and captures nearly all the improvement possible beyond scheduling the untransformed problem. The transformations significantly reduce the solution time of optimal schedulers, and more problems are solved within a bounded time. The transformations are described specifically for basic blocks, superblocks, and traces, and are generalized to arbitrary acyclic regions. Efficient algorithms are described for each transformation, and experimental data demonstrate that the algorithms are fast for real problems. Experimental results for scheduling regions from a set of industry standard benchmarks compiled using the GNU Compiler Collection (GCC) confirm the effectiveness of the transformations at improving heuristic and optimal instruction scheduling.