Optimal Global Instruction Scheduling for the Itanium

On the Itanium 2 processor, effective global instruction scheduling is crucial to high performance. At the same time, it poses a challenge to the compiler: This code generation subtask involves strongly interdependent decisions and complex trade-offs that are difficult to cope with for heuristics. We tackle this NP-complete problem with integer linear programming (ILP), a search-based method that yields provably optimal results. This promises faster code as well as insights into the potential of the architecture. Our ILP model comprises global code motion with compensation copies, predication, and Itanium-specific features like control/data speculation. In integer linear programming, well-structured models are the key to acceptable solution times. The feasible solutions of an ILP are represented by integer points inside a polytope. If all vertices of this polytope are integral, then the ILP can be solved in polynomial time. We define two subproblems of global scheduling in which some constraint classes are omitted and show that the corresponding two subpolytopes of our ILP model are integral and polynomial sized. This substantiates that the found model is of high efficiency, which is also confirmed by the reasonable solution times. The ILP formulation is extended by further transformations like cyclic code motion, which moves instructions upwards out of a loop, circularly in the opposite direction of the loop backedges. Since the architecture requires instructions to be encoded in fixed-sized bundles of three, a bundler is developed that computes bundle sequences of minimal size by means of precomputed results and dynamic programming. Experiments have been conducted with a postpass tool that implements the ILP scheduler. It parses assembly procedures generated by Intel’s Itanium compiler and reschedules them as a whole. Using this tool, we optimize a selection of hot functions from the SPECint 2000 benchmark. The results show a significant speedup over the original code.