Static speculation as post-link optimization for the grid alu processor

Abstract

In this paper we propose and evaluate a post-link-optimization to increase instruction level parallelism by moving instructions from one basic block to the preceding blocks. The Grid Alu Processor used for the evaluations comprises plenty of functional units that are not completely allocated by the original instruction stream. The proposed technique speculatively performs operations in advance by using unallocated functional units. The algorithm moves instructions to multiple predecessors of a source block. If necessary, it adds compensation code to allow the shifted instructions to work on unused registers, whose values will be copied into the original target registers at the time the speculation is resolved. Evaluations of the algorithm show a maximum speedup of factor 2.08 achieved on the Grid Alu Processor compared to the unoptimized version of the same program due to a better exploitation of the ILP and an optimized mapping of loops. © 2011 Springer-Verlag Berlin Heidelberg.