Register allocation for fine grain threads on multicore processor

Abstract

A multicore processor has multiple processing cores on the same chip. Unicore and multicore processors are architecturally different. Since individual instructions are needed to be scheduled onto one of the available cores, it effectively decreases the number of instructions executed on the individual core of a multicore processor. As each core of a multicore processor has a private register file, it results in reduced register pressure. To effectively utilize the potential benefits of the multicore processor, the sequential program must be split into small parallel regions to be run on different cores, and the register allocation must be done for each of these cores. This article discusses register allocating heuristics for fine grained threads which can be scheduled on multiple cores. Spills are computed and its effect on speed-up, power consumption and performance per power is compared for a RAW benchmark suite.