Decomposing Meeting Graph Circuits to Minimise Kernel Loop Unrolling

International audienceThis article studies an important open problem in backend compilation regarding loop unrolling after periodic register allocation. Although software pipelining is a powerful technique to extract fine-grain parallelism, variables can stay alive across more than one kernel iteration, which is challenging for code generation. The classical software solution that does not alter the computation throughput consists in unrolling the loop a posteriori (13; 12). However, the resulting unrolling degree is often unacceptable and may reach absurd levels. Alternatively, loop unrolling can be avoided thanks to software register renaming. This is achieved through the insertion of move operations. However, inserting those operations may increase the initiation interval (II) and nullifies the benefits of software pipelining itself. We propose in this article a new technique to minimise the loop unrolling degree generated after periodic register allocation. In fact, this technique consists on decomposing the generated meeting graph circuits by inserting move instructions without compromising the throughput benefits of software pipelining. The different experiments showed that the execution time is acceptable and good results can be produced when we have many functional units which can execute move operations