LoRA: a Package for Loop Optimal Register Allocation
Résumé
Instruction-level code parallelization increases the register pressure and renders the register allocation phase crucial. In the case of software pipelined loops, unrolling has to be performed when variables are alive during more than one iteration resulting in code size increases. Loop unrolling also influences the register pressure. LoRA is a package that implements several algorithms for trading the register pressure against code size. In LoRA either the register pressure or the unrolling degree can be constrained. We explain the different strategies used in LoRA and show experimental results on a large benchmark of loops. Our experiments show that in concrete cases the unrolling degree can be kept reasonable although the worst case is exponential in the number of registers thought.