Using the parallel complexity of programs to improve compaction
Marc Pouzet · 2002
This paper focuses on the compaction of tail-recursive programs-arbitrary programs with conditionals, loops and function calls appearing in terminal position-for fine-grain parallel machines like VLIW or superscalars. An iterative compaction method based on instruction move-up and fold/unfold transformations is studied. In order to speed-up compaction and produce shorter schedules, we propose to use some information about the parallel complexity of expressions. This information is used on the fly during the compaction process to decide whether to unfold a function or not.