Avoiding unnecessary remappings at run-time by means of a strategic placement of distribution assignments (DAP) is a major means for improving the run-time efficiency of data-parallel programs on distributed-memory architectures. Recently, we presented a novel and aggressive intraprocedural algorithm achieving this by eliminating partially redundant and partially dead distribution assignments. Here, we show how to enhance this approach interprocedurally. Surprisingly at first sight, it turns out that a straightforward adaption of the intraprocedural approach fails because central properties being valid for the intraprocedural case do not carry over to the interprocedural one revealing severe anomalies. After discussing the essential differences and analogies of DAP in the intraprocedural and interprocedural case, we show how to overcome these anomalies in order to arrive at a powerful and flexible approach for interprocedural DAP (IDAP). As in the intraprocedural case we get a hierarchy of IDAP-algorithms of varying power and efficiency supporting user-customized solutions. First practical experiences underline its importance and effectivity.

Top