Optimized Execution of Fortran 90 Array Language on Symmetric Shared-Memory Multiprocessors

Abstract

Past compilers have found it challenging to implement Fortran 90 array language on symmetric shared-memory multiprocessors (SMPs) so as to match, let alone beat, the performance of comparable Fortran 77 scalar loops. This is in spite of the fact that the semantics of array language is implicitly concurrent and the semantics of scalar loops is implicitly sequential. A well known obstacle to efficient execution of array language lies in the overhead of using array temporaries to obey the fetch-before-store semantics of array language. We observe that another major obstacle to supporting array language efficiently arises from the fact that most past compilers attempted to compile and optimize each array statement in isolation. In this paper, we describe a solution for optimized compilation of Fortran 90 array language for execution on SMPs. Our solution optimizes scalarized loops and scalar loops in a common framework. Our solution also adapts past work on array temporary minimization so as to avoid degradation of parallelism and locality. This solution has been implemented in the IBM XL Fortran product compiler for SMPs. To the best of our knowledge, no other Fortran 90 compiler performs such combined optimizations of scalarized loops and scalar loops. Our preliminary experimental results indicate that the performance of Fortran 90 array language can match, and even beat, the performance of comparable scalar loops. In addition to Fortran 90 array language, the approach outlined in this paper will be relevant to similar array language extensions that might appear in Java and other programming languages in the future.