Extending Index-Array Properties for Data Dependence Analysis

Automatic parallelization is an approach where a compiler analyzes serial code and identifies computations that can be rewritten to leverage parallelism. Many data dependence analysis techniques have been developed to determine which loops in a code can be parallelized. With code that includes indirect array accesses through what are commonly called index arrays, such data dependence analysis is restricted in the conclusions that can be drawn at compile time. Various approaches that use index array properties such as monotonicity have been shown to more effectively find parallel loops. In this paper, we extend the kinds of properties about index arrays that can be expressed, show how to convert loop-carried data dependence relations and relevant index-array properties to constraints that can be provided to the Z3 SMT solver, and evaluate the impact of using such index-array properties on identifying parallel loops in a set of numerical benchmarks.