Enhanced Parallelization via Analyses and Transformations on Array SSA Form

Array SSA form is a version of SSA form that captures precise element-level data flow information for array variables. As an example of program analysis using Array SSA form, we presented a conditional constant propagation algorithm that can lead to discovery of a larger set of constants than previous algorithms that analyze only scalar variables [7]. As an example of program transformation using Array SSA form, we showed that making its elementlevel functions manifest at run-time can increase the set of parallelizable loops [6]. In this paper, we extend our past work by showing how Array SSA form can be used to perform analyses and transformations that go beyond loop parallelization. These transformations (which include region/task parallelism, pipeline parallelism, speculation, and general forms of code reordering) rely on two important properties of Array SSA form. First, because the operation is truly functional in Array SSA form, it can be subjected to classical optimizations and transformations (e.g., copy propagation, reassociation, code reordering) like other operations. Second, Array SSA form, like scalar SSA form, creates a factoring of reaching definitions. However, in the array case this factoring can be refined by using element-level information. Specifically, we present a static analysis technique called resolution that reduces the number of definitions that appear to reach a use. Resolution is global in scope and is distinct from traditional dependence analysis which focuses on array references in loops.