Array SSA for Explicitly Parallel Programs

The usefulness and applicability of the Static Single Assignment (SSA) framework is undisputed. SSA was originally crafted for sequential programs manipulating scalars, but it has been separately extended to parallel programs on the one hand, and to sequential programs with arrays on the other. In an Array SSA framework, arrays are precisely handled on an element-per-element basis. This paper proposes an Array SSA form for parallel programs with either weak or strong memory consistency, with event-based synchronization or mutual exclusion, with parallel sections or indexed parallel constructs. 1 Introduction and Related Work The usefulness and applicability of the Static Single Assignment (SSA) framework 5] is undisputed, and still is the subject of active research (e.g.,,22]). SSA was originally crafted for sequential programs manipulating scalars, but it has been extended recently to sequential programs with arrays 11]. In an Array SSA framework, arrays are precisely handled on an element-per-element basis. Array SSA was already successfully applied to traditional optimizations such as constant propagation 21]. Moreover, extending reaching deenition analyses (RDAs) and SSA to parallel programs has been a very active area of research too. Data ow analyses for parallel programs, and in particular reaching deenition analyses, were proposed in 8, 10, 3]. SSAs for various kinds of parallel language models were proposed in 17, 18, 14, 23]. Contrasting papers presenting an RDA with papers about SSA is easy. Several other issues, however, make the comparison of related papers a tricky task. Questions that arise include: 1. Does the paper apply element-wise to arrays? This issue is extremely important for scientiic and \high-performance" programming, but most papers apply to scalars only or to arrays considered as one single variable (exceptions include 3, 11, 21]). 2. Does the paper consider both the weak and the strong