Optimizing the translation out-of-SSA with renaming constraints

Static Single Assignment form is an intermediate representation, that uses -functions to merge values at each confluent points of the control flow graph. functions are not machine instructions and should be renamed back to move operations when translating out-of-SSA form. Without a coalescing algorithm, out-of-SSA translation generates many move instructions. In this paper we propose an extension of the algorithm of Leung and George [7] to minimize the -related copies during the out-of-SSA translation. Leung et al. constructed SSA form for programs represented as native machine instructions, including the use of machine dedicated registers. For this purpose, the out-of-SSA translation contains renaming constraints that are represented using a pinning principle. Pinning the -function arguments and their corresponding definition to a common resource is a very attractive technique for coalescing variables, even if this is not a true minimization: this article presents a renaming-constraints aware and pinning-based coalescing algorithm. Even without renaming constraints, the move instructions minimization problem is still considered an open issue [7, 10]. This article provides also a discussion about the formulation of this problem, its complexity and its motivations. Finally, we implemented our algorithm in the STMicroelectronics Linear Assembly Optimizer [2]. This provides many interesting results when comparing several possible approaches. We also explain, using hand crafted examples, the limitations of Leung’s, Sreedhar’s and classical register coalescing [6] algorithms.