CAREER: rePLay: a Microarchitecture to Support Dynamic Program Optimization

University Of Illinois At Urbana-Champaign, Urbana IL

Investigators

Abstract

Dynamic optimization is a newly emerging area of computer systems research that offers good potential for large advances in computer performance. The crux of dynamic optimization is the incorporation of dynamic information available only at execution time into the optimization process. Examples of potentially useful dynamic information include the likely behavior of branches and the likely values of operands. More immediate access to run-time information gives dynamic optimizers a notable edge over traditional static optimizers. The rePLay Framework is a set of microarchitectural components for supporting dynamic optimization. RePLay contains hardware that allows the dynamic optimizer to run in parallel with program execution thereby reducing the large overheads associated with dynamic optimization. RePLay contains hardware for identifying candidate regions, for optimizing these regions, fetching regions, and sequencing between them. Furthermore, this microarchitecture creates a tight coupling between the optimizer and hardware that detects run-time events. Initial results indicate that dynamic information can be used to create very long (approx 100 instructions) atomic, single entry, single exit regions of code for optimization. The research focuses on the development of the rePLay Framework, specifically in the development of hardware-based optimizations and hardware schemes to support potentially speculative value-based optimizations.