Architectural specialization for inter-iteration loop dependence patterns

Abstract—Hardware specialization is an increasingly com-mon technique to enable improved performance and energy ef-ficiency in spite of the diminished benefits of technology scal-ing. This paper proposes a new approach called explicit loop specialization (XLOOPS) based on the idea of elegantly en-coding inter-iteration loop dependence patterns in the instruc-tion set. XLOOPS supports a variety of inter-iteration data-and control-dependence patterns for both single and nested loops. The XLOOPS hardware/software abstraction requires only lightweight changes to a general-purpose compiler to gen-erate XLOOPS binaries and enables executing these binaries on: (1) traditional microarchitectures with minimal performance impact, (2) specialized microarchitectures to improve perfor-mance and/or energy efficiency, and (3) adaptive microarchitec-tures that can seamlessly migrate loops between traditional and specialized execution to dynamically trade-off performance vs. energy efficiency. We evaluate XLOOPS using a vertically inte-grated research methodology and show compelling performance and energy efficiency improvements compared to both simple and complex general-purpose processors. I