Dependency Collapsing in Instruction-Level Parallel Architectures

Processors that employ instruction fusion can improve performance and energy usage beyond traditional processors by collapsing and simultaneously executing dependent instruction chains on the critical path. This paper describes compiler mechanisms that can facilitate and guide instruction fusion in processors built to execute fused instructions. The compiler support discussed in this paper includes compiler annotations to guide fusion, exploring multiple new fusion configurations, and developing scheduling algorithms that effectively select and order fusible instructions. The benefits of providing compiler support for dependent instruction fusion include statically detecting fusible instruction chains without the need for hardware dynamic detection support and improved performance by increasing available parallelism.