CAREER: Compiler and Run-time System Support for Modern Hardware

University Of Colorado At Boulder, Boulder CO

Investigators

Abstract

To achieve the best performance on modern computers, one must understand and exploit the interactions between program behavior, run-time systems, and hardware. This research explores compiler and run-time system support for improving load-value prediction and for reducing the energy consumption of programs. Load-value prediction exposes instruction-level parallelism in programs by allowing loads to execute in parallel with the instructions that depend on it. Energy reduction techniques reduce the energy requirements of programs without significantly degrading performance. Broadly speaking this research considers two kinds of techniques: (i) Compiler techniques that use easily-available information in programs (such as type information) to transform or annotate code to make better use of hardware; (ii) Run-time system techniques that reorganize and annotate data to make better use of hardware. This research will also consider new hardware mechanisms for load-value prediction and energy consumption that can be exploited in software.