Modern high-performance microprocessors with their deep pipelines, multi-level memory systems, and superscalar microarchitectures provide ample opportunities for profile-driven, machine-specific optimizations. In this talk, I will introduce Machine SUIF, an extension of the Stanford SUIF compiler, and describe our work in profile-driven, machine-specific optimization. We have used Machine SUIF to develop several optimizations that rely heavily on code motion and code duplication. To maximize the performance benefits of these optimizations, we have pioneered the use of profiles that summarize temporal information. Temporal profiles provide more information about the program execution than traditional profiles but are much less expense to analyze and store than full traces. In particular, we have used path information (information that summarizes how often execution follows a sequence of code blocks) to improve the static predictability of conditional branches and direct the trace selection portion of a trace-based global instruction scheduler. We have also employed temporal ordering information (information that summarizes the interleaving of code blocks in a program trace) in a new algorithm for procedure placement. As demonstrated by our simulation results, each of these optimizations can improve application performance on today's modern microarchitectures.

Machine SUIF is part of the NSF- and DARPA-funded National Compiler Infrastructure project.