Guest Editors' Introduction

We are delighted to present a selection of excellent papers in the areas of microarchitecture, compilation, and feed-back directed optimization in this special section of the Journal of Instruction-Level Parallelism.  We invited submissions from authors of the best papers in MICRO 32 and Second Annual Workshop on Feedback-Directed Optimization (FDO99).  These papers were each carefully reviewed by at least three reviewers and revised to ensure journal quality.

Five of the MICRO papers blend compiler and hardware techniques to solve dependence problems.  Papers by Rotenberg and Smith and Moshovos and Sohi implement the compiler-centric techniques of control dependence analysis and redundant load elimination in hardware.  Pai and Adve's paper proposes an analysis of dependence on the first access to cache lines to overlap read misses, and Bharadwaj, Menezes and McKinsey's paper uses the unified analysis of control and data dependences to enhance ILP.  Wolfe and Noonburg's paper applies out-of-order scheduling to a new domain, parallelism among graphics primitives.  The other three papers pertain to various design issues. Albonesi's cache paper addresses the increasingly-important topic of power conservation.  Monreal, et al's paper proposes a scheme that shortens the time that registers are allocated.  MICRO's best paper award went to Austin's paper on microarchitectural verification.

The feedback-directed optimization papers cover three of the key technologies for using information collected at run time on one or more binaries to improve performance on a new binary: combining and weighting multiple profiles to enhance performance of new data sets, determining how to map profile information from one binary to a new, optimized binary whose control flow and perhaps even source have changed, and lastly, a set of feedback-directed optimizations, each quantitatively analyzed.

The area of feedback-directed optimization is currently characterized by a convergence of academic theory and pragmatic engineering in a production environment.  The paper by Savari and Young uses arigorous, information-theoretic approach to gain insight into profiles.  The BMAT paper, awarded best paper, operates on volume-production x87 binaries and the paper by Cohn and Lowney describes techniques implemented in a mature production compiler for the Compaq Alpha.

This web-first special section of the Journal of Instruction-Level Parallelism provides an excellent means of disseminating papers of archival quality in a timely fashion.

Ilan Spillinger and Chris J. Newburn, Co-Guest-Editors