EAGER: Software Exposed Hardware Signatures for Code Analysis, Optimization and Debugging

North Carolina State University, Raleigh NC

Investigators

Abstract

Effective dynamic memory disambiguation limits many practical approaches for code analysis, optimization, and debugging. An efficient solution to this problem could unlock additional performance and increase the reliability of computer systems. A potential avenue for overcoming the traditional limits of hardware memory disambiguation is through the use of signature registers. Such registers can operate on hundreds of addresses simultaneously and can be exposed to software through a flexible and general interface for use in a wide variety of systems. However, because signatures represent sets of addresses imprecisely, they are prone to false positives which limit their accuracy and effectiveness. This research investigates the effectiveness of exposing signature registers to software. Using scenarios from code analysis, optimization, and debugging as case studies, the effectiveness of signatures is compared to other hardware memory disambiguation techniques to evaluate their performance, complexity, and power costs. The hardware implementation for signature registers is investigated and refined to reduce false positives.