Memory Models for Architects and Compiler Writers

点击翻译按钮获取中文摘要

A new memory model called Commit-Reconcile \& Fences (CRF) ispresented, and it is shown how other memory models can be described asrestricted versions of it. The CRF memory model has been designed sothat it is both easy for architects to implement, and rich enough thatcompilers for high-level parallel languages can use it as a targetmachine language. This proposal is for investigating 1. theintegration of CRF into modern microprocessors, 2. adaptive protocolsfor implementing CRF in Distributed Shared memory Systems (DSM) and3. efficient translation of high-level multithreaded programmingmodels (e.g. Cilk) into CRF. The impact of this research would be to eliminate a plethora ofimplementation-dependent memory models, and to provide a stableinterface for architects and compiler writers. The development ofmicro-architectures and cache-coherence protocols to support CRF couldhave an immediate impact on industry because these changes can beincorporated in PC's and servers without the loss of softwarecompatibility.Another theme in this proposal is the use of a new formal frameworkbased on Term Rewriting Systems (TRS) for describing computer systems.The biggest appeal of TRS's is its naturalness for designers. The TRSframework is useful in verification of system properties. It is alsopossible to synthesize hardware from TRS's. A unified synthesis andverification framework is likely to have a big impact on currentdesign practices.