compiler orchestrated prefetching via speculation and编译通过投机和预取的预取

《compiler orchestrated prefetching via speculation and编译通过投机和预取的预取》由会员上传分享，免费在线阅读，更多相关内容在教育资源-天天文库。

1、CompilerOrchestratedPrefetchingviaSpeculationandPredicationRodricM.Rabbah1,HariharanSandanagobalane2,MongkolEkpanyapong3,Weng-FaiWong21ComputerScienceandArtificialIntelligenceLaboratory,MassachusettsInstituteofTechnology,rabbah@mit.edu2DepartmentofComputerScience,NationalUniversityofSingapore,{ha

2、rihar1,wongwf}@comp.nus.edu.sg3SchoolofElectricalandComputerEngineering,GeorgiaInstituteofTechnology,pop@ece.gatech.eduABSTRACT1.INTRODUCTIONThispaperintroducesacompiler-orchestratedprefetchingsystemModerntrendsinthedesignofhigh-performancemicroproces-asaunifiedframeworkgearedtowardamelioratingth

3、egapbe-sorscontinuetowidenthegapbetweentherateofdataconsump-tweenprocessingspeedsandmemoryaccesslatencies.Wefo-tionintheprocessorandtherateofdatadeliveryfromthememorycusthescopeoftheoptimizationonspecificsubsetsofthepro-system.Thisratedisparityimpliesasystemmusteffectivelytol-gramdependencegrapht

4、hatsuccinctlycharacterizethememoryerate(ormask)potentiallylongmemorylatencies,lestitsufferse-accesspatternofbothregulararray-basedapplicationsandirregu-vereperformancedegradationastheprocessorstallsandawaitsthelarpointer-intensiveprograms.Weillustratehowprogramembed-deliveryofdatafrommemory.Thep

5、erformanceimplicationsarededprecomputationviaspeculativeexecutioncanaccuratelypre-exacerbatedwhenapplicationsmakeextensiveuseofdynamicallydictandeffectivelyprefetchfuturememoryreferenceswithnegli-allocatedobjectswhichcomplicatetheanalysisoftheapplicationgibleoverhead.Theproposedtechniquesreducet

6、hetotalrunningreferencepatternandofteninhibitthedevelopmentofstatictech-timeofsevenSPECbenchmarksandtwoOLDENbenchmarksbyniquesformaskinglongmemoryaccesslatencies.27%onanItanium2processor.TheimprovementsareinadditionInthispaper,weproposeandevaluateacompilertechniquetotoseveralstate-of-the-artopti

7、mizationsincludingsoftwarepipelin-orchestratetheprefetchingofdataaheadofeventualprocessorref-inganddataprefetching.Inaddition,weusecycle-accuratesimu-erences.Theprefetchingisdrivenbytheprecomputationoffuturelationstoidentify