Itanium+2上编译器性能调优方法的分析

《Itanium+2上编译器性能调优方法的分析》由会员上传分享，免费在线阅读，更多相关内容在学术论文-天天文库。

1、AbstractCompilerisasystemsoftwarethattranslateshighlevelprogramminglanguageinto．semanticallyequivalentmachinecode．Inthistranslationprocess，compileroftenperformsVDXiOU8optimizationstotheinputsourcecode．80thatbetterperformanceorsmallercodesizecanbeachieved．Forgeneralpurposeproc

2、essor’scompilers。performanceisalwaystheultimategoal．Alltheoptimizationsaredesignedtowardhigherperformance．However，duringthecompilerdevelopment，Weoftenmeetthecasethatafterahugeefforthasbeenspent；westillgainverylittleperformanceimprovement，sometimesevenperformancedecrease．Tosol

3、vethisproblem,weneedtodoperformanceanalysi8andtuning，whosemainpurposeistoidentifytheprogram’8bottleneck．Throughtheanalysisofthesebottlenecks．wec蛐findoutthedeficienciesofcurrentoptimizationalgorithmsandmakefurtherimprovement．Itaninm2isajointeffortbyIntelandHPforhighperformance

4、internetserversandworkstations．ItisthesecondimplementationofIA-64architecture．WithEPIC(ExplicitlyParallelInstructionComputing)stylefortightcouplingbetweenhardwareandsoftware，theItanium2hardware-softwareinterfacecanmoreefficientlyexploitthecompilationtimeinformation，thusputtin

5、gmoreemphasisOntheeffectsofcompileroptimiz眦ions．恤paperdealswiththeperformanceanalysisandtuningmethodologyoftheha-nium2architecture，especiallytheperformancetuningofcompileroptimizations．First，theconceptofperformancetuningandanalysisisintroduced．Then，thegeneralmeth-otisandcommo

6、nlyusedtoolsforperformancetuning．aswellastheerchitecturesupportprovidedbyItanium2，arediscussed．Finally,oIlecasestudyisshown．ThiscasestudyshowourexperiencesintunmgtheORC(openresearchcompiler)compiler,onSPEC2000FPca8e：mgrid．Duringthetuningprocess，wefoundout：1．Thelocalityhintgen

7、erationalgorithminORCcompilershouldbeimproved，be-cause8UthecurrenthintsarejustntawhichmeansnotemporallocalityinallleveIs．2．ORCcompiler'scodegenerationdoesnotutilizeItaniumarchitecturefeaturesverywell(e．g．theimmediatebaseupdateforminlfotchinstruction)3．Thetooaggressivesplittin

8、ginliverangesplittinginregisterallocation．Thismayresultinredundantco