Interprocedural optimization - Wikipedia

WPO and LTO Interproceduraloptimization FromWikipedia,thefreeencyclopedia Jumptonavigation Jumptosearch Computerprogramoptimizationmethod Interproceduraloptimization(IPO)isacollectionofcompilertechniquesusedincomputerprogrammingtoimproveperformanceinprogramscontainingmanyfrequentlyusedfunctionsofsmallormediumlength.IPOdiffersfromothercompileroptimizationbecauseitanalyzestheentireprogram;otheroptimizationslookatonlyasinglefunction,orevenasingleblockofcode. IPOseekstoreduceoreliminateduplicatecalculations,inefficientuseofmemory,andtosimplifyiterativesequencessuchasloops.Ifthereisacalltoanotherroutinethatoccurswithinaloop,IPOanalysismaydeterminethatitisbesttoinlinethat.Additionally,IPOmayre-ordertheroutinesforbettermemorylayoutandlocality. IPOmayalsoincludetypicalcompileroptimizationsonawhole-programlevel,forexampledeadcodeelimination(DCE),whichremovescodethatisneverexecuted.Toaccomplishthis,thecompilertestsforbranchesthatarenevertakenandremovesthecodeinthatbranch.IPOalsotriestoensurebetteruseofconstants.ModerncompilersofferIPOasanoptionatcompile-time.TheactualIPOprocessmayoccuratanystepbetweenthehuman-readablesourcecodeandproducingafinishedexecutablebinaryprogram. Forlanguagesthatcompileonafile-by-filebasis,effectiveIPOacrosstranslationunits(modulefiles)requiresknowledgeofthe"entrypoints"oftheprogramsothatawholeprogramoptimization(WPO)canberun.Inmanycases,thisisimplementedasalink-timeoptimization(LTO)pass,becausethewholeprogramisvisibletothelinker. Contents 1Analysis 2WPOandLTO 3Example 3.1Ingeneral 4History 5Flagsandimplementation 5.1Unix-like 5.1.1Non-LTOoptions 5.2Other 6Seealso 7References 8Externallinks Analysis[edit] Theobjectiveofanyoptimizationforspeedistohavetheprogramrunasswiftlyaspossible;theproblemisthatitisnotpossibleforacompilertocorrectlyanalyzeaprogramanddeterminewhatitwilldo,muchlesswhattheprogrammerintendedforittodo.Bycontrast,humanprogrammersstartattheotherendwithapurpose,andattempttoproduceaprogramthatwillachieveit,preferablywithoutexpendingalotofthoughtintheprocess. Forvariousreasons,includingreadability,programsarefrequentlybrokenupintoanumberofprocedures,whichhandleafewgeneralcases.However,thegeneralityofeachproceduremayresultinwastedeffortinspecificusages.Interproceduraloptimizationrepresentsanattemptatreducingthiswaste. SupposethereisaprocedurethatevaluatesF(x),andthecoderequeststheresultofF(6)andthenlater,F(6)again.Thissecondevaluationisalmostcertainlyunnecessary:theresultcouldhaveinsteadbeensavedandreferredtolater,assumingthatFisapurefunction.ThissimpleoptimizationisfoiledthemomentthattheimplementationofF(x)becomesimpure;thatis,itsexecutioninvolvesreferencetoparametersotherthantheexplicitargument6thathavebeenchangedbetweentheinvocations,orsideeffectssuchasprintingsomemessagetoalog,countingthenumberofevaluations,accumulatingtheCPUtimeconsumed,preparinginternaltablessothatsubsequentinvocationsforrelatedparameterswillbefacilitated,andsoforth.Losingthesesideeffectsvianon-evaluationasecondtimemaybeacceptable,ortheymaynot. Moregenerally,asidefromoptimization,thesecondreasontouseproceduresistoavoidduplicationofcodethatwouldproducethesameresults,oralmostthesameresults,eachtimetheprocedureisperformed.Ageneralapproachtooptimizationwouldthereforebetoreversethis:someorallinvocationsofacertainprocedurearereplacedbytherespectivecode,withtheparametersappropriatelysubstituted.Thecompilerwillthentrytooptimizetheresult. WPOandLTO[edit] Wholeprogramoptimization(WPO)isthecompileroptimizationofaprogramusinginformationaboutallthemodulesintheprogram.Normally,optimizationsareperformedonapermodule,"compiland",basis;butthisapproach,whileeasiertowriteandtestandlessdemandingofresourcesduringthecompilationitself,doesnotallowcertaintyaboutthesafetyofanumberofoptimizationssuchasaggressiveinliningandthuscannotperformthemeveniftheywouldactuallyturnouttobeefficiencygainsthatdonotchangethesemanticsoftheemittedobjectcode. Link-timeoptimization(LTO)isatypeofprogramoptimizationperformedbyacompilertoaprogramatlinktime.Linktimeoptimizationisrelevantinprogramminglanguagesthatcompileprogramsonafile-by-filebasis,andthenlinkthosefilestogether(suchasCandFortran),ratherthanallatonce(suchasJava'sjust-in-timecompilation(JIT)). Onceallfileshavebeencompiledseparatelyintoobjectfiles,traditionally,acompilerlinks(merges)theobjectfilesintoasinglefile,theexecutable.However,inLTOasimplementedbytheGNUCompilerCollection(GCC)orLLVM,thecompilerisabletodumpitsintermediaterepresentation(GIMPLEbytecodeorLLVMbitcode)todisksothatallthedifferentcompilationunitsthatwillgotomakeupasingleexecutablecanbeoptimizedasasinglemodulewhenthelinkfinallyhappens.Thisexpandsthescopeofinterproceduraloptimizationstoencompassthewholeprogram(or,rather,everythingthatisvisibleatlinktime).Withlink-timeoptimization,thecompilercanapplyvariousformsofinterproceduraloptimizationtothewholeprogram,allowingfordeeperanalysis,moreoptimization,andultimatelybetterprogramperformance. Inpractice,LTOdoesnotalwaysoptimizetheentireprogram--libraryfunctions,especiallydynamicallylinkedsharedobjects,areintentionallykeptouttoavoidexcessiveduplicationandtoallowforupdating.StaticlinkingdoesnaturallylendtotheconceptofLTO,butitonlyworkswithlibraryarchivesthatcontainIRobjectsasopposedtomachine-codeonlyobjectfiles.[1]Duetoperformanceconcerns,noteventheentireunitisalwaysdirectlyused--aprogramcouldbepartitionedinadivide-and-conquerstyleLTOsuchasGCC'sWHOPR.[2]Andofcourse,whentheprogrambeingbuiltisitselfalibrary,theoptimizationwouldkeepeveryexternally-available(exported)symbol,withouttryingtoohardatremovingthemasapartofDCE.[1] AmuchmorelimitedformofWPOisstillpossiblewithoutLTO,asexemplifiedbyGCC's-fwhole-programswitch.ThismodemakesGCCassumethatthemodulebeingcompiledcontainstheentrypoint(usuallymain())oftheentireprogram,sothateveryotherfunctioninitisnotexternallyusedandcanbesafelyoptimizedaway.Sinceitonlyappliestoonemoduleonly,itcannottrulyencompassthewholeprogram.(ItcanbecombinedwithLTOintheone-big-modulesense,usefulwhenthelinkerisnotcommunicatingbacktoGCCaboutwhatentrypointsorsymbolsarebeingusedexternally.)[1] Example[edit] Programexample; integerb;{Avariable"global"totheprocedureSilly.} ProcedureSilly(a,x) ifx<0thena:=x+belsea:=-6; EndSilly;{Referencetob,notaparameter,makesSilly"impure"ingeneral.} integera,x;{ThesevariablesarevisibletoSillyonlyifparameters.} x:=7;b:=5; Silly(a,x);write(x); Silly(x,a);write(x); Silly(b,b);write(b); Endexample; IftheparameterstoSillyarepassedbyvalue,theactionsoftheprocedurehavenoeffectontheoriginalvariables,andsinceSillydoesnothingtoitsenvironment(readfromafile,writetoafile,modifyglobalvariablessuchasb,etc.)itscodeplusallinvocationsmaybeoptimizedawayentirely,leavingthevalueofaundefined(whichdoesn'tmatter)sothatjusttheprintstatementsremain,andtheyforconstantvalues. Ifinsteadtheparametersarepassedbyreference,thenactiononthemwithinSillydoesindeedaffecttheoriginals.Thisisusuallydonebypassingthemachineaddressoftheparameterstotheproceduresothattheprocedure'sadjustmentsaretotheoriginalstoragearea. Thusinthecaseofcallbyreference,procedureSillyhasaneffect.Supposethatitsinvocationsareexpandedinplace,withparametersidentifiedbyaddress:thecodeamountsto x:=7;b:=5; ifx<0thena:=x+belsea:=-6;write(x);{aischanged.} ifa<0thenx:=a+belsex:=-6;write(x);{Becausetheparametersareswapped.} ifb<0thenb:=b+belseb:=-6;write(b);{TwoversionsofvariablebinSilly,plustheglobalusage.} Thecompilercouldtheninthisrathersmallexamplefollowtheconstantsalongthelogic(suchasitis)andfindthatthepredicatesoftheif-statementsareconstantandso... x:=7;b:=5; a:=-6;write(7);{bisnotreferenced,sothisusageremains"pure".} x:=-1;write(-1);{bisreferenced...} b:=-6;write(-6);{bismodifiedviaitsparametermanifestation.} Andsincetheassignmentstoa,bandxdelivernothingtotheoutsideworld-theydonotappearinoutputstatements,norasinputtosubsequentcalculations(whoseresultsinturndoleadtooutput,elsetheyalsoareneedless)-thereisnopointinthiscodeeither,andsotheresultis write(7); write(-1); write(-6); Avariantmethodforpassingparametersthatappeartobe"byreference"iscopy-in,copy-outwherebytheprocedureworksonalocalcopyoftheparameterswhosevaluesarecopiedbacktotheoriginalsonexitfromtheprocedure.IftheprocedurehasaccesstothesameparameterbutindifferentwaysasininvocationssuchasSilly(a,a)orSilly(a,b),discrepanciescanarise.So,iftheparameterswerepassedbycopy-in,copy-outinleft-to-rightorderthenSilly(b,b)wouldexpandinto p1:=b;p2:=b;{Copyin.Localvariablesp1andp2areequal.} ifp2<0thenp1:=p2+belsep1:=-6;{Thusp1maynolongerequalp2.} b:=p1;b:=p2;{Copyout.Inleft-to-rightorder,thevaluefromp1isoverwritten.} Andinthiscase,copyingthevalueofp1(whichhasbeenchanged)tobispointless,becauseitisimmediatelyoverwrittenbythevalueofp2,whichvaluehasnotbeenmodifiedwithintheprocedurefromitsoriginalvalueofb,andsothethirdstatementbecomes write(5);{Not-6} Suchdifferencesinbehaviorarelikelytocausepuzzlement,exacerbatedbyquestionsastotheorderinwhichtheparametersarecopied:willitbelefttorightonexitaswellasentry?Thesedetailsareprobablynotcarefullyexplainedinthecompilermanual,andiftheyare,theywilllikelybepassedoverasbeingnotrelevanttotheimmediatetaskandlongforgottenbythetimeaproblemarises.If(asislikely)temporaryvaluesareprovidedviaastackstoragescheme,thenitislikelythatthecopy-backprocesswillbeinthereverseordertothecopy-in,whichinthisexamplewouldmeanthatp1wouldbethelastvaluereturnedtobinstead. Theprocessofexpandingaprocedurein-lineshouldnotberegardedasavariantoftextualreplacement(asinmacroexpansions)becausesyntaxerrorsmayariseaswhenparametersaremodifiedandtheparticularinvocationusesconstantsasparameters.Becauseitisimportanttobesurethatanyconstantssuppliedasparameterswillnothavetheirvaluechanged(constantscanbeheldinmemoryjustasvariablesare)lestsubsequentusagesofthatconstant(madeviareferencetoitsmemorylocation)goawry,acommontechniqueisforthecompilertogeneratecodecopyingtheconstant'svalueintoatemporaryvariablewhoseaddressispassedtotheprocedure,andifitsvalueismodified,nomatter;itisnevercopiedbacktothelocationoftheconstant. Putanotherway,acarefullywrittentestprogramcanreportonwhetherparametersarepassedbyvalueorreference,andifused,whatsortofcopy-inandcopy-outscheme.However,variationisendless:simpleparametersmightbepassedbycopywhereaslargeaggregatessuchasarraysmightbepassedbyreference;simpleconstantssuchaszeromightbegeneratedbyspecialmachinecodes(suchasClear,orLoadZ)whilemorecomplexconstantsmightbestoredinmemorytaggedasread-onlywithanyattemptatmodifyingitresultinginimmediateprogramtermination,etc. Ingeneral[edit] Thisexampleisextremelysimple,althoughcomplicationsarealreadyapparent.Morelikelyitwillbeacaseofmanyprocedures,havingavarietyofdeducibleorprogrammer-declaredpropertiesthatmayenablethecompiler'soptimizationstofindsomeadvantage.Anyparametertoaproceduremightbereadonly,bewrittento,bebothreadandwrittento,orbeignoredaltogethergivingrisetoopportunitiessuchasconstantsnotneedingprotectionviatemporaryvariables,butwhathappensinanygiveninvocationmaywelldependonacomplexwebofconsiderations.Otherprocedures,especiallyfunction-likeprocedureswillhavecertainbehavioursthatinspecificinvocationsmayenablesomeworktobeavoided:forinstance,theGammafunction,ifinvokedwithanintegerparameter,couldbeconvertedtoacalculationinvolvingintegerfactorials. Somecomputerlanguagesenable(orevenrequire)assertionsastotheusageofparameters,andmightfurtheroffertheopportunitytodeclarethatvariableshavetheirvaluesrestrictedtosomeset(forinstance,6