小型JIT編譯器之最佳化技術(shù)評(píng)估指導(dǎo)教授:單智君老師指導(dǎo)學(xué)長:喬偉豪學(xué)長組員:鍾懿軒，蔣季融，李國丞小型JIT編譯器之最佳化技術(shù)評(píng)估指導(dǎo)教授:單智君老師1Outline觀察-----------------------------------鍾懿軒動(dòng)機(jī)-----------------------------------鍾懿軒目標(biāo)-----------------------------------鍾懿軒初步設(shè)計(jì)-----------------------------李國丞修改過的設(shè)計(jì)------------------------蔣季融未來進(jìn)度-----------------------------蔣季融Outline觀察---------------------2觀察Java是用stack運(yùn)算。因?yàn)榈讓訖C(jī)器是register-based非stack-based，用interpreter執(zhí)行，performance不佳。加入JIT將bytecode轉(zhuǎn)換為register-based的指令，可以增進(jìn)performance。若於JIT中加入一些機(jī)制，增進(jìn)其效能，則可進(jìn)一步增進(jìn)javaperformance。觀察Java是用stack運(yùn)算。3動(dòng)機(jī)加入的機(jī)制若能產(chǎn)生IR，使JITcompiler後端處理更方便。就有可能增進(jìn)JIT的速度。folding機(jī)制:讀入bytecode產(chǎn)生register-based的IR，簡便後段處理。因此決定於JIT中加入folding機(jī)制。動(dòng)機(jī)加入的機(jī)制若能產(chǎn)生IR，使JITcompiler後端處4目標(biāo)將folding機(jī)制加入JIT中，以最少的overhead，fold最多的pattern，達(dá)到增加performance的目的。目標(biāo)將folding機(jī)制加入JIT中，以最少的overhea5初步設(shè)計(jì)架構(gòu)圖Bytecode分類Pattern統(tǒng)計(jì)遇到的問題初步設(shè)計(jì)架構(gòu)圖6舊架構(gòu)圖舊架構(gòu)圖7舊folding架構(gòu)圖舊folding架構(gòu)圖8Bytecode分類1.定義:參考kim’spaper(AdvancedPOCModel-BasedJavaInstructionFoldingMechanism)P:非operation含有push。Op:operation含有push不含pop。Oc:operation含有pop不含push。C:非operation含有pop。On:不含pushpop,無法分類者。Bytecode分類1.定義:9Bytecode分類(續(xù))參考學(xué)長的paper(stackoperationsfoldinginjavaprocessors)P:transferingdatafromConstantRegisterorLocalVariabletotheoperandstack。O:getsdatafromtheoperandstackandthenperformsdifferenttasksOe:ALUtypeoperatorthatwritestheresultbacktotheoperandstack。Ob:Branchtypeoperator。Oc:Complextypeoperatorincludingarrayaccess,constantpoolaccess。Ot:unableorhardtojointhefoldingoperation。C:consumesdatafromtheoperandstack,andstoresdatabackintothelocalvariable。Bytecode分類(續(xù))參考學(xué)長的paper(stack10Bytecode分類(續(xù))2.統(tǒng)計(jì)bytecodepush/pop數(shù):檢視kvmsourcecode(bytecodes.c):查看每道bytecode的執(zhí)行碼中有多少push,pop的動(dòng)作。檢視jvmspec(chapter6)每道bytecode都已規(guī)定好stack情況，可直接紀(jì)錄。Bytecode分類(續(xù))11AnalyzepatternspossibilityinJavaclassfileGetinformationfromJavaclassfileDoitourself(inC++)BCELlibraryforJava

(/bcel)

Analyzepatternspossibilityi12Possibilityofpatterns(1)KIM所統(tǒng)計(jì)出的Patterns及出現(xiàn)百分比(2)自行統(tǒng)計(jì)”EmbeddedCaffeineMark”的結(jié)果

Possibilityofpatterns(1)KIM所13ProblemsClassificationcansavetimeofstringmatchingUsing“Hashing”isbetterthanstringmatching!Benefitofclassificationnolongerexistswhenusing“Hashing”!使用stringmatchingsequentialsearch:Toomuchfindingoverhead。ProblemsClassificationcansav14修改過的設(shè)計(jì)新架構(gòu)圖Folding方法修改過的設(shè)計(jì)新架構(gòu)圖15Structureof“JVMwithJIT”JavacodesourceJAVAcompilerJavabytecodeClassloaderBytecodeverifierHardwareOperatingSystemJavaclasslibrariesHotspotInterpretNativecodeJVMInterpreterJITYNStructureof“JVMwithJIT”Jav16Interactionbetween

Interpreter&JITTimeOnemethodHotspotDetectInterpretTimeComplicatedcodeJITEndofmethod(Startofmethod)CodeblockCodeblockInteractionbetween

Interpret17StrutureofourJITJITIRgeneratorFoldingCodegeneratorIR1IR2NativecodeMethodStrutureofourJITIRgenerato18方法(4-1)-OverviewO-oriented.

SearchbytecodesforO-typebytecodethenfindfoldingpatternforthisO-typebytecode.Datastructure:array.Table:Usedforstoringinformationofbytecodes.方法(4-1)-OverviewO-oriented.

Se19方法(4-2)-DataStructureTable:usebytecodeIDnumberasindex.Buffer:length=4;Bytecodeinformation:TypeP_num:PushnumberPostive,stackgrowC_num:PopnumberNegative,stackfall方法(4-2)-DataStructureTable:u20方法(4-3)-Algorithm經(jīng)由bytecode定義可得:P-type:C_num=0,P_num=1.C-type:P_num=0,C_num=1.O-type:不一定，視其功能而定。以O(shè)-type的attribute為尋找pattern的依據(jù)。P_num:向後尋找P_num個(gè)bytecodesC_num:向前尋找C_num個(gè)bytecodes方法(4-3)-Algorithm經(jīng)由bytecode定義可21方法(4-4)Match:向前找C_num個(gè)bytecodes的P_num剛好和O-type的C_num相抵銷。向後找P_num個(gè)bytecodes的C_num剛好和O-type的P_num相抵銷。整個(gè)pattern的attribute和要為零。無法match放棄此O-typebytecode，繼續(xù)找下一個(gè)。針對(duì)continuouspattern設(shè)計(jì)。方法(4-4)Match:22未來進(jìn)度TracingcodeModifyingcodeSimulationPerformance未來進(jìn)度Tracingcode23TracingcodeEnvironmentconfigurationTracecodeConcentrateon“HowtoaddFoldingintoJIT?”TracingcodeEnvironmentconfig24ModifyingcodeAddingFoldingtoJITModulizingFoldingandaddingitintoJITbetween“IRgenerator”and“Codegenerator”IRGeneratorCodeGeneratorBytecodeNativecodeFoldingModifyingcodeAddingFoldingt25SimulationRunbenchmarkonARMsimulatoronLinuxworkstationBenchmark:”EmbeddedCaffeineMark”SimulationRunbenchmarkonARM26PerformanceFindout“PerformanceSpeedup”afterusingfoldingSpeedup=

(time_with_folding)/