intg_1=1;intg_2=0;Thread1Thread2intfunc_1(void){intl;for(l=0;(l!=4);l+ - PDF document

intg_1=1;intg_2=0;Thread1Thread2intfunc_1(void){intl;for(l=0;(l!=4);l++){if(g_1)returnl;�for(g_2=0;(g_2=26);++g_2);}}voidfunc_2(void){g_2=42;printf("%d",&g_2);}Shouldalwaysprint42asfunc_1doesnotaectg_2. GCC4.7.0-O2:theabovecodesometimesprints0! 2/28 CcodeAssemblygeneratedintfunc_1(void){intl;for(l=0;(l!=4);l++){if(g_1)returnl;�for(g_2=0;(g_2=26);++g_2);}}func_1:movlg_1(%rip),%edxmovlg_2(%rip),%eaxtestl%edx,%edxjne.L2movl$0,g_2(%rip)ret.L2:movl%eax,g_2(%rip)xorl%eax,%eaxret GCCsavesandrestoresg_2,sometimesrestoringitwronglyto0 3/28 Thread1Thread2func_1:movlg_1(%rip),%edxmovlg_2(%rip),%eaxtestl%edx,%edxjne.L2movl$0,g_2(%rip)ret.L2:movl%eax,g_2(%rip)xorl%eax,%eaxretvoidfunc_2(void)g_2=42;printf("%d",&g_2); I1:savesg_2=0in%eax I2:g_2=42; I1:restores%eaxtog_2 I2:printsg_2 4/28 Thread1Thread2func_1:movlg_1(%rip),%edxmovlg_2(%rip),%eaxtestl%edx,%edxjne.L2movl$0,g_2(%rip)ret.L2:movl%eax,g_2(%rip)xorl%eax,%eaxretvoidfunc_2(void)g_2=42;printf("%d",&g_2); I1:savesg_2=0in%eax I2:g_2=42; I1:restores%eaxtog_2 I2:printsg_2 4/28 Thread1Thread2func_1:movlg_1(%rip),%edxmovlg_2(%rip),%eaxtestl%edx,%edxjne.L2movl$0,g_2(%rip)ret.L2:movl%eax,g_2(%rip)xorl%eax,%eaxretvoidfunc_2(void)g_2=42;printf("%d",&g_2); I1:savesg_2=0in%eax I2:g_2=42; I1:restores%eaxtog_2 I2:printsg_2 4/28 CompilertestingStateoftheartofdierentialcompilertesting(Regehretal.):randomwell-denedprogramGCCvv CLang(( executableexecute executableexecute nalstate? nalstate Hardonconcurrentprograms:IRandomnon-racyinterestingprogramsIGettingallbehaviors:non-determinismIFinalstateequivalenceisnotenoughanymore 6/28 KeyremarksGCCuseseparatecompilation,soitmusttreateverysequentialfunctionasifitmaybeusedinaconcurrentprogram.Itispossibletodescribewhichoptimisationsarecorrectinanynon-racyconcurrentcontext.7/28 ContributionsAsoundnessproofforseveraltransformationsofC11executions(alongthelinesofSevcik'sworkonanidealisedDRFmodel)cmmtest:atoolforrandomdierentialtestingofcompilersforconcurrencybugs.9/28 Whatisanoptimisation?y=42inthisexample.for(inti=0;i2;i++){z=i;;x[i]+=y+1;}=)tmp=y+1;for(inti=0;i2;i++){z=i;x[i]+=tmp;}Wz0sb Ry42sb Wx[0]43sb Wz1sb Ry42sb Wx[1]43=)Ry42sb Wz0sb Wx[0]43sb Wz1sb Wx[1]4311/28 AtomicaccessesC11isDRF:IRacesareforbiddenIRace-freecode:traceisaninterleavingoftracesofthreadsEscapemechanism:\atomiclocations"IRacesareallowedontheseIAttributesonaccessesgivetheirsemantics13/28 Release-acquiremessagepassingInitiallyx=y=f=0inthisexample:Thread1 Thread2 x=1;f.store(1,release); while(0==f.load(acquire));y=x+1;Synchronisationbetweenareleasestoreandanacquireload.Soy=2attheend.15/28 Release/acquirepairs:denitionAnactionisareleaseifitisapossiblesourceofasynchronisation(unlock,releaseatomicwrite,scatomicwrite)Anactionisanacquireifitisapossibletargetofasynchronisation(lock,acquireatomicread,scatomicread)Asame-threadrelease-acquirepairisareleasefollowedbyanacquireinprogramorder(sb).16/28 Eliminations:criterionAnactionaisanOverwritten-Writeif:IaisawritetolIwisawritetolIaandwinthesamethread,withasbw(sbisinstructionorderinathread)I@(rel;acq):asbrelsbacqsbwwithrelareleaseandacqanacquire Similardenitions(andsamecriterion)forRead-after-Read,Read-after-Write,Write-after-ReadandWrite-after-Write.WealsoprovedthesoundnessofeliminatingIrrelevant-Reads 17/28 Eliminations:Whythiscriterion? Thread1Thread2x=1;sb while(0==f 1.load(acquire));sb f 1.store(1,release); sb rf�!;sw;hb11 printf("%i",&x);sb while(0==f 2.load(acquire)); sb f 2.store(1,release);rf�!;sw;hboo x=2; Twosynchronisationsarerequiredforanon-racydiscriminatingcontext.18/28 ReorderingsLoop-invariantcodemotionforexamplerequirereorderingsinadditiontoeliminations.Criterion:IUnrelatednon-atomicactionscanbefreelyreorderedIActionscanbepushedincriticalsections,notout.I.e.theycangobelowanacquire,orabovearelease,butneverthecontrary(notproven).19/28 SoundoptimisationsCmmtestConclusionandfuturework 21/28 ResultsFourconcurrencybugsfoundinthecurrentversionofGCC.Writeintroductions,asintherstexample.InteractionwithGCCdevellopers:ITheywereusingcriteriauselesslyconservativeITheythoughtGCCwasnotdoinganyreorderingwithatomics...23/28 SummaryIWefoundcriteriaofsoundnessforcodetransformationsintheC++11memorymodelIWehaveacompleteproofoftheirsoundnessIAndbuiltatoolforautomaticallyndingwhencompilersdoothertransformations26/28 Release-sequenceAWx1&#xREL0;sb;mo sw;hb!! AWx2&#xRLX0;sb;mo AWx3&#xRLX0; rf�!** ARx3¬Q0;28/28