course news - university of washingtoncourses.cs.washington.edu/courses/csep590a/17sp/...dan...

5/16/17

1

CSE341: Programming Languages

Lecture 1 Course Mechanics

ML Variable Bindings

Dan Grossman Fall 2011





CSEP590–ProgrammingSystemsUniversityofWashington

Lecture6:Potpourri

MichaelRingenburgSpring2017









CourseNews•  PresentaJons

–  Startnextweek!–  Schedulepostedoncourseweb

•  Today–mixofinteresJngtopicsthatwehaven’tcoveredyet–  TypeChecking–  LoopParallelism–  JVMandJITcompilaJon–  QueryopJmizaJon,ifJmepermits

•  FinalHomeworkpostedtoday–  Dueendofquarter(June2)

Spring2017 UWCSEP590(PMPProgrammingSystems):Ringenburg 1

5/16/17

2










TypesfromtheCompiler’sPerspec3ve









Types

•  Typesplayakeyroleinmostprogramminglanguages.E.g.,– Run-Jmesafety– Compile-JmeerrordetecJon–  Improvedexpressiveness(inheritance,overloading,etc)

– ProvideinformaJontoopJmizer•  Stronglytypedlanguages–whatdatamightbeusedwhere

•  Typequalifiers(e.g.,constandrestrictinC)


5/16/17

3









TypeCheckingTerminology

StaJcvs.dynamictyping•staJc:checkingdonepriortoexecuJon(e.g.compile-Jme)•dynamic:checkingduringexecuJon

Strongvs.weaktyping•strong:guaranteesnoillegaloperaJonsperformed•weak:can’tmakeguarantees

static dynamic

strong Java, ML Scheme, Ruby

weak C PERL










TypeSystems

•  BaseTypes– Fundamental,atomictypes– Typicalexamples:int,double,char,bool

•  Compound/ConstructedTypes– Builtupfromothertypes(recursively)viaconstructors

– Constructorsincludearrays,records/structs/classes,pointers,enumeraJons,funcJons,modules,…


5/16/17

4









TypesvsASTs•  TypesarenottypicallyASTnodes

– ASTnodesofenhaveatypefield,however•  AST=abstractrepresentaJonofsourceprogram(includingsourceprogramtypeinfo)

•  Types=abstractrepresentaJonoftypesemanJcsfortypechecking,inference,etc.–  CanincludeinformaJonnotexplicitlyrepresentedinthesourcecode,ormaydescribetypesinwaysmoreconvenientforprocessing

•  Needaseparate“type”classhierarchyinyourcompilerdisJnctfromtheAST










BaseTypes

•  Foreachbasetype(int,boolean,etc),cancreateasingleobjecttorepresentit–  BasetypesinsymboltableentriesandASTnodesaredirectreferencestotheseobjects

–  Basetypeobjectsusuallycreatedatcompilerstartup•  Usefultocreateatype“void”objecttotagfuncJonsthatdonotreturnavalue

•  Alsousefultocreateatype“unknown”objectforerrors–  (“void”and“unknown”typesreducetheneedforspecialcasecodeinvariousplacesinthetypechecker)


5/16/17

5









CompoundTypes

•  Basicidea:useaappropriate“typeconstructor”objectthatreferstothecomponenttypes– Limitednumberofthese–corresponddirectlytotypeconstructorsinthelanguage(record/struct/class,array,funcJon,…)










ArrayTypes•  ForregularJavathisissimple:onlypossibilityis#ofdimensionsandelementtypeclassArrayTypeextendsType{ intnDims; TypeelementType;}

•  Lengthnotpartoftype•  MoreinteresJnginlanguageslikePascal(morecomplexarrayindexing–indextypes!)


5/16/17

6









Methods/FuncJons

•  Typeofamethodisitsresulttypeplusanorderedlistofparametertypes

classMethodTypeextendsType{ TyperesultType; //typeor“void” ListparameterTypes;}










ClassTypes

•  Typefor:classId{fieldsandmethods}classClassTypeextendsType{TypebaseClassType;//reftobaseclassMapfields; //typeinfoforfieldsMapmethods; //typeinfoformethods(later)

}–  Baseclasspointer,sowecancheckfieldreferencesagainstbaseclass

ifwedon’tfindinthisclass.–  (Note:maynotwanttodothisliterally,dependingonhowclass

symboltablesarerepresented;i.e.,classsymboltablesmightbeusefulorsufficientastherepresentaJonoftheclasstype.)


5/16/17

7









TypeEquivalance

•  Forbasetypesthisissimple–  Ifyouhavejustasingleinstanceofeachbasetype(asrecommend),thentypesarethesameifandonlyiftheyareidenJcal

•  Pointer/referencecomparisoninthetypechecker– NormallytherearewelldefinedrulesforcoercionsbetweenarithmeJctypes

•  Dependingonlanguagerules,compilerinsertstheseautomaJcallyorwhenrequestedbyprogrammer(casts)–ofeninvolvesinserJngcast/conversionnodesinAST










TypeEquivalenceforCompoundTypes

•  Twobasicstrategies–  Structuralequivalence:twotypesarethesameiftheyarethesamekindoftypeandtheircomponenttypesareequivalent,recursively

•  E.g.,twostructtypes,eachwithexactlytwointfields– Nameequivalence:twotypesarethesameonlyiftheyhavethesamename.Iftheirstructuresmatch,buthavedisJnctnames,theyarenotequal.

•  Differentlanguagedesignphilosophies– Mixiscommon,e.g.,C/C++nameforstructs/classes,structuralotherwise.


5/16/17

8









StructuralEquivalence

•  Structuralequivalencesaystwotypesareequalifftheyhavesamestructure–  IdenJcalbasetypesclearlyhavethesamestructure–  iftypeconstructors:

•  sameconstructor•  recursively,equivalentargumentstoconstructor

•  Ex:atomictypes,arraytypes,pointertypes•  ImplementwithrecursiveimplementaJonofequals










NameEquivalence

•  Nameequivalencesaysthattwotypesareequalifftheycamefromthesametextualoccurrenceofatypeconstructor–  Ex:classtypes,Cstructtypes(structtagname),datatypesinML

–  specialcase:typesynonyms(e.g.typedefinC)donotdefinenewtypes–usesstructuralequivalence

•  Implementwithpointer/referenceequalityassumingappropriaterepresentaJonoftypeinfo


5/16/17

9









TypeEquivalenceandInheritance

•  SupposewehaveclassBase{…}classExtendedextendsBase{…}

•  AvariabledeclaredwithtypeBasehasacompile-6metypeofBase

•  DuringexecuJon,thatvariablemayrefertoanobjectofclassBaseoranyofitssubclasseslikeExtended(orcanbenull)–  SomeJmescalledtherun6metype–  Subclassesguaranteedtohaveallfields/methodsofbaseclass,sotypecheckingasbaseclasssuffices










TypeCasts

•  Inmostlanguages,onecanexplicitlycastanobjectofonetypetoanother– someJmescastmeansaconversion(e.g.,castsbetweennumerictypes)

– someJmescastmeansachangeofstaJctypewithoutdoinganycomputaJon(castsbetweenpointertypesorpointerandnumerictypes)

– Withclasstypes,mayalsomeanupcast(free)ordowncast(runJmecheck)s


5/16/17

10









TypeConversionsvsCoercions

•  InJava,wecanexplicitlyconvertanvalueoftypedoubletooneoftypeint– Canrepresentasunaryoperator– Typecheck,generatecodenormally

•  InJava,canimplicitlycoerceanvalueoftypeinttooneoftypedouble– Compilermustinsertunaryconversionoperators,basedonresultoftypechecking










CandJava:typecasts•  InC:safety/correctnessofcastsnotchecked

– AllowswriJnglow-levelcodethat’stype-unsafe–  ResultisofenimplementaJondependent/undefined.Notportable,butsomeJmesuseful.

•  InJava:downcastsfromsuperclasstosubclassneedrun-Jmechecktopreservetypesafety– Otherwise,mightusefield(orcallmethod)thatisnotpresentinsuperclass

–  StaJctypecheckerallowsthecast–  Codegeneratorintroducesrun-Jmecheck–  Java’smainformofdynamictypechecking


5/16/17

11









FinalNote:VariousNoJonsofEquivalance

•  ThereareusuallyseveralrelaJonsontypesthatcompilerneedstodealwith:– “isthesameas”– “isassignableto”– “issameorasubclassof”– “isconverJbleto”

•  Besuretocheckfortherightone(s)










UsefulCompilerFuncJons

•  CreateahandfulofmethodstodecidedifferentkindsoftypecompaJbility,e.g.:–  TypesareidenJcal–  Typet1isassignmentcompaJblewitht2–  ParameterlistiscompaJblewithtypesofexpressionsinthecall

•  Usualmodularityreasons:isolatesthesedecisionsinoneplaceandhidestheactualtyperepresentaJonfromtherestofthecompiler

•  ProbablybelongsinthesamepackagewiththetyperepresentaJonclasses(packagefordealingwithtypes)


5/16/17

12










LoopParallelism









Whatisloopparallelism?

•  Aserial(non-parallelized)loopconsistsofaseriesofiteraJonsthatrunoneataJme(inorder)onasinglethread.


y A “normal” (non-parallelized) loop consists of a series of iterations that run one at a time (in order) on a single thread.

for(i=0; i<N; i++) {foo[i] = bar[i];

}453721

000000

453721

bar foo



}453721

000000

453721

bar foo1

2

3

4

5

6

5/16/17

13









Whatisloopparallelism?

•  AparallelizedloopconsistsofaseriesofiteraJonsthatmayrunsimultaneouslyonmulJplethreads.

•  EverythreadexecutesadisJnctsubsetoftheiteraJons

•  IteraJonsnotordered.




}453721

000000

453721

bar foo



}453721

000000

453721

bar foo1

1

2

2

2

1









MoJvaJonsforLoopParallelizaJon

•  Takeadvantageofavailableparallelisminarchitecture– MulJ-coreandmany-coreprocessors– VectorizaJoninstrucJons– Hyperthreading

•  Latencyhiding– SwitchcontextsratherthanwaiJng


5/16/17

14









CondiJonsforParallelizaJon

•  TypicalnecessarycondiJonsforcompilertoauto-parallelizealoop–  ItcanfigureouthowtocomputethenumberofiteraJonspriortoexecuJngtheloop

–  ItcanprovethattherearenodependencesbetweeniteraJons

–  TherearenofuncJoncallswithunknownsideeffects(e.g.,output)

–  Theloophasasimplestructure(e.g.,nomulJpleexits)•  UsersmayinsertextrainformaJontohelpthecompilerestablishthatthesecondiJonshold.–  Compilerreliesoninfo:iffalse,compiledprogrammaybehaveunepectedly










Examples

•  Parallelizableloop:


void foo(int n) { int i; int my_array[n]; for (i = 0; i < n; i++) { my_array[i] = i; } return; }

5/16/17

15









Examples

•  Non-parallelizableloop:

•  aandbmaypointtooverlappingmemory:


void foo(int *a, int *b) { int i; for (i = 0; i < 10000; i++) { a[i] = b[i]; }

}

foo(x, x+5000)









HelpingtheCompiler

•  Commontypesofhints/informaJon–  Thispointerisnotaliasedwithanyotherpointers(doesn’tpointtodatathatoverlapswithanotherpointer).E.g.,restrictkeywordinC

–  TherearenodependenciesbetweeniteraJons(loop-carrieddependencies)introducedbythispointer

–  Trustme,thisloopcanbeparallelized•  OfenbeuertogivethecompilerinformaJonaboutwhyitissafetoparallelize(allowsmoreopJmizaJon–“trustmeit’ssafe”onlysayssafeaswriuen)


5/16/17

16









CompilerCanHelpItself,Too

•  Ofencompilerswillauempttorestructurecodetofindorenhanceparallelism

•  Somecommonexamples– Scalarexpansion– ReducJons– Loopcollapse










ScalarExpansion•  Thisloopcannotbeparallelizedaswriuenbecauseof

dependencesbetweenthereadsandwritesoftindifferentiteraJons(wriJngtinoneiteraJonmayoverwritethevalueoftfromanotheriteraJonbeforeitisused):

Spring2017 UWCSEP590(PMPProgrammingSystems):

Ringenburg 31

int t; for (i = 0; i < n; ++i) { t = sqrt(b[i]); ... a[i] = t + 5; }

5/16/17

17









ScalarExpansion•  Thisloopcannotbeparallelizedaswriuenbecauseof

dependencesbetweenthereadsandwritesoftindifferentiteraJons(wriJngtinoneiteraJonmayoverwritethevalueoftfromanotheriteraJonbeforeitisused):

•  CompilercansolvethisbyconverJngthescalarintegertintoanarrayofintegers.


int t[n]; for (i = 0; i < n; ++i) { t[i] = sqrt(b[i]); ... a[i] = t[i] + 5; }









ReducJons•  Compilersofenauempttorecognizeloopsthatcalculatesums,products,minimums,andmaximumsoveranarray.E.g.:

•  ThecompilercanconvertthesetoreducJons–  Eachthreadcomputesthemin/max/sum/productoverasub-secJonofthearray

–  Threadsthencombineresultstodeterminethefinalvalue(canusetree-structureforefficiency)


int min = MAX_VAL; for (i = 0; i < n; i++) { if (x[i] < min) min = x[i]; }

5/16/17

18









LoopCollapse

•  Howdowehandlenestedparallelloops?•  OpJon1:Goparallelfortheouterloop,andthenagainfor

theinnerloop.–  Inefficient–thereisasignificantoverheadtogoingparallel.Ifwenest,theneveryiteraJonoftheouterloophastopaythatoverhead.

–  MaylimiteffecJvenessoftheloadbalancingobtainedbysomeloopiteraJonschedulingmethods.


void foo(int* restrict num_bars, int size_x, int* restrict x, int* restrict bar) { for (int i = 0; i < size_x; i++) for (int j = 0; j < num_bars[i]; j++) x[i] += bar[i + j]; }









LoopCollapse

•  OpJon2:Loopcollapse.ConvertthenestedpairofparallelloopstoasingleparallelloopthatsimulatestheexecuJonofthenestedloops.

•  Manhauanstyle,wheninnerloopiteraJonsmayvary–  CreateanewparallellooptocalculatethetotalnumberofiteraJonsofthe

innerloop(acrossalliteraJonsoftheouterloop).–  ConvertthepairofloopsintoasingleloopwhereeachiteraJoncorresponds

toadisJnctouter/inneriteraJonpair.•  IfinnerloopiteraJonsarefixed,asimplerectangularcollapsesuffices

–  MiteraJonsnestedinNiteraJons=MxNcollapsedloopiteraJons•  BigperformancewinSpring2017 UWCSEP590(PMPProgrammingSystems):

Ringenburg 35

void foo(int* restrict num_bars, int size_x, int* restrict x, int* restrict bar) { for (int i = 0; i < size_x; i++) for (int j = 0; j < num_bars[i]; j++) x[i] += bar[i + j]; }

5/16/17

19









ManhauanCollapsePsuedocode


// t[i] = total # of inner loop iterations // in first i iterations of outer loop t[0] = 0; for (i = 0; i < size_x; i++) t[i + 1] = t[i] + num_bars[i]; for (k = 0; k < t[size_x]; k++) { // Set i to index of largest element of t // less than k (use binary search) // Optimization: Don’t recompute every time i = max_element_less_than(t, k); j = k - t[i]; x[i] += bar[i + j]; // original loop body }









Example


bool foo(int *a, int *b, int n, int sought, int *old_val) { int i; for (i = 0; i < n; i++) { if (b[i] == sought) break; a[i] = b[i]; } return (i < n); }

5/16/17

20









Example


1 X | for (i = 0; i < n; i++) { ** loop exit ** multiple exits 1 X | if (b[i] == sought) | break; 1 X | a[i] = b[i]; | }

•  CompilerfeedbacktoolsofenprovidedtotellyouwhereopJmizaJon/parallelizaJonoccurredordidn’t,andwhy.









Example


bool foo(int *a, int *b, int n, int sought, int *old_val) { int i; int found_index = n; for (i = 0; i < n; i++) { if (b[i] == sought) if (i < found_index) found_index = i; } for (int i = 0; i < found_index; i++) a[i] = b[i]; return (found_index < n); }

5/16/17

21









Example


| for (i = 0; i < n; i++) { 3 P:$| if (b[i] == sought) ** reduction moved out of 1 loop | if (i < found_index) | found_index = i; | } | for (int i = 0; i < found_index; i++) 4 S | a[i]=b[i];









Example


bool foo(int * restrict a, int *b, int n, int sought, int *old_val) { int i; int found_index = n; for (i = 0; i < n; i++) { if (b[i] == sought) if (i < found_index) found_index = i; } for (int i = 0; i < found_index; i++) a[i] = b[i]; return (found_index < n); }

5/16/17

22









Example


| for (i = 0; i < n; i++) { 3 P:$| if (b[i] == sought) ** reduction moved out of 1 loop | if (i < found_index) | found_index = i; | } | for (int i = 0; i < found_index; i++) 4 P | a[i]=b[i];










TheJVMandJIT

5/16/17

23









Java Implementation Overview

•  Java compiler (javac et al) produces machine-independent .class files – Target architecture is Java Virtual Machine

(JVM) – simple stack machine •  Java execution engine (java)

– Loads .class files (often from libraries) – Executes code

•  Either interprets stack machine code or compiles to native code (JIT)










JVM Architecture •  Abstract stack machine

–  Bytecodes pop operands, –  and push results

•  Implementation not required to use JVM specification literally –  Only requirement is that

execution of .class files has specified effect

–  Multiple implementation strategies depending on goals

•  Compilers vs interpreters •  Optimizing for servers vs

workstations


iconst_1

5/16/17

24















workstations


iconst_1

1















workstations


sipush100

1

5/16/17

25















workstations


100

sipush100

1















workstations


100

iadd

1

5/16/17

26















workstations


iadd

101









JVM Data Types

•  Primitive types – byte, short, int, long, char, float, double,

boolean

•  Reference types – Non-generic only (more on this later)


5/16/17

27









JVM Runtime Data Areas

•  Semantics defined by the JVM Specification – Implementer may do anything that preserves

these semantics •  Per-thread data

– pc register – Stack

•  Holds frames (details below) •  May be a real stack or may be heap allocated










JVM Runtime Data Areas

•  Per-VM data – shared by all threads – Heap – objects allocated here – Method area – per-class data

•  Runtime constant pool •  Field and method data •  Code for methods and constructors


5/16/17

28









Frames

•  Created when method invoked; destroyed when method completes

•  Allocated on stack of creating thread •  Contents

–  Local variables –  Operand stack for JVM instructions –  Reference to runtime constant pool

•  Symbolic data that supports dynamic linking

–  Anything else the implementer wants










JVM Instruction Set

•  Stack machine •  Byte stream •  Instruction format

– 1 byte opcode – 0 or more bytes of operands

•  Instructions encode type information – Verified when class loaded


5/16/17

29









Instruction Sampler

•  Load/store – Transfer values between local variables and

operand stack – Different opcodes for int, float, double,

addresses – Load, store, load immediate

•  Special encodings for load0, load1, load2, load3 to get compact code for first few local vars










Instruction Sampler

•  Arithmetic – Again, different opcodes for different types

•  byte, short, char & boolean use int instructions – Pop operands from operand stack, push result

onto operand stack – Add, subtract, multiply, divide, remainder,

negate, shift, and, or, increment, compare •  Stack management

– Pop, dup, swap


5/16/17

30









Instruction Sampler

•  Type conversion – Widening – int to long, float, double; long to

float, double, float to double – Narrowing – int to byte, short, char; double to

int, long, float, etc.










Instruction Sampler

•  Object creation & manipulation – New class instance – New array – Static field access – Array element access – Array length – Instanceof, checkcast


5/16/17

31









Instruction Sampler

•  Control transfer – Unconditional branch – goto – Conditional branch – ifeq, iflt, ifnull, etc. – Compound conditional branches - switch










Instruction Sampler

•  Method invocation –  invokevirtual –  invokeinterface –  invokespecial (constructors, superclass, private) –  invokestatic

•  Method return –  Typed value-returning instructions –  Return for void methods


5/16/17

32









BytecodeExample


outer:for(inti=2;i<1000;i++){for(intj=2;j<i;j++){if(i%j==0)con3nueouter;}System.out.println(i);}

0:iconst_21:istore_12:iload_13:sipush10006:if_icmpge449:iconst_210:istore_211:iload_212:iload_113:if_icmpge3116:iload_1…

17:iload_218:irem19:ifne2522:goto3825:iinc2,128:goto1131:getsta3c#84;System.out34:iload_135:invokevirtual#85;println38:iinc1,141:goto244:return









Execution Engines

•  Basic Choices – Interpret JVM bytecodes directly – Compile bytecodes to native code, which then

executes on the native processor •  Just-In-Time compiler (JIT)


5/16/17

33









JITLevels

•  C1:Fast,simplelight-weightopJmizaJons– Ofenusedforshortercodes(“client”mode)

•  C2:MoreaggressiveopJmizaJon,significantlyslower– Ofenusedforlongrunningcodes(“server”mode)

•  Butbothcauseoverheadwheninvoked










JITProfiling

•  JITcompilaJontypicallyprofile-driven–  CompilaJon(evenC1)hasacost–  CountexecuJonsofmethods,idenJfyhotloopnests–  JITonlyhotcode

•  TieredJIT– MulJplehot-nessthresholds– Uselight-weightJIT(C1)oncehitlowerthreshold–  Paycostofheavy-weightJIT(C2)ifhithigherthreshold


5/16/17

34









SpeculaJveJIT•  ProfilingcandomorethanjustidenJfycodeblocksexecuted

–  Certainbranchesalways/nevertaken–  Actualtypesused/methodimplementaJonscalled–  DetectifNULLpointersneverpassed–  Etc…

•  CanspeculaJvelyopJmizebasedonprofile–  Removeunusedbranches–  InlineparJcularimplementaJonsofvirtualmethods–  RemoveNULLchecks

•  MustdetectandbackoffifspeculaJonincorrect–  Detectvia“guards”,e.g.,checkingtypeorcondiJon,handlingSIGSEGV–  “DeopJmizaJon”…switchbacktointerpreter–  Switchingbackcanbeexpensive










SpeculaJonExample


o = receiver object ; x = receiver class (o) ; if ( x == expected-class ) { // virtual guard x.foo(a, b, c); // direct call can be inlined } else { o.foo(a, b, c); // guard failed, virtual call }

o.foo(a, b, b);

FromIBMJust-In-TimeCompiler(JIT)forJava:BestpracJcesandcodingguidelinesforimprovingperformance

5/16/17

35









Escape Analysis

•  Another optimization based on observation that many methods allocate local objects as temporaries

•  Idea: Compiler tries to prove that no reference to a locally allocated object can “escape” – Not stored in a global variable or object – Not passed as a parameter










Using Escape Analysis

•  If all references to an object are local, it doesn’t need to be allocated on the heap in the usual manner – Can allocate storage for it in local stack frame

•  Essentially zero cost

– Still need to preserve the semantics of new, constructor, etc.


5/16/17

36









OtherTechniques•  SaveprofileinformaJonfrompreviousexecuJons–  CanalsosaveJIT’edresults–  Eliminate“warm-up”–  Possiblybeuerprofileinfothanfakewarm-upssomeJmesemployed

•  AzulFalcon:recentlyannouncedJITusingLLVM–  Takeadvantageofpowerfulopensourcecompiler– MoreopJmizaJonspotenJallyavailable– MoreprocessorspecificopJmizaJons–e.g.,vectorizaJoninstrucJons











SQLQueryOp3miza3on(ExcerptsfromSparkSummitCatalystOp3mizerDeepDive,SparkSummit2016)

course news - university of washingtoncourses.cs.washington.edu/courses/csep590a/17sp/...dan...

Documents