Compilation 2007 Code Generation Michael I. Schwartzbach BRICS, University of Aarhus

2 Code Generation Code Generation Phases  Computing resources, such as: layout of data structures offsets register allocation  Generating an internal representation of machine code for statements and expressions  Optimizing the generated code (ignored for now)  Emitting the code to files in assembler format  Assembling the emitted code to binary format

3 Code Generation Joos Code Generation  Compute offsets and signatures  Generate code for static initializers  Generate code for statements and expressions  Optimize the generated code (ignored for now)  Compute locals and stack limits  Emit Jasmin code  Assemble Jasmin code to class files

4 Code Generation Computing Offsets  Each formal and local variables must have an offset in the stack frame  The this object always has offset 0  The naive solution: enumerate all formals and locals  The better solution: reuse offsets for locals in disjoint scopes  The clever solution: exploit liveness information must still respect the runtime types of locals

5 Code Generation Naive Offsets public void m(int p, int q, Object r ) int x = 42; int y ; { int z ; z = 87; } { boolean a ; Object b ; { boolean b ; int z ; b = true; boolean c ; c = b && (x==87); } { int y ; y = x; } max = 12

6 Code Generation Better Offsets public void m(int p, int q, Object r ) int x = 42; int y ; { int z ; z = 87; } { boolean a ; Object b ; { boolean b ; int z ; b = true; boolean c ; c = b && (x==87); } { int y ; y = x; } max = 10

7 Code Generation Clever Offsets public void m(int p, int q, Object r ) int x = 42; int y ; { int z ; z = 87; } { boolean a ; Object b ; { boolean b ; int z ; b = true; boolean c ; c = b && (x==87); } { int y ; y = x; } max = 6

8 Code Generation  The function sig(σ) encodes a type: sig( void ) = V sig( byte ) = B sig( short ) = S sig( int ) = I sig( char ) = C sig( boolean ) = Z sig(σ [] ) = [ desc(σ) sig(C 1. C C k ) = C 1 / C 2 /... / C k desc( void ) = V desc( byte ) = B desc( short ) = S desc( int ) = I desc( char ) = C desc( boolean ) = Z desc(σ [] ) = [ desc(σ) desc(C 1. C C k ) = L C 1 / C 2 /... / C k ; Computing Signatures (1/2)

9 Code Generation Computing Signatures (2/2)  This extends to fields, methods, and constructors  The field named x in class C: sig(C) / x  The method σ m ( σ 1 x 1,..., σ k x k ) in class C: sig(C) / m ( desc(σ 1 )...desc(σ k ) ) desc(σ)  The constructor C ( σ 1 x 1,..., σ k x k ) in class C: sig(C) / ( desc(σ 1 )...desc(σ k ) )V

10 Code Generation Static Initializers (1/2)  Initialization of static fields is performed when the class is loaded by the JVM  All static fields are first given default values  The code for initialization is written in a special method with the name  Fields that are static final and constant valued must then be initialized  Finally, all other static fields are initialized

11 Code Generation Static Initializers (2/2) public class A { public static int x = A.y+1; public static final int y = 42; public static void main(String[] args) { System.out.print(A.x); } public class A { public static int x = A.y+1; public static int y = 42; public static void main(String[] args) { System.out.print(A.x); } public class A { public static int x = A.y+1; public static final int y = A.fortytwo(); public static int fortytwo() { return 42; } public static void main(String[] args) { System.out.print(A.x); }

12 Code Generation Generating Code  Each statement and expression generates a sequence of bytecodes  A code template shows how to generate bytecodes for a given language construct  The template ignores the surrounding context  This yields a simple, recursive strategy for the code generation

13 Code Generation Code Template Invariants  A statement and a void expression leaves the stack height unchanged  A non- void expression increases the stack height by one  This is a local property of each template  The generated code must be verifiable  This is not a local property, since the verifier performs a global static analysis

14 Code Generation Code Templates (1/12) if( E ) S E ifeq false S false: if( E ) S 1 else S 2 E ifeq false S 1 goto endif false: S 2 endif: nop while( E ) S goto cond: loop: S cond: E ifne loop while(true) S loop: S goto loop

15 Code Generation { σ n ; S } S Code Templates (2/12) E ; E { σ n = E ; S } E σ store offset(n) S type(E) = void E ; E pop type(E) ≠ void throw E ; E athrow return E ; E σ return return; return σ store is either istore or astore depending on σ

16 Code Generation Code Templates (3/12) new C ( E 1,..., E k δ new sig(C) dup E 1... E k invokespecial sig(δ) this( E 1,..., E k δ aload 0 E 1... E k invokespecial δ indicates that δ is the corresponding resolved declaration

17 Code Generation Code Templates (4/12) super( E 1,..., E k δ aload 0 E 1... E k invokespecial sig(δ) aload 0 I 1 putfield sig(x 1 ) desc(σ 1 )... aload 0 I n putfield sig(x n ) desc(σ n ) The current class contains the non-static field initializations: σ 1 x 1 = I 1 ;... σ n x n = I n ;

18 Code Generation Code Templates (5/12) E. m ( E 1,..., E k δ E E 1... E k invokevirtual sig(δ) E. m ( E 1,..., E k δ E E 1... E k invokeinterface sig(δ) C. m ( E 1,..., E k δ E 1... E k invokestatic sig(δ) sig(δ) is an interface sig(δ) is a class

19 Code Generation (char) E E i2c E instanceof CE instanceof sig(C) Code Templates (6/12) ( C ) E E checkcast sig(C)

20 Code Generation Code Templates (7/12) thisaload 0 n σ load offset(n) E.f E getfield sig(f) desc(type(E.f)) E 1 [ E 2 ] E 1 E 2 σ aload type(E 1 [ E 2 ] ) = σ type(n) = σ C.f getstatic sig(f) desc(type(C.f)) σ aload is either iaload, baload, saload, caload, or aaload depending on σ

21 Code Generation Code Templates (8/12) n = E E dup σ store offset(n) E 1. f = E 2 E 1 E 2 dup_x1 putfield sig(f) desc(type(E 1.f)) E 1 [ E 2 ] = E 3 E 1 E 2 E 3 dup_x2 σ astore type(E 1 [ E 2 ] ) = σ type(n) = σ C. f = E E dup putstatic sig(f) desc(type(C.f))

22 Code Generation Code Templates (9/12) new σ [ E ] E multianewarray desc(σ) 1 E.length E arraylength E.clone() E invokevirtual sig(type(E)) /clone()Ljava/lang/Object;

23 Code Generation Code Templates (10/12) 42 ldc_int 42 true ldc_int 1 null aconst_null "abc" ldc_string "abc"

24 Code Generation Code Templates (11/12) E 1 + E 2 E 1 E 2 iadd type(E 1 + E 2 ) = int E 1 + E 2 E 1 E 2 invokevirtual S /concat(L S ;)L S ; type(E 1 + E 2 ) = String S  java/lang/String - E E ineg

25 Code Generation Code Templates (12/12) E 1 || E 2 E 1 dup ifne firsttrue pop E 2 firsttrue: E 1 && E 2 E 1 dup ifeq firstfalse pop E 2 firstfalse:

26 Code Generation Stack and Locals Limits  The generated code must explicitly state: the maximal number of local and formal offsets the maximal local stack height  This is used to determine the size of the frame  The locals limit is the maximal offset + (1 or 0)  The stack limit is computed by a static analysis

27 Code Generation Stack Limit Analysis  Consider the control flow graph of the bytecodes  succ(S i ) denotes the set of successor bytecodes  Δ(S i ) denotes the change in stack height by S i  S 0 denotes the first bytecode  For every bytecode S i we define the following integer-valued properties: B[[S i ]] denotes the stack height before S i A[[S i ]] denotes the stack height after S i

28 Code Generation Dataflow Constraints  B[[S 0 ]] = 0  A[[S i ]] = B[[S i ]] + Δ(S i )   x  succ(S i ): A[[S i ]] = B[[x]]  A[[S i ]]  0  These constraints must have a solution  The stack limit is the largest value of any A[[S i ]]

29 Code Generation Jasmin Class Format (1/3)  The overall structure of a Jasmin file is:.source sourcefile.class modifiers name.super sig( superclass ).implements sig( interface ).field modifiers desc( type ) constructors methods

30 Code Generation Jasmin Class Format (2/3)  The structure of a constructor is:.method modifiers sig( constructor ).throws sig( exception ).limit stack stacklimit.limit locals localslimit bytecodes.end method

31 Code Generation Jasmin Class Format (3/3)  The structure of a method is:.method modifiers sig( method ).throws sig( exception ).limit stack stacklimit.limit locals localslimit bytecodes.end method

32 Code Generation A Tiny Class public class Foo { public int y = 42; public Foo(int z) { y = y+z; } public String print(int n) { if (n==0) return new Integer(y).toString(); else return new Foo(y).print(n-1); }

33 Code Generation The Generated Code.source Foo.java.class public Foo.super java/lang/Object.field public "y" I.method public (I)V.limit stack 3.limit locals 2 aload_0 invokespecial java/lang/Object/ ()V aload_0 bipush 42 putfield Foo/y I aload_0 getfield Foo/y I iload_1 iadd dup_x1 putfield Foo/y I pop return.end method.method public print(I)Ljava/lang/String;.limit stack 3.limit locals 2 iload_1 iconst_0 if_icmpeq true2 iconst_0 goto end3 true2: iconst_1 end3: ifeq false0 new java/lang/Integer dup aload_0 getfield Foo/y I invokespecial java/lang/Integer/ (I)V invokevirtual java/lang/Integer/toString()Ljava/lang/String; areturn false0: new Foo dup aload_0 getfield Foo/y I invokespecial Foo/ (I)V iload_1 iconst_1 isub invokevirtual Foo/print(I)Ljava/lang/String; areturn.end method

34 Code Generation The Binary Class File cafe babe e 001e 0c00 0a a f6c 616e 672f 496e a f6c 616e 672f 4f62 6a c69 6e69 743e c d f6f e 740c c 0a00 1d f a53 6f c a a c f e67 0c00 0f f 6f2e 6a c6a f 6c61 6e67 2f e 673b 0a a 001d 000b c 6a f6c 616e 672f e67 3b f e ab a10 2ab a2a b b 605a b b a e a e1b 039f a bb00 1d59 2ab b700 0cb6 001b b0bb a b400 04b7 001a 1b04 64b b