Ownership and Immutability in Generic Java (OIGJ)

YoavZibin+, Alex Potanin* Paley Li*, Mahmood Ali^, and Michael Ernst$ Presenter: Yossi Gil+ +IBM *Victoria,NZ^MIT $Washington Ownership and Immutability in Generic Java (OIGJ)

Ownership + Immutability Our previous work OGJ: added Ownership to Java IGJ: added Immutability to Java This work OIGJ: combine Ownership + Immutability The sum is greater than its parts IGJ could not type-check existing code for creating immutable cyclic data-structures (e.g., lists, trees) We found a non-trivial connection between ownership and immutability

Contributions No refactoring of existing code Prototype implementation No syntax changes (uses type-annotations in Java 7) No runtime overhead Backward compatible Verified that Java’s collection classes are properly encapsulated (using few annotations) Flexibility OIGJ can type-check more code than previous work: cyclic structures, the factory and visitor design patterns Formalization Formalized the concepts of raw/cooked immutable objects and wildcards as owner parameters Proved soundness

Problem 1: Representation exposure Internal representation leaks to the outside private doesn’t offer real protection! Real life example! class Class { private List signers; public List getSigners() { return this.signers; } } Forgot to copy signers! • http://java.sun.com/security/getSigners.html • Bug: the system thinks that code signed by one identity is signed by a different identity

Solution for Representation Exposure Ownership! Class should own the list signers No outside alias can exist Ownership can be nested: note the tree structure Class X signers X … entry1 entry2 entryN X … elem1 elem2 elemN

Ownership: Owner-as-dominator Dominators in graph theory Given: a directed rooted graph X dominates Y if any path from the root to Y passes X Owner-as-dominator Object graph; roots are the static variables An object cannot leak outside its owner, i.e., Any path from a root to an object passes its owner Conclusion: No aliases to internal state

Problem 2: Unintended Modification Modification is not explicit in the language can Map.get() modify the map? for (Object key : map.keySet()) { map.get(key); }throwsConcurrentModificationExceptionfor the following mapnew LinkedHashMap(100, 1, true) Reorders elements according to last-accessed (like a cache)

Solution: Immutability Varieties of Immutability Class immutability (like String or Integer in Java) Object immutability The same class may have both mutable and immutable instances Reference immutability A particular reference cannot be used to mutate its referent (but other aliases might cause mutations) Example in IGJ syntax class Student { @Immutable Date dateOfBirth; … void setTutor(@ReadOnly Student tutor) @Mutable { … } } Method may modify the this object

Objects vs. References Objects mutable or immutable Creation of an immutable object Raw state: Fields can be assigned Cooked state: Fields cannot be assigned References mutable, immutable, or readonly

Challenge: Cyclic Immutability • Cooking a cyclic data-structure is complicated • Many objects must be raw simultaneously to manipulate backward pointers • Then everything must become immutable simultaneously • OIGJ’s novel idea: • Prolong the cooking phase by using ownership information • Enables creation of immutable cyclic structures

Cooking immutable objects Previous work An object becomes cooked when its constructor finishes OIGJ’s observation An object becomes cooked when its owner’s constructor finishes The outside world will not see this cooking phase The complex object with its representation becomes immutable simultenously

Cooking LinkedList (1 of 2) No refactoring – the original code must compile in OIGJ Sun’s code is similar 1 : LinkedList(Collection<E> c){ 2 : this(); 3 : Entry<E> succ = this.header, pred = succ.prev; 4 : for (E e : c) { 5 : Entry<E> entry = 6 : new Entry<E>(e,succ,pred); 7 : // An entry is modified after it’s constructor finished 8 : pred.next = entry; pred = entry; 9 : } 10: succ.prev= pred; 11: }

Cooking LinkedList (2 of 2) The list owns its entries Therefore, it can mutate them, even after their constructor finished Sun’s code is similar 1 : LinkedList(@ReadOnlyCollection<E> c) @Raw { 2 : this(); 3 : @This @I Entry<E> succ = this.header, pred = succ.prev; 4 : for (E e : c) { 5 : @This @I Entry<E> entry = 6 : new @This @I Entry<E>(e,succ,pred); 7 : // An entry is modified after it’s constructor finished 8 : pred.next = entry; pred = entry; 9 : } 10: succ.prev= pred; 11: } Code in OIGJ; Annotations next slide.

Hierarchies in OIGJ ReadOnly World Raw Immut This Mutable Ownership hierarchy World – anyone can access This – this owns the object Immutability hierarchy ReadOnly– no modification Raw– object under construction

OIGJ syntax: fields (1 of 2) Two annotations per type 1:class Foo{ 2: // An immutable reference to an immutable date. @O @ImmutDate imD = new @O @ImmutDate (); 3: // A mutable reference to a mutable date. @O @Mutable Date mutD = new @O @Mutable Date(); 4: // A readonly reference to any date. Both roD and imD cannot mutate // their referent, however the referent of roD might be mutated by an // alias, whereas the referent of imD is immutable. @O @ReadOnlyDate roD= ... ? imD : mutD; 5: //A date with the same owner and immutability asthis @O @I Date sameD; 6: // A date owned bythis; it cannot leak. @This @I Date ownedD; 7: // Anyone can access this date. @World @I Date publicD;

OIGJ syntax: methods (2 of 2) Method receiver’s annotation has a dual purpose: Determines if the method is applicable. Inside the method, the bound of @Iis the annotation. 8 : // Can be called on any receiver; cannot mutate this. intreadonlyMethod() @ReadOnly{...} 9 : // Can be called only on mutable receivers; can mutate this. void mutatingMethod() @Mutable {...} 10: // Constructor that can create (im)mutable objects. Foo(@O @I Date d) @Raw { 11: this.sameD = d; 12: this.ownedD = new @This @I Date (); 13: // Illegal, because sameDcame from the outside. // this.sameD.setTime(...); 14: // OK, because Rawis transitive for owned fields. this.ownedD.setTime(...); 15: }

Formalization: Featherweight OIGJ • Novel idea: Cookers • Every location l in the heap is of the form: • l' is the owner of l • l” is the cooker of l, i.e., when the constructor of l” finishes then l becomes cooked • We keep track of the set of ongoing constructors • Subtyping rules connect cookers and owners or lFoo<l’,Mutable> lFoo<l’,Immut > l" • Proved soundness and type preservation

Case studies • Implementation uses the checkers framework • Only 1600 lines of code (but still a prototype) • Requires type annotations available in Java 7 • Java’s Collections case study • 77 classes, 33K lines of code • 85 ownership-related annotations • 46 immutability-related annotations

Case studies conclusions • Verified that collections own their representation • Method clone is problematic • Clone makes a shallow copy that breaks ownership • Our suggestion: compiler-generated clone that nullifies fields, and then calls a copy-constructor

Previous Work Universes Relaxed owner-as-dominator to owner-as-modifier ReadOnly references can be freely shared Constrains modification instead of aliasing, i.e., only the owner can modify an object Reference immutability: C++’s const Javari

Future work Inferring ownership and immutability annotations Bigger case study Extending OIGJ owner-as-modifier uniqueness or external uniqueness

Conclusions • Ownership Immutability Generic Java (OIGJ) • Simple, intuitive, small • Static – no runtime penalties (like generics) • Backward compatible, no JVM changes • Case study proving usefulness • Formal proof of soundness • Paper submitted to OOPSLA. Links: • http://ecs.victoria.ac.nz/twiki/pub/Main/TechnicalReportSeries/ • http://code.google.com/p/checker-framework/ • http://code.google.com/p/ownership-immutability/

OIGJ typing rules Ownership nesting Field access Field assignment Method invocation Method guards Raw parameter can be used only in method guards (see the paper for all rules, such as inner classes, covariant subtyping)

Ownership example in OGJ This-owned fields/methods can be accessed only via this class Class { @This List signers; // This-owned field public @This List getSigners1() { return this.signers; } public @World List getSigners2() { return new @World LinkedList(this.signers); } public void example(Class other) { this.signers = …; // Ok other.signers = …; // Illegal this.getSigners1(); // Ok other.getSigners1(); // Illegal other.getSigners2(); // Ok }

OIGJ syntax: fields (1 of 2) Two new generic parameters were added 1:class Foo<O extends World, I extends ReadOnly> { 2: // An immutable reference to an immutable date. Date<O,Immut>imD = new Date<O,Immut>(); 3: // A mutable reference to a mutable date. Date<O,Mutable>mutD = new Date<O,Mutable>(); 4: // A readonly reference to any date. Both roD and imD cannot mutate // their referent, however the referent of roD might be mutated by an // alias, whereas the referent of imD is immutable. Date<O,ReadOnly>roD = ... ? imD : mutD; 5: // A date with the same owner and immutability as this Date<O,I>sameD; 6: // A date owned by this; it cannot leak. Date<This,I>ownedD; 7: // Anyone can access this date. Date<World,I>publicD;

OIGJ syntax: methods (2 of 2) Method guard <T extends U>?has a dual purpose: The method is included only if T extends U Inside the method, the bound of Tis U. 8 : // Can be called on any receiver; cannot mutate this. ? intreadonlyMethod(){...} 9 : // Can be called only on mutable receivers; can mutate this. ? void mutatingMethod(){...} 10: // Constructor that can create (im)mutable objects. ? Foo(Date<O,I> d) { 11: this.sameD = d; 12: this.ownedD = new Date<This,I>(); 13: // Illegal, because sameDcame from the outside. // this.sameD.setTime(...); 14: // OK, because Raw is transitive for owned fields. this.ownedD.setTime(...); 15: }

LinkedList Example 1 : class Entry<O,I,E> { 2 : E element; 3 : Entry<O,I,E> next, prev; … 4 : } 5 : class LinkedList<O,I,E> { 6 : Entry<This,I,E> header; … 7 : ?LinkedList( 8 : Collection<?,ReadOnly,E> c) { 9 : this(); this.addAll(c); 10: } 11: ? void addAll( 12: Collection<?,ReadOnly,E> c) { 13: Entry<This,I,E> succ = this.header, 14: pred = succ.prev; 15: for (E e : c) { 16: Entry<This,I,E> en = 17: new Entry<This,I,E>(e,succ,pred); 18: pred.next = en; pred = en; } 19: succ.prev = pred; 20: }

Ownership nesting The main owner parameter must be inside any other owner parameter. List<This, I,Date<World,I>> l1; // Legal nesting List<World,I,Date<This, I>> l2; // Illegal! It’s illegal because we can store l2 in this variable: public static Object<World,ReadOnly> alias_l2;

Field access/assignment this-owned fields can be accessed/assigned only via this 1: class Foo<O extends World, I extends ReadOnly> { 2: Date<This,I> ownedD; // this-owned field 3: Date<O,I> sameD; 4: ? void bar(Foo<This,I> other) { 5: this.ownedD = …; // Legal: assign via this 6: other.ownedD = …; // Illegal: not via this 7: other.sameD = …; // Legal: not this-owned 8: }

Field assignment A field can be assigned only if the object is Raw or Mutable. If it is Raw, then the object must be this or this-owned. 1 : class Foo<O extends World, I extends ReadOnly> { 2 : Date<O,I> sameD; 3 : ? void bar( 4 : Foo<?,Mutable> mutableFoo, 5 : Foo<?,ReadOnly> readonlyFoo, 6 : Foo<?,I> rawFoo1, 7 : Foo<This,I> rawFoo2) { 8 : mutableFoo.sameD = …; // Legal: object is Mutable 9 : readonlyFoo.sameD = …; // Illegal: object is not Raw nor Mutable 10: rawFoo1.sameD = …; // Illegal: object is not this nor this-owned 11: rawFoo2.sameD = …; // Legal: object is Raw and this-owned 12: this.sameD = …; // Legal: object is Raw and this 13: }

Method invocation Method invocation is the same as field access/assignment: If any parameter is this-owned, then the receiver must be this. If the guard is Raw and the object is Raw, then the receiver must be this or this-owned. 1 : class Foo<O extends World, I extends ReadOnly> { 2 : Date<This,I> m1() { … } // Parameter is this-owned 3 : ? void m2() { … } 4 : ? void bar( 5 : Foo<?,I> rawFoo1, 6 : Foo<This,I> rawFoo2) { 7 : this.m1(); // Legal: object is this 8 : rawFoo2.m1(); // Illegal: object is not this 9 : rawFoo1.m2(); // Illegal: object is not this nor this-owned 10: rawFoo2.m2(); // Legal: both Raw and object is this-owned 11: this.m2(); // Legal: both Raw and object is this 12: }

Method guards Guard “<T extends U>?” has a dual purpose: The receiver’s T must be a subtype of U. Inside the method, the bound of T is U. 1: class Foo<O extends World, I extends ReadOnly> { 2: ? void rawM() { … } 3: ? void bar( 4: Foo<?,ReadOnly> readonlyFoo, 5: Foo<?,I> mutableFoo) { 6: readonlyFoo.rawM(); // Illegal: ReadOnly is not a subtype of Raw // The bound of I in this method is Mutable 7: mutableFoo.rawM(); // Legal: Mutableis a subtype of Raw 8: this.rawM(); // Legal: Mutable is a subtype of Raw 9: } Conditional Java (cJ) proposed method guards for Java

Ownership and Immutability in Generic Java (OIGJ)