Summary

Use boxing to support language enhancements that treat primitive types more like reference types. This is a preview language and VM feature.

Goals

Allow boxing of primitive values when they are used as the "receiver" of a field access, method invocation, or method reference.

Allow unboxed return types when overriding a method with a reference-typed return.

Support primitive types as type arguments, implemented via boxing at the boundaries with generic code.

Support conversions between primitive- and reference-typed arrays, with automatic boxing and unboxing on reads and writes.

Motivation

Java's classes and interfaces provide an expressive mechanism to model data and associated operations. But the primitive types of the language—booleans, integers, and floating-point numbers—do not make use of this mechanism. Instead, they support a predetermined set of operations and conversions, and cannot otherwise interoperate with other types.

As a workaround, the standard library provides wrapper classes, instances of which store a single primitive value and present it as an object. In Java 5, implicit boxing and unboxing conversions were introduced, automatically converting the primitive values to wrapper class instances, and vice versa, as required by the program.

Conventionally, boxing is to be avoided when possible, because wrapper class instances perform significantly worse than bare primitives and can introduce subtle dependencies on object identity.

But the value objects feature, enhanced with [null-restricted types][jep-storage], eliminates most of these problems. As a result, boxing can be used more liberally, and arbitrary distinctions in the Java language between primitive and reference types can be minimized.

Ideally, it should be possible to eliminate most of the limitations of primitives, using boxing as needed to allow developers to use primitive types much like other types.

Description

The features described below are preview features, enabled with the --enable-preview compile-time and runtime flags.

Member accesses

An expression of a primitive type may appear to the left of a . in a field access or method invocation, and to the left of :: in a method reference. The members of the corresponding wrapper class are searched for a matching field or method.

int i = 12;
int iSize = i.SIZE;
double iAsDouble = i.doubleValue();
Supplier<String> iSupp = i::toString;

(To do: does a literal to the left of the . introduce any parsing problems?)

At run time, boxing is applied to the primitive value before the member access occurs.

The name of a primitive type may also be used in a field access, method invocation, or method reference.

int max = int.MAX_VALUE;
int zeros = int.numberOfLeadingZeros(max);
ToIntFunction<String> parser = int::parseInt;

Primitive return overriding

A method with a primitive return type may override a method with a reference return type, or vice versa, as long as the boxed or unboxed return type of the overriding method would be a legal return type.

interface Option {
    String name();
    Object value();
}

interface BooleanOption extends Option {
    String name();
    boolean value();
}

When the overridden method is invoked, the result is boxed or unboxed before returning. (This is implemented via a bridge method that performs the conversion.)

Primitive type arguments

A primitive type can be used as a type argument.

In order to properly handle null, the [nullness][nullness-jep] of type variable uses affects how primitive type argument substitution behaves. A null-restricted or parametric type variable use (T!, T*) maps to the primitive type (int). A nullable type variable use (T?) maps to the nullable boxed type (Integer?). And an unspecified-nullness type variable use maps to an unspecified-nullness boxed type (Integer).

interface Foo<T> {
    T* get(); // Foo<char> returns char
    T! getNonNull(); // Foo<char> returns char
    T? getOrNull(); // Foo<char> returns Character?
    T getOrAlternate(Supplier<T> alt); Foo<char> returns Character
}

For the purpose of bounds checking—testing for either parameterized type well-formedness or wildcard containment—boxing conversion is applied to the primitive type before the comparison to bounding types.

At the use site of a primitive-parameterized type, boxing and unboxing is implicit as needed for primitive values to interoperate with reference-typed type variables. Within the body of the generic class or method, type variables continue to be understood to range over reference types.

A new kind of unchecked boxing conversion and unchecked unboxing conversion allows Foo<int> to be converted to Foo<Integer!>, and vice versa. (This applies to both top level type arguments and types mentioned as nested type arguments or wildcard bounds.) These conversions can be thought of as lazily boxing and unboxing at the generic API boundaries.

At run time, generics continue to be implemented via erasure, so a List<int> is no more performant than a List<Object>—unlike in many other uses of boxing, these Integer instances will be heap-allocated. But future JVM enhancements will allow for specialized performance optimizations for primitive parameterizations (see Dependencies).

Overriding, overloading, and type arguments inference

In general, it has always been the case that a primitive type in a method signature is distinct from the boxed type. The two methods can be overloaded, and the compiler prefers to invoke the overload (if any) that doesn't require any boxing/unboxing conversions.

However, a type variable instantiated with a primitive type requires special treatment.

If a class extends a primitive-parameterized class type, then superclass methods with the corresponding type variable as a parameter type effectively represent both the boxed and unboxed variants of the signature. Either the primitive type or its boxed counterpart may be used to override the method. The usual rule about conflicting erased signatures applies: a method that doesn't override the superclass method may conflict with it if the method makes use of either the primitive type or its boxed counterpart.

Just like the handling of overriding with different return types, overriding in this case is implemented via bridge methods that perform any necessary boxing/unboxing. For binary compatibility, a bridge method is always generated with the boxed type instantiation of parameter and return types.

interface Box<T> {
    T get();
    void set(T val);
}

interface IntBox extends Box<int>
                 // formerly extends Box<Integer>
{
    int get();
    void set(int val);
}

class AnIntBox implements IntBox {
    int val;
    public Integer get() { return val; }
    public void set(Integer val) { this.val = val; }
}

Type argument inference may produce primitive results if primitive types appear at the use site. Primitive types are boxed before comparing them to other types.

For the purpose of overload resolution, type variables instantiated with primitive types are treated like reference types (for example, a method with an explicit parameter type int is preferred over a method with a generic parameter type instantiated with int).

<T> List<T> pair(T x, T y);
<T> List<T> singleton(T x);
IntList singleton(int x);

var l1 = pair(1, 2); // List<int>
List<Integer> l2 = pair(1, 2); // List<Integer>
var l3 = pair(1, 2.0); // List<Number & Comparable & ...>
var l4 = singleton(23); // IntList
var l5 = singleton((Integer) 23); // List<Integer>

Covariant arrays

Unchecked boxing and unboxing conversions also allow an int[]</code> to be
treated as an <code class="prettyprint" >Integer![], and vice versa.

int[] ints = new int[]{ 1, 2, 3 };
Object[] objs = ints;

assert objs[2] instanceof Integer;

Integer![] integers = new Integer![]{ 4, 5, 6 };
ints = integers;
assert ints[2] == 6;

At run time, this behavior requires JVM support: arrays allocated as int[] need to respond to aaload and aastore, while arrays allocated as Integer[] need to respond to iaload and iastore. Since the encoding of values stored in the array ought to be the same in either case, the behavior of these accessor instructions doesn't actually need to change, but verification must allow for these conversions.

Alternatives

It's tempting to entirely eliminate the distinction between primitive and reference types, making int and Integer! equivalent. But the distinction can't be eliminated at the JVM level, so certain seams are unavoidable. Compatibility with the existing language rules (such as method overloading) and previously-compiled binaries (which reference primitive types) also makes a wholesale transition difficult.

Dependencies

This JEP depends on Value Classes and Objects, which establishes the semantics of identity-free objects. Some details also depend on Null-Restricted Value Class Types, which separates nullness from boxing and introduces null-restricted type arguments.

In the future, JVM class and method specialization (JEP 218, with revisions) will allow generic classes and methods to specialize field, array, and local variable layouts when parameterized by primitive types.