为什么这个带有绑定的泛型方法可以返回任何类型？

Question

Why does the following code compile? The method IElement.getX(String) returns an instance of the type IElement or of subclasses thereof. The code in the Main class invokes the getX(String) method. The compiler allows to store the return value to a variable of the type Integer (which obviously is not in the hierarchy of IElement ).

public interface IElement extends CharSequence {
  <T extends IElement> T getX(String value);
}

public class Main {
  public void example(IElement element) {
    Integer x = element.getX("x");
  }
}

Shouldn't the return type still be an instance of IElement - even after the type erasure?

The bytecode of the getX(String) method is:

public abstract <T extends IElement> T getX(java.lang.String);
flags: ACC_PUBLIC, ACC_ABSTRACT
Signature: #7                           // <T::LIElement;>(Ljava/lang/String;)TT;

Edit: Replaced String consistently with Integer .

Answer 1

This is actually a legitimate type inference*.

We can reduce this to the following example ( Ideone ):

interface Foo {
    <F extends Foo> F bar();
    
    public static void main(String[] args) {
        Foo foo = null;
        String baz = foo.bar();
    }
}

The compiler is allowed to infer a (nonsensical, really) intersection type String & Foo because Foo is an interface. For the example in the question, Integer & IElement is inferred.

It's nonsensical because the conversion is impossible. We can't do such a cast ourselves:

// won't compile because Integer is final
Integer x = (Integer & IElement) element;

---

Type inference basically works with:

• a set of inference variables for each of a method's type parameters.
• a set of bounds that must be conformed to.
• sometimes constraints , which are reduced to bounds.

At the end of the algorithm, each variable is resolved to an intersection type based on the bound set, and if they're valid, the invocation compiles.

The process begins in 8.1.3 :

When inference begins, a bound set is typically generated from a list of type parameter declarations P 1 , ..., P p and associated inference variables α 1 , ..., α p . Such a bound set is constructed as follows. For each l (1 ≤ l ≤ p) :

• […]

• Otherwise, for each type T delimited by & in a TypeBound , the bound α l <: T[P 1 :=α 1 , ..., P p :=α p ] appears in the set […].

So, this means first the compiler starts with a bound of F <: Foo (which means F is a subtype of Foo ).

Moving to 18.5.2 , the return target type gets considered:

If the invocation is a poly expression, […] let R be the return type of m , let T be the invocation's target type, and then:

• […]

• Otherwise, the constraint formula ‹R θ → T› is reduced and incorporated with [the bound set].

The constraint formula ‹R θ → T› gets reduced to another bound of R θ <: T , so we have F <: String .

Later on these get resolved according to 18.4 :

[…] a candidate instantiation T i is defined for each α i :

• Otherwise, where α i has proper upper bounds U 1 , ..., U k , T i = glb(U 1 , ..., U k ) .

The bounds α 1 = T 1 , ..., α n = T n are incorporated with the current bound set.

Recall that our set of bounds is F <: Foo, F <: String . glb(String, Foo) is defined as String & Foo . This is apparently a legitimate type for glb , which only requires that:

It is a compile-time error if, for any two classes ( not interfaces ) V i and V j , V i is not a subclass of V j or vice versa.

Finally:

If resolution succeeds with instantiations T 1 , ..., T p for inference variables α 1 , ..., α p , let θ' be the substitution [P 1 :=T 1 , ..., P p :=T p ] . Then:

• If unchecked conversion was not necessary for the method to be applicable, then the invocation type of m is obtained by applying θ' to the type of m .

The method is therefore invoked with String & Foo as the type of F . We can of course assign this to a String , thus impossibly converting a Foo to a String .

The fact that String / Integer are final classes is apparently not considered.

---

_{* Note: type erasure is/was completely unrelated to the issue.}

_{Also, while this compiles on Java 7 as well, I think it's reasonable to say we needn't worry about the specification there. Java 7's type inference was essentially a less sophisticated version of Java 8's. It compiles for similar reasons.}

---

As an addendum, while strange, this will likely never cause a problem that was not already present. It's rarely useful to write a generic method whose return type is solely inferred from the return target, because only null can be returned from such a method without casting.

Suppose for example we have some map analog which stores subtypes of a particular interface:

interface FooImplMap {
    void put(String key, Foo value);
    <F extends Foo> F get(String key);
}

class Bar implements Foo {}
class Biz implements Foo {}

It's already perfectly valid to make an error such as the following:

FooImplMap m = ...;
m.put("b", new Bar());
Biz b = m.get("b"); // casting Bar to Biz

So the fact that we can also do Integer i = m.get("b"); is not a new possibility for error. If we were programming code like this, it was already potentially unsound to begin with.

Generally, a type parameter should only be solely inferred from the target type if there is no reason to bound it, eg Collections.emptyList() and Optional.empty() :

private static final Optional<?> EMPTY = new Optional<>();

public static<T> Optional<T> empty() {
    @SuppressWarnings("unchecked")
    Optional<T> t = (Optional<T>) EMPTY;
    return t;
}

This is A-OK because Optional.empty() can neither produce nor consume a T .