Generic functions that have a function as a parameter expand inferred type instead of narrowing

[nomcopter]

TypeScript Version: typescript@3.1.0-dev.20180801

Search Terms: Generic Type Inference Function Parameter Union

Code

interface Foo {
    foo: number;
}

declare function acceptUnion(foo: Foo | number): void;
declare var foo: Foo;

declare function erroneouslyExpands<T>(value: T, func: (t: T) => void): T;

 // Errors, TS infers the result as `Foo | number`
const test: Foo = erroneouslyExpands(foo, acceptUnion);

Note: This case is common when passing comparators into functions.

Expected behavior:

T should be inferred as Foo.

Actual behavior:

T is expanded based on what func accepts.

Playground Link: https://www.typescriptlang.org/play/index.html#src=interface%20Foo%20%7B%0A%20%20%20%20foo%3A%20number%3B%0A%7D%0A%0Adeclare%20function%20acceptUnion(foo%3A%20Foo%20%7C%20number)%3A%20void%3B%0Adeclare%20var%20foo%3A%20Foo%3B%0A%0Adeclare%20function%20erroneouslyExpands%3CT%3E(value%3A%20T%2C%20func%3A%20(t%3A%20T)%20%3D%3E%20void)%3A%20T%3B%0A%0Aconst%20test%3A%20Foo%20%3D%20erroneouslyExpands(foo%2C%20acceptUnion)%3B%0A

Related Issues:

• Unexpected union inference because of generic function argument #16107
• Type inference issue when using generic function and generic constraint #24747 may reduce to this issue..?

[RyanCavanaugh]

See #14829

[ahejlsberg]

In --strictFunctionTypes mode we should definitely infer just Foo, so I'm changing this back to a bug.

[ahejlsberg]

However, fixing this causes #19576 to again break. I seems that having #19576 be an error is the lesser of two evils, but it would be nice to somehow keep around contra-variant inferences when the only co-variant inference we have is a never resulting from a freshly created array literal.

[thw0rted]

I'm still having issues with this, I think. You can see the error on this Playground example. Note that bad gives a compiler error.

The problem that lead me here is that const s: BehaviorSubject<number[]> = new BehaviorSubject([]); is an error if I turn strictFunctionTypes on. ("Type 'BehaviorSubject<never[]>' is not assignable to type 'BehaviorSubject<number[]>'.") I can't imagine that we really want everybody to have to write new BehaviorSubject([] as number[]) all over the place. Am I missing something?