Conditional type prevents assignability

[dylhunn]

Bug Report

🔎 Search Terms

Index signature assignability, generic index signatures, conditional types

🕗 Version & Regression Information

• This is the behavior in every version I tried, and I reviewed the FAQ for entries about conditional types

⏯ Playground Link

Playground link with relevant code

💻 Code

// Use of the conditional type `string extends T` prevents assignability.
// However, the conditional type `T extends string` is accepted.

declare class Foo<T>  {
    bar: string extends T ? number : boolean;
}

declare let indexable: Foo<{[key: string]: boolean}>;
declare let namedProp: Foo<{baz: boolean}>;

indexable = namedProp;

🙁 Actual behavior

When a conditional type is present in Foo, I sometimes cannot assign Foo<{a: any}> to Foo<[x: string]: any>, depending on the condition.

🙂 Expected behavior

Regardless of the conditional type, I should be able to assign Foo with named properties to Foo with an index signature.

[RyanCavanaugh]

Likely we're considering a type extends T to mark Foo<T> as invariant in T

[dylhunn]

Likely we're considering a type extends T to mark Foo as invariant in T

Is this on a roadmap or existing issue? I'd like to know whether to wait for this fix, or design around it. Thanks!

[dylhunn]

(More context: we'd be using this as part of the solution for this longstanding Angular issue.)

[RyanCavanaugh]

It's very possible a straightforward fix here (if one is even possible) could have side effects that result in the fix being rejected. I'd recommend designing around it if possible.

[jihndai]

After walking through the compiler a bit, it does seem to be marking type extends T's case as invariant like you said, while T extends type's case gets marked as bivariant.

Is the expected behavior for type extends T to be flagged as bivariant like T extends type? Although it might satisfy the above test case, I think that it would probably lead to a weird behavior similar to how bivariance currently interacts with T extends type, for example:

declare class A { a: string; }
declare class B extends A { b: string; }
declare class Foo<T> { foo: T extends B ? number : boolean }

declare let fooA: Foo<A>;
declare let tempA: Foo<A>;
declare let fooB: Foo<B>;

fooA.foo; // boolean
fooB.foo; // number

tempA = fooA;
fooA = fooB;
fooB = tempA;

fooA.foo; // incorrectly boolean
fooB.foo; // incorrectly number

Or should we try something else, like somehow applying a structural comparison between the classes when we encounter those kinds of conditional types with generics? For example, below I use a duplicate of class Foo<T> to cause the compiler to perform a structural comparison instead of checking type parameters variance:

declare class A { a: string; }
declare class B extends A { b: string; }
declare class Foo<T> { foo: B extends T ? number : boolean }
declare class Boo<T> { foo: B extends T ? number : boolean }

declare let fooA: Foo<A>;
declare let tempA: Foo<A>;
declare let fooB: Boo<B>;

tempA = fooA;
// both assignment works because fooA.foo and fooB.foo are both number
// even though the type parameters aren't bivariant
fooA = fooB;
fooB = tempA;

[dylhunn]

Here's a more robust and extremely surprising example of what can happen as a result of this invariance, in particular as it interacts with boolean widening. Playground link.

Code:

class FormArray<T extends FormControl<any>> {
  constructor(public readonly controls: Array<T>) {}

  // When T is on the RHS of a conditional type, FormArray<T> becomes invariant in T.
  // https://github.com/microsoft/TypeScript/issues/48033
  foo!: string extends T ? true : false;
}

class FormControl<T> {
  constructor(public readonly value: T) {}
}

function test() {
  // Notice that TypeScript infers this as `FormControl<boolean>` -- there are some
  // (likely special-cased) rules for inferring `boolean` instead of the narrow literal type `true`
  const fc1 = new FormControl(true);
  // Here, explicitly specifying the boolean on the LHS blocks the special casing above,
  // causing the RHS constructor to be inferred as `FormControl<true>`.
  const fc2: FormControl<boolean> = new FormControl(true);

  // This infers as `FormArray<FormControl<boolean>>`, as above.
  const fa1 = new FormArray([new FormControl(true)]);
  // Error! Again, providing an explicit type causes the RHS control to be inferred
  // as `FormControl<true>` instead of `FormControl<boolean>`. Because `FormArray<T>`
  // is invariant in T, `FormArray<FormControl<true>>` is not assignable to `FormArray<FormControl<boolean>>`.
  const fa2: FormArray<FormControl<boolean>> = new FormArray([new FormControl(true)]);
}

[yseymour]

This also affects mapped types:

interface M {
    A: number;
    B: number;
}

declare class Foo<T extends keyof M> {
    bar: M[T]
}

declare let foo1: Foo<"A">;
let foo2: Foo<"B"> = foo1; // "A" is not assignable to "B"

[RyanCavanaugh]

@yseymour that's a correct error. When M[T] is anything other than a roundabout way of writing number, the assignment is unsafe

interface M {
    A: number;
    B: string; // <- changed
}

declare class Foo<T extends keyof M> {
    bar: M[T]
}

// Legal initialization
let foo1: Foo<"A"> = { bar: 32 }
// Purported legal assignment
let foo2: Foo<"B"> = foo1;
// Unsound - number inhabits string binding
const m: string = foo2.bar;

[yseymour]

@RyanCavanaugh, agreed, I'd expect an assignability error between the two Foos in your example. But in mine, Foo<"A">["bar"] and Foo<"B">["bar"] are both number, so intuitively I'd expect the Foos to be fungible. Is it perhaps the case that checker doesn't look through the type argument to do a structural check on the individual members?

Maybe this is a different issue to the OP.