anything is assignable to T[P] where T extends null/undefined/void/object/never

[zpdDG4gta8XKpMCd]

case 1

export function test<T extends object, P extends keyof T>(
    whatever: T, name: P
): T[P] {
    const result = Math.random() > 0.5 ? whatever[name] : 'Hello!'; // <-- expected T[P] | string, actual T[P]
    return result; // <-- returning unsound result
}

case 2

export function oneOrAnother<A, B>(one: A, another: B): A | B {
    return Math.random() > 0.5 ? one : another;
}

export function test<T extends object, P extends keyof T>(
    whatever: T, name: P
): T[P] {
    const result = oneOrAnother(whatever[name], 'Hello!'); // <-- result is clearly string | T[P] as expected
    return result; // <-- expected a type error, actual returning string | T[P] instead of T[P] no problem
}

[MrAndersen1]

You probably already found this out but these issues arise from extending object.
What one gains from extending object and why these issues arise when you do it is beyond me however.

[sandersn]

Here's a smaller repro:

function test<T extends object, P extends keyof T>(s: string, tp: T[P]): void {
    tp = s; // should error
}

[sandersn]

To check assignability of string to T[P], the compiler checks whether string is assignable to the constraint of T[P]. But getConstraintOfIndexedAccess results in anyType:

• the type constraint of T[P] is object[string]
• But object doesn't have a string index signature.
• so getPropertyTypeForIndexType returns anyType.

So string is assignable to T[P] because the apparent type of T[P] is any, which is incorrect. I think getConstraintOfIndexedAccess is incorrect but I'm not yet sure what the fix is.

[sandersn]

Fix is up at #17870.

[sandersn]

The previous PR got merged by mistake and is now reverted; #17912 is the PR with the fix now.

[zpdDG4gta8XKpMCd]

looks like this pr might be the reason to #12215 (comment)