Has the behaviour that was added in #27490 been taken out? The following function did error, but not under the new rules.

function fun<U extends { a: T }, T, K extends 'a'>(u: U, k: K, shambles: T): U[K] {
  u[k] = shambles; // was error
  return u[k];
}

const result: number = fun<{ a: number }, unknown, 'a'>({ a: 42 }, 'a', 'a string');

I think the reason that the test-cases from #27470 didn't regress is because the restrictive instantiation has been subsequently added to conditional types, which catches the error too.

Has the behaviour that was added in #27490 been taken out?

Yes, but that behavior wasn't right. For an indexed access T[K] we flat out didn't explore any constraints on the target side when both T and K were generic, which is quite inconsistent. For example, if you change the assignment in your example from u[k] = shambles to u['a'] = shambles there is no error before or now.

The real (and orthogonal) issue in your example is that we treat an assignment to T[K] as an assignment to C[K], where C is the constraint of T. I discuss the reasons in my comment here. The unsoundness caused by that rule continues to exist (and would be quite a bit harder to remove).

What has changed with this PR is that for a T[K] on the target side of a relation we now ignore index signatures in the constraint of T. That's definitely the right thing to do (because a type argument for T isn't actually required to have an index signature, it is merely required to have properties with matching types) and that's what really fixes the issue in #27470.

It's all rather involved!

Wesley Wigham (@weswigham) Can you take a look at why this is failing on PragmaPsuedoMap?

Looks like we're pulling out some unexpected index signatures from the target or something? The bug should repro in the test harness, will just need to copy the relevant types into a test file.

Wesley Wigham (@weswigham) It's actually identifying an unsafe coercion. It used to be that PragmaPseudoMap["reference"] was (unsoundly) assignable to PragmaPseudoMap[K], where K extends keyof PragmaPseudoMap, which, in combination with parameter bi-variance (because the compiler isn't yet using --strictFunctionTypes), caused the explicit type annotation on (arg: PragmaPseudoMap["reference"]) => ... to be accepted. That's no longer the case, and that's a good thing.

The easy fix is to explicitly cast the array parameter to forEach to the appropriate type.

TypeScript Bot (@typescript-bot) test this

Heya Anders Hejlsberg (@ahejlsberg), I've started to run the extended test suite on this PR at e1fd5e5. You can monitor the build here. It should now contribute to this PR's status checks.

TypeScript Bot (@typescript-bot) run dt

Heya Anders Hejlsberg (@ahejlsberg), I've started to run the Definitely Typed test suite on this PR at e1fd5e5. You can monitor the build here. It should now contribute to this PR's status checks.

Thanks for the detailed write-up Anders Hejlsberg (@ahejlsberg).

The real (and orthogonal) issue in your example is that we treat an assignment to T[K] as an assignment to C[K], where C is the constraint of T. I discuss the reasons in my comment here. The unsoundness caused by that rule continues to exist (and would be quite a bit harder to remove).

True, though did you not add behaviour to prevent pulling down C when T and K were both generic variables which is no longer in this PR? I understand that the original issue that prompted the fix is now correctly handled for other (more sensible) reasons, but discarding a sound rule seems a shame! Though, the rule is abit ad-hoc!

With latest commits there are three RWC projects with new errors. The errors in two of the projects are minor and identify unsoundness that we previously didn't catch. The last project, fp-ts, which already has massive inverted pyramids of error messages, now looks even more daunting. It makes heavy use of indexed access types and when they occur in target positions we create intersection types instead of union types. There are several new errors that appear to be correct with the stricter checking, but it's hard to tell.

TypeScript Bot (@typescript-bot) test this

• Given a type variable T with a constraint C, when an indexed access T[K] occurs on the target side of a type relationship, index signatures in C are now ignored. This is because a type argument for T isn't actually required to have an index signature, it is just required to have properties with matching types.

Anders Hejlsberg (@ahejlsberg) If "a type argument for T isn't actually required to have an index signature", why is indexed access still allowed on the source side of a type relationship?

If "a type argument for T isn't actually required to have an index signature", why is indexed access still allowed on the source side of a type relationship?

An indexed access T[K] doesn't actually require T to have an index signature. It's more helpful to think of it as a property access. As long as K is known to reference the name of a property in T (i.e. because K is a literal type like "foo"), then T[K] is permitted. However, an indexed access T[string] does require T to have an index signature.

Could the errors in this test case be overly strict?

function f3<T extends { [key: string]: any }>(obj: T) {
    let foo = obj['foo'];
    let bar = obj['bar'];
    obj['foo'] = 123;  // Error  - (its writable and matches the parameters why error?)
    obj['bar'] = 'x';  // Error - (ditto)
}

This readonly function format exists if errors are desired:

function f5<T extends Readonly<{ [key: string]: any }>>(obj: T) {
    let foo = obj['foo'];
    let bar = obj['bar'];
    obj['foo'] = 123;  // Error
    obj['bar'] = 'x';  // Error
}
declare const obj: {
  foo: boolean;
  bar: boolean;
};
f5(obj);

Shouldn't TypeScript assume the user is reading the function parameters and trust them not to call f3 because of the any type and the fact that it is non-readonly?

Conversely, the user may have a special need for writing a writable any typed function like f3, and now they are not allowed to write it. That could be frustrating.