Instantiations of constrained generic signatures allow all assignments

[weswigham]

🔎 Search Terms

variance generic signature measurement

🕗 Version & Regression Information

This has been our behavior forever as far as I can tell - in fact, I've seen this pattern used once or twice to work around other bugs, even though the behavior is in no way safe.

⏯ Playground Link

https://www.typescriptlang.org/play?#code/C4TwDgpgBAYgdgHgCoD4oF4oIKpQgD2AjgBMBnKVACnwC4psBKDNbAbgCgPRIoBhAPZwAxgCcIRAIwYsuAkVIU4AVwC2AIwiiUNekxYNOPaIJHiiAJhk48hYuShlgogJZwA5jroNm6Vpw4AM2URYBchKEDkSQAaSgsdYEl6eGiUOOALFMQkBLjhZP4hMQkIWKhhLKKzUotmAG8OKGaoAHpWvAA3LRBHN3cAG2ghYcCoYAALCDJoAEMyMhd3OFViYAoyCYFlAZI8UVEBUS4W8elMTLY2js3t3f3D0Tj1ZWAoEgFpuAByYCaWzIyJJXdqOLY7PZaR7PV7vT5kH5-f7NAoySqcU6VNGSAKnJLYjEtVEXHEcAC+QA

💻 Code

type Fn<T> = <U extends T>(x: U) => U;

type Concrete1 = <U extends number>(x: U) => U;
type Concrete2 = <U extends string>(x: U) => U;

function f<T1, T2>(t1: Fn<T1>, t2: Fn<T2>, c1: Concrete1, c2: Concrete2) {
    // every single one of these assignments should error

    t1 = t2; // should error, but doesn't
    t2 = t1; // should error, but doesn't

    c1 = c2;
    c2 = c1;

    t1 = c1;
    c1 = t1;
}

🙁 Actual behavior

Missing two errors on the generic relationships.

🙂 Expected behavior

All of the assignments are errors, as the generic relationship should be even stricter than the non-generic one.

Additional information about the issue

The issue lies in our variance analysis - we currently incorrectly measure T as independent, which is quite different from the invariant it should adopt (via U's invariant usages). This discrepancy arises because of some oddities in how we do generic signature relationships.

[johanrosenkilde]

I came across the following beauty today (whittled down to a minimal):

type C = {
  a: string;
};

function t<A extends C, B extends C>(x: A, y: B): B["a"] {
  return x.a;
}


const a = { a: "a" } as const;
const b = { a: "b" } as const;
const res = t(a, b);  // res: "b"
                      // but actually res = "a"

I'd have expected that the function t doesn't type check, in particular the line return x.a.

Seems to be related/equal to your bug here?