NonNullable is not resolved for a mapped type in a contextual parameter type from instantiated type alias

[Andarist]

Bug Report

🔎 Search Terms

NonNullable, Omit, mapped type, contextual type

🕗 Version & Regression Information

• This changed between versions 4.7 and 4.8

⏯ Playground Link

Playground link with relevant code
Original repro case

💻 Code

type GestureKey = "drag";
type DragState = { movement: [number, number]; };

interface State {
  drag?: DragState;
}

type SharedGestureState = {
  dragging?: boolean;
};

type FullGestureState<Key extends GestureKey> = SharedGestureState &
  NonNullable<State[Key]>;

type Handler<Key extends GestureKey> = (
  state: Omit<FullGestureState<Key>, "event">
) => void;

const works = (state: Omit<FullGestureState<"drag">, "event">) => {
  state;
  // ^? (parameter) state: Omit<SharedGestureState & DragState, "event">
  
  type StateKeys = keyof typeof state;
  //   ^? type StateKeys = "dragging" | "movement"
  console.log(state.movement);
};

const doesntWork: Handler<"drag"> = (state) => {
  state; (parameter) state: Omit<SharedGestureState & DragState, "event">
  // ^? (parameter) state: Omit<SharedGestureState & DragState, "event">

  type StateKeys = keyof typeof state;
  //   ^? type StateKeys = "dragging"
  console.log(state.movement); // Property 'movement' does not exist on type 'Omit<SharedGestureState & DragState, "event">'.(2339)
};

🙁 Actual behavior

We can observe a type error in the doesntWork function and the keyof reporting incorrect result there for the given parameter type.

🙂 Expected behavior

There should be no error as the type for the state parameter in both functions is the same.

cc @ahejlsberg

[typescript-bot]

The change between release-4.6 and main occurred at 51b346d.

[ahejlsberg]

The core issue here is the normalization we perform on index types. Specifically, we generally normalize keyof (A & B) into keyof A | keyof B. Turns out that doesn't always hold when A and B are generic types:

type Foo = { x: string } | undefined;
type Bar = { y: string };

type FooAndBar = Foo & Bar;

type Keys<T, U> = keyof (T & U);  // Normalized into keyof T | keyof U

type K1 = keyof Foo;  // never
type K2 = keyof Bar;  // "y"
type K3 = Keys<Foo, Bar>;  // "y"
type K4 = keyof (Foo & Bar);  // "x" | "y"

Above, keyof Foo | keyof Bar is just "y", whereas keyof (Foo & Bar) is "x" | "y" because Foo & Bar is reduced to just { x: string} & { y: string }, thus eliminating the undefined type.

We already have logic in place to defer types like keyof (T & {}), but it looks like we need to defer all keyof (T & U) where one or both of the types are generic. We can then simplify the type to keyof T | keyof U where appropriate. I'm going to put together a PR and we'll see what the effects are.