Design Meeting Notes, 12/1/2023

[DanielRosenwasser]

Ignoring undefined/null types fed into keyof

#51331

• Intersections of undefined and object types get evaporated away.

  type Foo = { x: string } | undefined;
  type Bar = { y: string };
  
  type FooAndBar = Foo & Bar;
  //   ^? { x: string } & Bar
  // No `undefined` because `undefined & Bar` is `never`

• Okay, what's the problem?

  type Foo = { x: string } | undefined;
  type Bar = { y: string };
  
  type KF = keyof Foo;
  type KB = keyof Bar;
  
  type K1 = keyof (Foo & Bar);
  //   ^? "x" | "y"
  type K2 = keyof Foo | keyof Bar;
  //   ^? "x"

• Uh oh.

• Could make the argument that keyof (T | undefined) should probably just be keyof T - if you have a type like keyof (T | undefined), being able to actually index into a T depends on actually being able to do something with the T itself.

• You can already write (T | undefined)[keyof T] in type space without any issues. So the idea of (T | undefined)[keyof (T | undefined)] being allowed makes sense.

• Are we sure we're not saving someone from a mistake?

  • Feels like as long as you're preventing the indexing in the value space, that's where the value is.

• Seems reasonable, we'd like to experiment with that.

• [[Revisiting - the getProp function actually doesn't make enough guarantees for us to do this safely.]]

  function getProp<T, K extends keyof T>(obj: T, key: K) {
      return obj[key]; // allowed!
  }
  
  let obj: { foo: string } | undefined
  getProp(obj, "foo"); // uh oh!

Node.js module interop changes

nodejs/node#50981
nodejs/node#50096

• Prologue directive to specify to Node what named exports a CommonJS provides.
• Node can't do the same analyses that bundlers perform, so this directive is intended to help here.
• Gives CJS modules a path forward for compat.
• Maybe not too bad because we serialize __esModule markers.
• For our emit, specifying these when given as inputs would "just work" because we preserve all prologue directives.
• The fact that this is in a JavaScript file complicates things for us. We would need to see some sort of marker in the .d.ts file.
  • Or we would need to resolve the JavaScript file and scan prologue directives.
• Apart from the complications, we are not a big fan of adding more and more and more and more to try to fix CJS/ESM interop. Every fix is another edge case for devs. It continues to complicate the options and configuration people need to manage.