Discriminated union containing optional properties and generic types is not narrowed

[TomerAberbach]

Bug Report

I would have expected doesNotNarrow (below) to also be able to narrow x.b to NonUndefined<B>.

🔎 Search Terms

generic, narrow, discriminated, union, optional

🕗 Version & Regression Information

This is the behavior in every version I tried, and I reviewed all the FAQ for entries.

⏯ Playground Link

Playground link with relevant code

💻 Code

 type NonUndefined<T> = T extends undefined ? never : T

type X<A, B> = { a: NonUndefined<A>, b?: never } | { a?: never, b: NonUndefined<B> }

function doesNarrow(x: X<number, number>) {
  if (x.a === undefined) {
    const y = x.b
      // ^? number
  } 
}

function doesNotNarrow<A extends number, B extends number>(x: X<A, B>) {
  if (x.a === undefined) {
    const y = x.b
      // ^? NonUndefined<B> | undefined
  } 
}

🙁 Actual behavior

x.b was not narrowed.

🙂 Expected behavior

I expected x.b to be narrowed because if x.a is undefined, then x must be the latter part of the discriminated union because x.a can only be undefined there. TypeScript seems to be able to work that out when the type parameters are constant/non-generic (see doesNarrow), but for some reason it's not able to work it out for the generic case (even though NonUndefined should make it feasible).

[RyanCavanaugh]

TypeScript doesn't have negated types, so attempts to recreate them through constructs like NonUndefined are sort of doomed to fail.

[typescript-bot]

This issue has been marked as "Duplicate" and has seen no recent activity. It has been automatically closed for house-keeping purposes.

[TomerAberbach]

Even without NonUndefined and with the generics constrained to number it can't figure it out

These two functions probably should behave identically in terms of narrowing

[TomerAberbach]

Even without NonUndefined and with the generics constrained to number it can't figure it out

These two functions probably should behave identically in terms of narrowing

@RyanCavanaugh wondering if you'd be able to take a second look at this issue? Specifically my comment above? I think the issue isn't really a duplicate of the negated types feature request given the above playground link. What do you think?