regression in 5.4: error TS2604: JSX element type 'Cmp' does not have any construct or call signatures. about typescript

[DanielRosenwasser]

Don't know how that label got added - I haven't gotten the chance to actually look at this issue. My apologies.

[RyanCavanaugh]

#57117 (comment)

[wmertens]

@RyanCavanaugh yes but that error doesn't appear any more, only the 2604 one

[RyanCavanaugh]

This is, for all intents and purposes, a correct error that was missed in 5.3. Prior versions were incorrectly (within the bounds of correct analysis that actually exist) allowing the call, though I'm still not exactly sure why.

Further reduced (bisect on this one is still #56004):

declare const React: any;
namespace JSX {
  export interface IntrinsicElements {
    div: {
      name: string;
      as?: "div";
    };
  }
}

// This is required for the issue, changing it to return T makes it a different error
type ObjectProps<T> = T extends Record<any, any> ? T : never;
type FunctionComponent<P> = (props: P) => any;
// string|fn is required for the issue, changing it to only check for fn makes it a different error
type PropsOf<COMP> = COMP extends keyof JSX.IntrinsicElements
  ? JSX.IntrinsicElements[COMP]
  : COMP extends FunctionComponent<infer OrigProps>
  ? ObjectProps<OrigProps & Record<any, any>>
  : never

const Poly1 = <C extends string | FunctionComponent<unknown>>(thing: { as: C } & PropsOf<C>) => {
  const Cmp = thing.as;
  const p1 = <Cmp {...(null as any)} />;;
  if (typeof Cmp === "string") {
    const p2 = <Cmp {...(null as any)} />;
  } else {
    const p3 = <Cmp {...(null as any)} />;
  }
}

Poly1({ as: "div", name: "foo" });

In 5.3, p2 is an error, but p1 isn't. That doesn't make any sense -- if p2 is an error, then p1 has to be one too. In 5.4 they're consistently both errors.

Correct analysis of this function basically boils down to being able to reason about the fact that

Poly1<"baz">(...);

can't actually happen since if C = baz then props = never, but TS has never had that capability.

[wmertens]

I'm a bit lost - so you're saying that this polymorphic behavior should never have worked and cannot work?

[wmertens]

BTW in the original implementation if C=="Baz" then the props would be the ones from span

[wmertens]

@RyanCavanaugh I'm not sure what to do here - is a polymorphic component impossible, or are we doing it wrongly?

[wmertens]

Ok - with some inspiration here I managed to make it pass in newer TS versions, will merge soon. Since I understand that it was failing correctly, I'll close this issue.

[RyanCavanaugh]

Sorry for the delay btw, we were double-checking a few other things to make sure this wasn't secretly a bug of some kind. Glad to hear you got an acceptable solution on your end.