Generic type inference prefers conditional types

[bowenni]

TypeScript Version: 3.4.5 and 3.5.1

Search Terms: Conditional types, Type inference

Code

interface I {a: string}
type Cond<T> = T extends any ? Partial<T>: T;

declare function f<T>(x1: T, x2: Cond<T>): T;
function g(p1: I, p2: Cond<I>): I {
  return f(p1, p2); // error here
}

Expected behavior:
No errors. A trivial type inference is T=I. This is also the behavior of TypeScript 3.3.

Actual behavior:
error TS2322: Type 'Partial<I>' is not assignable to type 'I'.
From that error message the actual inference is T=Partial<I> (or Cond<I>), which seems like a regression to me.

Playground Link:
playground link

Related Issues:
There are some issues around conditional types but nothing quite a duplication, I think.
#31601 requires function overloads
#24085 is a problem with enums
#30489 is more about type inference on the conditional type itself, not the generic type when the conditional type is used
#30341 is similar but it's marked as fixed

@rkirov

[RyanCavanaugh]

What is the intent of writing T extends any ?

[bowenni]

What is the intent of writing T extends any ?

To create a conditional type. I think the problem only happens with a conditional type.

If the example is oversimplified and doesn't make sense, below is another example, without the T extends any.

interface I { a: string }
type DeepPartial<T> =
    T extends object ? {[K in keyof T]?: DeepPartial<T[K]>} : T;

declare function f<T>(t: T, partial: DeepPartial<T>): T;
function g(p1: I, p2: Partial<I>): I {
  return f(p1, p2); // error Type 'Partial<I>' is not assignable to type 'I'.
}

[RyanCavanaugh]

interface I { a: string }
type DeepPartial<T> =
    T extends object ? {[K in keyof T]?: DeepPartial<T[K]>} : T;

declare function f<T>(t: T, partial: DeepPartial<T>): T;
function g(p1: I, p2: Partial<I>): I {
    // T inferred to be Partial<I>; should be I
  return f(p1, p2); // error Type 'Partial<I>' is not assignable to type 'I'.
}

[ahejlsberg]

This is caused by #30010. Part of why we inferred to a union type was to get lower priority candidates for inferences to naked type variables, but that got lost with the fix in #30010. We just need to reinstate that part.

[weswigham]

This is caused by #30010.

To be fair, at the time I'm pretty sure we had done away with the naked type variable priority; we just didn't think to add it to the conditional inference bit when we added it back. :P