Type guard of union with conditional type not working since 4.3

Bug Report

🔎 Search Terms

type guard conditional type

🕗 Version & Regression Information

• This changed between versions 4.2.3 and 4.3.2

⏯ Playground Link

Playground link with relevant code

💻 Code

interface A<S> {
    a: S
}
interface B<S> {
    b: S
}
interface C<S> {
    c: S
}

type U1<S> = A<S> | B<S> | C<S>
function f1<S>(u: U1<S>): u is B<S> | C<S> {
    return false
}
function test1<S>(x: U1<S>) {
    if (!f1(x)) {
        x.a //OK
    }
}

type Cond<S> = S extends number ? B<S> : C<S>
type U2<S> = A<S> | Cond<S>
function f2<S>(u: U2<S>): u is Cond<S> {
    return false
}
function test2<S>(x: U2<S>) {
    if (!f2(x)) {
        x.a //ERROR
    }
}

🙁 Actual behavior

The type guard is unable to narrow the type of x to A

🙂 Expected behavior

The type guard should narrow the type of x to A because it is not assignable to Cond<S>

[andrewbranch]

This was caused by #43183, because the conditional type Cond<S> gets substituted for its constraint, B<S> | C<S>, in the type of x before it gets narrowed, so we are effectively trying to narrow from A<S> | B<S> | C<S>. Most of the time, that’s way more useful than trying to narrow with Cond<S>, because it means you can actually discriminate between the two branches of the conditional. Cond<S> just won’t relate with many useful types during narrowing. However, it will relate to itself, which is what your example is doing, but in 4.3 we can no longer tell that’s what’s happening since we substituted the constraint earlier on.

I don’t know if it would make sense for your real code @pedro-pedrosa, but a workaround would be to expand the conditional type to a union in your type predicate:

function f2<S>(u: A<S> | B<S> | C<S>): u is B<S> | C<S> {
    return false
}

The only way I can think to fix this without undoing the very useful improvements of #43183 would be to carry the original, non-substituted type through narrowing (in addition to the usually-more-narrowable constraint-substituted type), trying it if narrowing the constraint-substituted type had no effect. But that sounds like it could be expensive, for what feels like quite an edge case to me.

[RyanCavanaugh]

This doesn’t meet our bar for a patch release, but I’ll move it to 4.4