Fix #5750: Type then part of If with no else part as Unit

[allanrenucci]

No description provided.

[smarter]

Wow, that was quick! Doing it in Typer seems like a good idea.
Before merging, please delete the () and associated comment in https://github.com/dotty-staging/dotty/blob/2ee9994efaf6b3b8b97377002c4a237fe9b5ce00/compiler/src/dotty/tools/dotc/ast/Positioned.scala#L56

[allanrenucci]

Actually I am not sure this is the correct fix. Consider:

class Test {
  def foo: String = ""
  def test(cond: Boolean): Int = {
    if (cond) foo else foo
    1
  }
}

This compiles to:

public class Test {
    public String foo() {
        return "";
    }

    public int test(boolean cond) {
        String string = cond ? this.foo() : this.foo();
        return 1;
    }
}

which is not optimal.

I believe the correct fix it to typecheck if in statement position with Unit as expected type

[smarter]

I believe the correct fix it to typecheck if in statement position with Unit as expected type

Careful though, we don't want val x = if ... else ... and if ... else ... to do different things because a different expected type was used leading to different type inference. So when there's an if/else we should first type it with the usual expected type computed through harmonic, then if it's a statement we can do something special.

[allanrenucci]

Careful though, we don't want val x = if ... else ... and if ... else ... to do different things because a different expected type was used leading to different type inference

Why not? This is already the case:

def foo: String = if (...) ... else ... // typed with `String` as expected type
val x = if (...) ... else ... // typed with no expected type

So it seems reasonable to me to implicitly type an if with Unit as expected type when it is in a statement position.

And we already have expressions that are implicitly typed with an expected type when in a certain position:

throw expr // typed as `Throwable`
if (cond) // cond typed as `Boolean`

[smarter]

hmm, ok, maybe it's not a big deal.

[odersky]

I think we tried checking statements with Unit as expected type at some point, but that got us into weirdness with type inference, so we backed out. So I would not open that can of worms at this point. I think the changes are OK, I just have two proposed simplifications.

[allanrenucci]

Comments addressed. Maybe we should keep the issue open since it doesn't solve the problem for if with else branch and matches