[Breaking Change] Make type inference of `e1 ?? e2` consistent. about sdk

[stereotype441]

@Hixie

This is just about how the type on the RHS of ?? is inferred, not how the type of the whole expression is inferred, right?

Correct, it's just about how the RHS of ?? is inferred.

As in:

T? nope<T>() => null;

void main() {
  var x = nope<int>() ?? ''; // x is inferred to be Object rather than int, even after this change
}

Yes. In fact, this code is unaffected by the change for two reasons: (1) because the type of the RHS ('') is always String, regardless of the context in which it's inferred, and (2) since the variable declaration has an implicit type, the context for the whole if-null expression nope<int>() ?? '' is _, not dynamic.

[itsjustkevin]

@Hixie @grouma @vsmenon for review as a follow on to #55418.

[Hixie]

This is just about how the type on the RHS of ?? is inferred, not how the type of the whole expression is inferred, right?

As in:

T? nope<T>() => null;

void main() {
  var x = nope<int>() ?? ''; // x is inferred to be Object rather than int, even after this change
}

Assuming I understand correctly, SGTM.

[vsmenon]

lgtm

[lrhn]

(Shouldn't the context type of e2 be NonNull(S) where S is the static type of e1, not just the type itself? It does seem to consistently be the static type itself, but that means that

void main() async {
  List<int>? maybeList = null as dynamic;
  var x = maybeList ?? (await Future.value([]));
  print([x].runtimeType);
}

gives x a static type of List<int>? for no good reason. Probably not something to fix in this "consistency CL", since it's something both implementations agree on today.)

[stereotype441]

@lrhn

(Shouldn't the context type of e2 be NonNull(S) where S is the static type of e1, not just the type itself? It does seem to consistently be the static type itself, but that means that

void main() async {
  List<int>? maybeList = null as dynamic;
  var x = maybeList ?? (await Future.value([]));
  print([x].runtimeType);
}

gives x a static type of List<int>? for no good reason.

Yeah, considering that this is an example of an aspirational context, I agree. It makes sense to do something as close as possible to what we think the user probably wants, and NonNull(S) seems like a better model of what the user probably wants than S.

Probably not something to fix in this "consistency CL", since it's something both implementations agree on today.)

Agreed 😃

[stereotype441]

@grouma ping

[grouma]

cc @leonsenft who will handle breaking change requests for ACX going forward.

[leonsenft]

LGTM assuming

A prototype of this change caused zero test failures in Google's internal codebase

holds.

[stereotype441]

LGTM assuming

A prototype of this change caused zero test failures in Google's internal codebase

holds.

I will recheck before landing the change.

[itsjustkevin]

@stereotype441 lock it in!