`call` method invocation and `noSuchMethod` forwarder

[sgrekhov]

This issue derieved from the test mentioned in #55803 (comment). I didn't find a separate github item for this issue. Feel free to close this one as a duplicate if there is such one.

class BNSMC {
  // Should be NSM-forwarder.
  int Function(int) get call;

  Object? noSuchMethod(Invocation i) {
    if (i.memberName == #call) {
      print("isMethod=${i.isMethod}"); // false
      print("isGetter=${i.isGetter}"); // true
      return (int x) => x;
    }
    return super.noSuchMethod(i);
  }
}

void main() {
  (BNSMC() as dynamic)(42);
}

Here we have an attempt to invoke a call method, not a getter, so isMethod and isGetter should have opposite values.
cc @lrhn @eernstg to confirm.

[dart-github-bot]

Summary: noSuchMethod incorrectly identifies a call method invocation as a getter, resulting in isMethod being false and isGetter being true. This contradicts expected behavior.

[lrhn]

Not sure if we have specified this particular interaction between dynamic and noSuchMethod. (But Erik will know ).

If the call should not invoke the call getter, what should it do?
It can either throw, or invoke noSuchMethod with an Invocation, which should have member name #call and be a method invocation. That's the invocation that failed to happen.

It'll be the only place where you can get an Invocation with the same name as a getter. Any other attempt to invoke a getter as a method would attempt to .call the getter value.
It might be surprising to some code that you can get a failed invocation with the same name as a getter.

[eernstg]

It is a tricky case! ;-)

I think the actual behavior of the CFE matches the specified behavior:

We're considering an invocation of the expression e of the form (BNSMC() as dynamic), which means that we're applying the following rules:

• im.isMethod evaluates to true.
• ...

We're applying these rules because the previous cases do not match the given situation. The static analysis goes through the following steps: e is not a type literal, it is not an identifier, it is not a property extraction, and the static type of e is not an interface type that has a method named call. Hence, the invocation is not a compile-time error, and it is not "treated as" any other expression. At run time, the value of e is not a function object, and it doesn't have a method whose name is call.

(It doesn't help that it has a getter named call, we're only looking for a method.)

Following the rules at that point, we'll then invoke noSuchMethod with an Invocation whose isMethod is true.

There is another part of the language specification which is concerned with the behavior of the call getter, namely this location. However, these rules make no difference in this case because we're never invoking the getter named call.

It could be claimed that it is underspecified which value isGetter will have in the dynamic invocation. However, the only reasonable choice, I think, would be to follow the rule specified for the noSuchMethod forwarders. This means that isGetter will be false. Similarly, isSetter should be false because the given member name isn't a setter name (it doesn't end in =).

It might be surprising to some code that you can get a failed invocation with the same name as a getter.

Yes, I was thinking about that, too. However, it's probably the most consistent behavior, measured against the behavior of statically checked invocations.

[eernstg]

Turns out that DartPad (DDC) yields isMethod=true, but all other backends yield isMethod=false. So those other backends should be adjusted.

[alexmarkov]

VM behavior was fixed in 4305541.

[lrhn]

Closed too soon. Dart2js still prints false.