const_eval_throws_exception error

[tvolkert]

Consider the work-in-progress changes in flutter/flutter#30646 in which I change Flutter's Row widget to have a const constructor.

In the associated test, I have the following code:

testWidgets('Row is const constructable', (WidgetTester tester) async {
  const Row(
    children: <Widget>[
      Text('Billable Hours'),
    ],
  );
});

When run over that code, the analyzer produces the following error:

  error • Evaluation of this constant expression throws an exception • test/widgets/row_test.dart:32:5 • const_eval_throws_exception

Yet when I run the test, it passes:

$ flutter test test/widgets/row_test.dart 
00:05 +28: All tests passed!                                  

• Dart VM version: 2.2.1-dev.3.0.flutter-None (Tue Apr 2 13:54:28 2019 +0000) on "macos_x64"

[tvolkert]

/cc @stereotype441 @scheglov

[vsmenon]

Do we know if this is a regression? (2.3 is already past due.)

[stereotype441]

Found the problem. It is not a regression. Explanation to follow.

[stereotype441]

Ok, there is an error in @tvolkert's code, which the analyzer successfully detects. Unfortunately, the analyzer does a pretty horrible job of reporting where the error is (which is why I had to spend a few hours this morning sleuthing to even figure it out). To make matters worse, the VM fails to detect the error at all, making it seem like the analyzer is at fault.

The error in @tvolkert's code is located here, in the constructor for the class Flex:

class Flex extends MultiChildRenderObjectWidget {
  const Flex({
    Key key,
    @required this.direction,
    this.mainAxisAlignment = MainAxisAlignment.start,
    this.mainAxisSize = MainAxisSize.max,
    this.crossAxisAlignment = CrossAxisAlignment.center,
    this.textDirection,
    this.verticalDirection = VerticalDirection.down,
    this.textBaseline,
    List<Widget> children = const <Widget>[],
  }) : assert(direction != null),
       assert(mainAxisAlignment != null),
       assert(mainAxisSize != null),
       assert(crossAxisAlignment != null),
       assert(verticalDirection != null),
       assert(crossAxisAlignment != CrossAxisAlignment.baseline || textBaseline != null), // ERROR
       super(key: key, children: children);

The line marked // ERROR is problematic because it attempts to compare two values of type CrossAxisAlignment. You can't do this safely in a const constructor because that assertion has to be able to be evaluated as part of constant evaluation, and constant evaluation won't allow you to do equality comparison on values of non-primitive types. (Note that once "constant update 2018" is enabled, the rules will be relaxed slightly: you'll be able to compare a value of a non-primitive type to null, but you still won't be able to compare two non-primitive values).

The reason the error is being reported at the site of the evaluation of const Row(...) is because Row extends Flex. Unfortunately, the analyzer doesn't report anywhere near enough information to help the user understand what went wrong. It just unhelpfully says that an exception was thrown. I'll file a separate bug report to track that.

The other problem that's happening here is that the VM does not properly detect the bug. It is allowing two enumerated values to be equality compared during constant evaluation without complaint. I'll file a separate bug report to track that.

[stereotype441]

Ok, I've filed #36526 to track the problem of the analyzer not giving enough information about the location of the error, and #36528 to track the problem of the VM failing to detect the error. Since the analyzer is correct in reporting an error, I'm closing this bug.

[dnfield]

Is this a case where if we used identical it would work? e.g. !(identical(crossAxisAlignment, CrossAxisAlignment.baseline))?

[stereotype441]

Is this a case where if we used identical it would work? e.g. !(identical(crossAxisAlignment, CrossAxisAlignment.baseline))?

@dnfield Yes! Sorry, I should have mentioned that. Switching to identical eliminates the error.

[dnfield]

Thanks! This seems relatedto #29278