NNBD: I-2-b works in different ways in analyzer and dart for the function with unused type parameter.

[iarkh]

Dart VM version: 2.9.0-3.0.dev (dev) (Thu Apr 23 09:59:01 2020 +0200) on "windows_x64"

The following source code causes a compile error in analyzer and runtime error in dart:

typedef F<X> = void Function<Y extends X>();

F<X> toF<X>(X x) => throw "It's runtime!";
Type typeOf<X>() => X;

typedef G<X extends num> = void Function();

main() {
  G? source;
  var fsource = toF(source);
  F<G<num>?>? target1 = fsource;
  F<G<Object>?>? target2 = fsource;
}

It seems like both tools should behave in a similar way here.

Sample output is:

$> dartanalyzer --enable-experiment=non-nullable test.dart
Analyzing test.dart...
error - 'Object' doesn't extend 'num'. - test.dart:12:7 - type_argument_not_matching_bounds
hint - The value of the local variable 'target1' isn't used. - test.dart:11:15 - unused_local_variable
hint - The value of the local variable 'target2' isn't used. - test.dart:12:18 - unused_local_variable
1 error and 2 hints found.

$> dart --enable-experiment=non-nullable test.dart
Unhandled exception:
It's runtime!
#0 toF (file:///D:/DART/sdk/tests/co19/src/LanguageFeatures/Simple-bounds/static/test.dart:3:21)
#1 main (file:///D:/DART/sdk/tests/co19/src/LanguageFeatures/Simple-bounds/static/test.dart:10:17)
#2 _startIsolate. (dart:isolate-patch/isolate_patch.dart:301:19)
#3 _RawReceivePortImpl._handleMessage (dart:isolate-patch/isolate_patch.dart:168:12)

[mraleph]

Looks like a missing compile time error in the CFE

/cc @johnniwinther @stefantsov

[johnniwinther]

This is triggering a latent bug in the CFE and is not a priority for Null Safety Tech Preview.

[leafpetersen]

Should this be separated from the null safety release or do we want to have this fixed for stable.

[eernstg]

The analyzer treats the unused type variable as if it were covariant (which is also what is proposed in this PR). So there's an error with G<Object> because that's a bound violation and not a correct super-bounded type, but G<Object?> is accepted. All good.

For the CFE it looks like it hasn't been updated to use Object? rather than Object when it checks whether a given substitution should be performed during the check for being a correct super-bounded type. I would expect that to be needed as part of the null-safety release. That's addressed directly by #42429.

I think that makes this issue a duplicate of #42429, and I'll close this one because #42429 addresses the remaining issue more directly.