[dart2wasm] co19/LanguageFeatures/int-to-double/assignment_class_member_t01,2 fail on dart2wasm about sdk

[osa1]

Smaller repro:

class C {
  double m1 = 42;
}

main() {
  C c = new C();
  c?.m1 ??= 42;
}

This looks similar to #55341 which also had an unnecessary null check.

[osa1]

I wonder if this is a TFA bug, post-TFA kernel looks wrong. Before TFA:

class C extends core::Object {
  field core::double m1 = 42.0;

  synthetic constructor •() → test::C
    : super core::Object::•()
    ;

  static method _#new#tearOff() → test::C
    return new test::C::•();
}

static method main() → dynamic {
  test::C c = new test::C::•();

  let final test::C #t1 = c
    in #t1 == null
         ?{core::double?} null
         : #t1.{test::C::m1}{core::double} == null ?{core::double} #t1.{test::C::m1} = 42.0 : null;
}

After TFA:

static method main() → dynamic {
  test::C c = new test::C::•();
  let final test::C #t1 = c in _in::unsafeCast<core::double?>(_in::unsafeCast<core::double>( block {
     #t1.{test::C::m1}{core::double};
  } =>null));
}

The problematic part is unsafeCast<double>(block { ... } => null), which tries to cast null to double.

@alexmarkov is my reading of the kernel correct? Could this be a TFA bug?

[alexmarkov]

Kernel AST before TFA doesn't look correct:

?{core::double} #t1.{test::C::m1} = 42.0 : null

The static type of this ?: is double, while it can result in null. When pruning this expression, TFA keeps its static type by inserting unsafeCast (as changing the static type may alter semantics of enclosing expressions).

So the problem is likely in CFE or in another transformation prior to TFA.

/cc @johnniwinther @chloestefantsova

[johnniwinther]

@chloestefantsova Is this the same problem that you fixed today?

[osa1]

Note: d4d7884 didn't fix this issue, repro works with the current main (671f271).