Strong mode analyzer inconsistent about when types must match

[Hixie]

Consider this code:

class A { }
class B extends A { }

void test<T>(T v) { }

void main() {
  final A x = new A();
  test<B>(x); // no error

  test<B>(new A()); // strong_mode_invalid_cast_new_expr
}

Why are these different?

[Hixie]

For the record, I would have expected an error in both cases.

[bwilkerson]

@leafpetersen

[leafpetersen]

You've asked me about this before. :) Our previous discussion of which led to these two issues being filed:

#29548
#29547

Strong mode makes implicit downcasts an error when it can immediately prove that the cast will fail. It's not smart enough to track the fact that x is definitely an A and not a B in your example, but it does notice when you try to cast a fresh instance of A to a B.

The --no-implicit-casts flag will make both of those an error.

[Hixie]

Why would we not make implicit downcasts an error in general? They're not sound, right? I thought the point was to be sound.

[Hixie]

So as @leafpetersen said, you can turn off implicit casts, that's implicit-casts: false in your analysis options. Unfortunately that isn't ready for prime time yet. Filed #30402 on some issues I found trying to turn that on.

[jmesserly]

removing implicit downcasts is now tracked in #31410