bug: Class constructor <class_name> cannot be invoked without 'new' about sdk

[lrhn]

Probably a name mangling issue.

[sigmundch]

Thanks @Diaga for the very nice repro!

Here is a simplified example illustrating the problem in a standalone application:

class B<T> {}
class B$ {}

main() => B<int>();

It appears DDC introduces a helper declaration when defining generic classes, but it is not checking whether the name for the helper conflicts with other definitions in the same module. In the example above, it introduces a helper called B$ to represent the generic instantiation of B, but B$ is also the declaration of the second class. When the latter overrides the former, it makes uses of the former invalid (like the instantiation of B<int>())

[sigmundch]

It's worth noting that this also happens with other conflicting declarations. For example:

class B<T> {}
final B$ = 0;
main() => B<int>();

So it's not sufficient to look for conflicts among classes. The error is slightly different in this case, but the cause is the same.

[Markzipan]

DDC emits auxiliary classes (and other top level identifiers) for generics and other structures with a $ suffix (see here). Lemme take a look at making these temporary Ids instead, which obey the built-in renaming semantics.

Thanks for the great repros, everyone!

[Markzipan]

Closing the loop here, DDC implicitly emits a generic object constructors as their original class name + $. The renaming logic here is a tricky to clean up (we can't just use temporaries since this "API" needs to be stable across our runtime). The workaround is simple (i.e., avoid creating top level elements named X$ if X is a generic class), but we can obfuscate the name in the short term if this becomes a blocker.

[lrhn]

Consider working towards a fix. I expect names with $ in then may get more prevalent with macros.
(You could change all $s in the original name to $$_ and then any $ you insert yourself will not cause it to be equal to a source $).