Asymmetric generic signatures for HKT in covariant/contravariant types

[WojciechMazur]

Based on apache/pekko#1474

In Scala 2.13 all generic param types of HKT (both covariant and invariant) are transformed into wildcard types.
In 3.3.4-RC1 / 3.5.0 contravariant types are preserved but erased.
In 3.3.3- both contravariant and covariant types were preserved and not erased

Compiler version

3.3.4 / 3.5.0

Minimized code

import scala.annotation.unchecked.uncheckedVariance

trait SubFlowDef[+F[+_]]
final class Flow1[-In]{
  type Repr[+O] = Flow1[In @uncheckedVariance]
}
final class Flow2[+Out]{
  type Repr[+O] = Flow2[O]
}
class SubFlow[In, Out](
    delegate1: SubFlowDef[Flow1[In]#Repr],
    delegate2: SubFlowDef[Flow2[Out]#Repr] 
)

Output

Based on javap outputs

Signatures:
Scala 3.3.4-RC1 / 3.5.0

  public SubFlow(SubFlowDef<?>, SubFlowDef<Flow2>);
    descriptor: (LSubFlowDef;LSubFlowDef;)V
    Signature: #9                           // (LSubFlowDef<*>;LSubFlowDef<LFlow2;>;)V

Scala 2.13

  public SubFlow(SubFlowDef<?>, SubFlowDef<?>);
    descriptor: (LSubFlowDef;LSubFlowDef;)V
    Signature: #12                          // (LSubFlowDef<*>;LSubFlowDef<*>;)V

Scala 3.3.3

  public SubFlow(SubFlowDef<Flow1<In>>, SubFlowDef<Flow2<java.lang.Object>>);
    descriptor: (LSubFlowDef;LSubFlowDef;)V
    Signature: #9                           // (LSubFlowDef<LFlow1<TIn;>;>;LSubFlowDef<LFlow2<Ljava/lang/Object;>;>;)V

Expectation

Probably both covariant and contravaraint types should have been replaced with wildcard types in generic signatures

[WojciechMazur]

The bisect shows that behaviour between 3.5.0 and 3.3.3 changed in #19931
The change in typer led to a change in generic signatures emitted by the backend.

[WojciechMazur]

The current generic signature for covariant HKT is incorrect (LSubFlowDef<LFlow2;>) and is invalid, it should be never emitted. We need to either restore old behaviour or emit wildcard types instead.
It means that both 3.5.0 and 3.5.1 releases are producing invalid generic signatures. We should fix in the next release

The PR introducing the change to LTS was reverted (#21609)