releaseFence needs to be added at the end of constructors of classes that contain mixin fields

[smarter]

Since #8652, fields in classes coming from traits are no longer marked final because they're set outside of the class constructor. When this was changed in scalac, they also had to add calls to releaseFence at the end of these constructors, this is justified with a comment:

              // If the field is mutable, it won't be final, so we can write to it in a setter.
              // If it's not, we still need to initialize it, and make sure it's safely published.
              // Since initialization is performed (lexically) outside of the constructor (in the trait setter),
              // we have to make the field mutable starting with classfile format 53
              // (it was never allowed, but the verifier enforces this now).

from scala/scala#7028, which we should port.

Tentatively assigning to @sjrd since unfortunately for him he has become our resident mixin expert :).

[smarter]

Also relevant is scala/scala@5e109ff from scala/scala#7270, but tha's only an optimization (if I understand correctly) and one that depends on #9181

[sjrd]

I won't have time to do that in the three weeks. If someone else wants to pick it up, I suggest to do that in Mixin. In transformTemplate, for a non-trait class, if we process any trait setter in that class, then add an additional statement at the end of the new Template's body that calls releaseFence(). Then Constructors will naturally move that statement to the body of the constructor.

[anatoliykmetyuk]

Is it still a blocker for 3.0.0-M1? @smarter ?

[smarter]

We can move it to M2