Fix mapping of annotations

[odersky]

We previously did not do an (expensive) TreeTypeMap on an annotation if the mapped versions of all types of subtrees of the annotation tree were =:= to the original types. But it turns out this is too coarse. In the test we have capture set variables where we intend to map a TypeRef to a TypeParamRef but the two were considered as =:= because of the bounds they had. So no mapping took place.

We now use eql instead of =:=, which is structural comparison with eq for references to corresponding binders.

[odersky]

@mbovel You looked at this code last, and I remember we discussed several alternatives. Do you see a reason why eql would not work? I know I was skeptical about changing this before, but this failure convinced me that comparing types with frozen_=:= is not correct.

[mbovel]

@mbovel You looked at this code last, and I remember we discussed several alternatives. Do you see a reason why eql would not work? I know I was skeptical about changing this before, but this failure convinced me that comparing types with frozen_=:= is not correct.

I concluded in #19957 (comment) that tp1 frozen_=:= worked for comparing TermRef and TermParamRef. I remember also doing some tests with TypeRef and TypeParamRef, but I guess I didn't stumble across the situation you have here.

Using eql sounds okay to me.

[odersky]

It was a very special setup where we now have TypeParam/Refs that might contain each other in their lower as well as upper bounds. That's what made frozen_=:= pass. It was a recent improvement to capture set variables by @noti0na1. Which shows that no good deed goes unpunished 😜 .