cc @DougGregor

@rjmccall

The protocol-extension implementation of the typeDescription getter has this:

define hidden { i8*, i64, i64 } @_TZFeRq_1t15TypeDescribable_S_S0_g15typeDescriptionSS(%swift.type*, %swift.type* %Self, i8** %Self.TypeDescribable) #0 {
entry:
%Self1 = alloca %swift.type*, align 8
%1 = alloca %swift.type*, align 8
store %swift.type* %Self, %swift.type** %Self1, align 8
store %swift.type* %Self, %swift.type** %1, align 8
%2 = bitcast %swift.type** %1 to %swift.opaque*
%3 = call %swift.type* @swift_getMetatypeMetadata(%swift.type* %Self) #4

which looks like it's passing the 'Self' type of the witness table rather than the 'Self' type passed into this static member.

The problem is that the Self archetype is not being bound to the dynamic type of the class, and the fix is that protocol witness thunks of protocol extension methods for conformances by non-final classes need to open the dynamic type and use that in the substitution they use when invoking the extension method.

That's going to require some non-trivial implementation work, though.

Comment by Marcus Smith (JIRA)

I'm running into this exact same problem as well, has there been any progress on this?

I'm happy to help in any way I can, but my code looks almost identical to what's already in the gist in the description.

@swift-ci create

#12174