crystal choose wrong initialize method

[kostya]

class A
  def initialize
  end
end

class A::B(T) < A
  def initialize(@v : T)
  end
end

class A::C::B(T) < A
  def initialize(@v : T, @v2 : Int32)
  end
end

A::C::B(Int32).new(1, 10)

wrong number of arguments for 'A::B(Int32)#initialize' (given 2, expected 1)
Overloads are:
 - A::B(T)#initialize(v : T)

  def initialize(@v : T, @v2 : Int32)
  ^

[asterite]

Closing as duplicate of #2665 (generics inheritance is still broken)

[HertzDevil]

A::C::B.new is expanded into the following:

class A::C::B(T) < A
  def self.new(v : T, v2 : Int32)
    obj = B(T).allocate
    obj.initialize(v, v2)
    ::GC.add_finalizer(obj) if obj.responds_to?(:finalize)
    obj
  end
end

Here path lookup for B(T) in the constructor returns A::B(T) because for some reason lookup proceeds in A::C::B's superclass first. It actually can happen without generics at all:

class A
  class B
  end

  module C
    class B < A
      def self.foo
        # lookup tries the following types in the given order:
        # B relative to ::A::C::B (direct namespace child)
        # B relative to ::A (ancestors)
        # B relative to ::A::C (enclosing namespace)
        B
      end
    end
  end
end

A::C::B.foo # => A::B
# same code in Ruby produces A::C::B

We can either use the fully qualified name (::A::C::B(T)) in the expanded new, or change the path lookup rules to always consider the current scope before the parent types; that is, only for the first component of the path, if a direct child cannot be found but this component is equal to the last component of the current scope's name, we use that immediately and skip the parents. (The namespace lookup will always succeed and find the current scope again, but that is skipped here too.)