promote_rule with automatic return type casting?

[RainerHeintzmann]

I am wondering how to make a julia type, which behaves exaclty like existing array classes but (for example) uses a different show function. Ideally I would not want to rewrite all existing methods but automatically all exiting methods should also apply to my new datatype.

In OOP based thinking this is only very few lines of code. However, in Julia it seems next to impossible to achieve this. On the web I found several implementations, which redefined all operations on arrays (e.g. +,-,.*) in a more or less automated way, but there really should be a simple solution to express this.

Here is a bit of code that demonstrates my problem:

type Img{T,N} <: DenseArray{T,N}  # If you derive from DenseArray, then at least the plus works after defining it. However promotion seems to be ignored
    data::DenseArray{T,N}
end
a=[4 5 6 75 34.2];
b=Img(a);
b.data

Base.ndims{T,N}(animg::Img{T,N}) = Base.ndims(animg.data)
Base.size{T,N}(animg::Img{T,N}) = Base.size(animg.data)
Base.getindex{T,N}(animg::Img{T,N}, anind::Real) = Base.getindex(animg.data, anind)
Base.getindex{T,N}(animg::Img{T,N}, anind::DenseArray{T,N}) = Base.getindex(animg.data, anind)
Base.getindex{T,N}(animg::Img{T,N}, anind::AbstractArray{T,N}) = Base.getindex(animg.data, anind)
Base.getindex{T,N}(animg::Img{T,N}, varargs...) = Base.getindex(animg.data, varargs...)
Base.setindex!{T,N}(animg::Img{T,N}, varargs...) = Base.setindex!(animg.data, varargs...)

Base.similar{T,N}(animg::Img{T,N},::Type{T}) = Base.similar(animg.data,varargs...)
Base.similar{T,N}(animg::Img{T,N},varargs...) = Base.similar(animg.data,varargs...)
c=b+b;
c
# This is now a  1x5 Array{Float64,20}, but I really want it to be of type Img{Float64,2}
# How can this be achieved?

Interestingly assignments do work, if c is first defined as an Img{Float64,2] datatype.
So how can one enforce that creation of new variables will have to Img datatype rather than the Array datatype? Or is this behavior a bug after all?

An alternative might be that promote_rule is somewhat extended, such that julia can know under what circumstances a result of a function call should automatically be cast to the type defined in promote_rule. E.g. one could use a tupel where the second type is the one to which the returned value should be cast to:

promote_rule{T,N}(::Type{Img{T,N}},::Type{Img{T,N}} ) = (Type{Img{T,N}},Type{Img{T,N}})
promote_rule{T,N}(::Type{Img{T,N}},::Type{DenseArray{T,N}} ) = (Type{Img{T,N}},Type{Img{T,N}})

Would this be doable?
Any ideas?

[JeffBezanson]

See also #2248

[JeffBezanson]

I think the promotion issue is the same as #2326. However, your similar method should return something of the same type as animg; that's what + is using to make its result.

This is captured by existing issues; closing as dup.