Ha! It seems all I needed was a type-assertion at the end. Both of these
work the way I want them to:
construct_instance{T<:Tuple}(::Type{T}) = tuple([construct_instance(t) for t in
T.parameters]...)::T
construct_instance{T<:Tuple}(::Type{T}) = ntuple(i ->
construct_instance(T.parameters[i]), length(T.parameters))::T
// T
On Monday, October 5, 2015 at 11:33:25 AM UTC+2, Tomas Lycken wrote:
I managed to get rid of the list comprehension by using ntuple:
>
> construct_instance{T<:Tuple}(::Type{T}) = ntuple(i ->
> construct_instance(T.parameters[i]), length(T.parameters))
>
> However, this suffers from the same instability problem (because of the
> anonymous function?). I’ll keep trying, but I would gladly accept any
> advice on how to do this.
>
> // T
>
> On Monday, October 5, 2015 at 10:28:15 AM UTC+2, Tomas Lycken wrote:
>
> I’ve tried to do the following
>>
>> construct_instance{T}(::Type{T}) = T()
>> construct_instance{T<:Tuple}(::Type{T}) = tuple([construct_instance(t) for t
>> in T.parameters]...)
>>
>> as a way to dynamically and recursively create an instance of a tuple
>> type. The tuple type might be nested, it might be any length > 0, and all
>> (supported) non-tuple types in the hierarchy are leaf-types with an empty
>> constructor defined.
>>
>> This works insofar as it gives me the tuple I want back, but it’s not
>> type stable; the inferred return type is Tuple rather than T. I also
>> tried wrapping it in convert(T, tuple(...)), but that only gave me e.g.
>> Tuple{Any,
>> Any} for a two-tuple type, which I assume is because type info is lost
>> in the list comprehension.
>>
>> Is there a way to help type inference realize that it will get a T back,
>> even when T is a tuple type?
>>
>> // T
>>
>>
>
>
