Here is my take for what it is worth as an observer of the Julia language:
I think your second method foo(x::Number) is unreachable because your first
method foo{T<:Number}(x::T) is more specific and there are no arguments to
which only the second method is applicable. Having both methods would make
more sense if the upper bound on T was a proper subtype of Number.
I guess that the first method is considered more specific because T could
be a proper subtype of Number. You won't get that from the documentation.
The key Julia concept of "more specific method" is completely undefined in
the documentation. I spent way too much time this spring trying to
reverse-engineer its actual definition. Eventually I concluded that it is
impossible to state clearly and succinctly what rules Julia currently
follows. It just does what it does, and now and then gets patched when
someone complains and the complaint seems to have merit.
This seems like an area where the language needs rethinking. Multiple
dispatch is great, and the transitive more specific method relation seems
like a good idea, but it needs to be based on semantics that are both easy
to understand and don't often produce surprising results. That doesn't
seem like an easy thing to figure out. Maybe someone should do a PhD
thesis on it.
Someone other than me should answer "When would I want to use one
definition over the other?"
On Thursday, July 3, 2014 3:16:41 PM UTC-4, Andrew McKinlay wrote:
>
> Why does
> foo{T<:Number}(x::T) = 1
> foo(x::Number) = 2
>
> println(foo(3))
> result in
> 1
> ?
>
> Is there any difference between the first method definition and the second?
>
> Does the first method completely shadow the second?
>
> When would I want to use one definition over the other?
>
