bug in dispatch

[CarloLucibello]

The following code works correctly

 julia> f{T<:Number}(x::T, y...) = (println(length(y)); f(y[1], y[2:end]...))
f (generic function with 1 method)

julia> f{T<:Number}(x::T) = "hi"
f (generic function with 2 methods)

julia> f(1,rand(10)...)
10
9
8
7
6
5
4
3
2
1
"hi"

But it fails when I change to <:Union the parametrization

julia> f{T<:Union{Float32,Float64}}(x::T, y...) = (println(length(y)); f(y[1], y[2:end]...))
f (generic function with 1 method)

julia> f{T<:Union{Float32,Float64}}(x::T) = "hi"
f (generic function with 2 methods)

julia> f(1.,rand(10)...)
10
9
8
7
6
5
4
3
2
1
0
ERROR: BoundsError: attempt to access ()
  at index [1]
 in f(::Float64) at ./REPL[1]:1 (repeats 11 times)
 in eval_user_input(::Any, ::Base.REPL.REPLBackend) at ./REPL.jl:64
 in macro expansion at ./REPL.jl:95 [inlined]
 in (::Base.REPL.##3#4{Base.REPL.REPLBackend})() at ./event.jl:68

At the last step we have a wrong dispatch to the first method instead of the second

Bye,
Carlo

[nalimilan]

That's because y... can match zero arguments too. Change the first function to this (which is actually more natural here):

f{T<:Union{Float32,Float64}}(x::T, y, z...) = (println(length(z)+1); f(y, z...))

[nalimilan]

That said, a redefinition warning should probably be printed in that case.

[CarloLucibello]

isn't the vararg signature less specialized than the other one, so that f(x) should always be preferred to f(x, y...)?

Also, the fact that T<:Number works and T<:Union is clearly an incoherent behaviour (bug?)

[Gnimuc]

I'm curious: what's the subtype of Union{Float32,Float64}? is it necessary to use parametric method here? Why not just write f(x::Union{Float32,Float64}, y...) = (println(length(y)); f(y[1], y[2:end]...))?

FYI, using typealias will suppress the error:

julia> typealias F32F64 Union{Float32, Float64}
Union{Float32,Float64}

julia> f{T<:F32F64}(x::T) = "hi"
f (generic function with 1 method)

julia> f{T<:F32F64}(x::T, y...) = (println(length(y)); f(y[1], y[2:end]...))
f (generic function with 2 methods)

julia> f(1.,rand(10)...)
10
9
8
7
6
5
4
3
2
1
"hi"

[yuyichao]

I think this is a dispatch error.

This is different from the default argument case and there are only two methods and the second one should always be more specific than the first one. Of course, defining it the way @nalimilan mentioned above is still preferred since it's easier to handle for the compiler and is less likely to cause ambiguity when there's other methods.

[CarloLucibello]

just to point to a real case scenario, I encountered this issue here JuliaCollections/Iterators.jl#85 while trying to define iteratorsize for chained iterators

[JeffBezanson]

I believe this is fixed by #18457 - @CarloLucibello can you confirm?

[CarloLucibello]

It's working now, nice! @JeffBezanson

julia> f{T<:Union{Float32,Float64}}(x::T, y...) = (println(length(y)); f(y[1], y[2:end]...))
f (generic function with 1 method)

julia> f{T<:Union{Float32,Float64}}(x::T) = "hi"
f (generic function with 2 methods)

julia> f(1.,rand(10)...)
10
9
8
7
6
5
4
3
2
1
"hi"