Inference failure for typed concatenation

[GunnarFarneback]

The typed concatenation T[a:b] is deprecated in favor of T[a:b;]. However, the latter fails to infer result type correctly.

   _       _ _(_)_     |  A fresh approach to technical computing
  (_)     | (_) (_)    |  Documentation: http://docs.julialang.org
   _ _   _| |_  __ _   |  Type "?help" for help.
  | | | | | | |/ _` |  |
  | | |_| | | | (_| |  |  Version 0.4.0-rc2+10 (2015-09-20 15:38 UTC)
 _/ |\__'_|_|_|\__'_|  |  Commit d8930a0* (0 days old release-0.4)
|__/                   |  x86_64-linux-gnu

julia> f(n) = Float64[1:n]
f (generic function with 1 method)

julia> g(n) = Float64[1:n;]
g (generic function with 1 method)

julia> f(2)
WARNING: T[a:b] concatenation is deprecated; use T[a:b;] instead
 in depwarn at ./deprecated.jl:73
 in getindex at ./deprecated.jl:662
 in f at ./none:1
while loading no file, in expression starting on line 0
2-element Array{Float64,1}:
 1.0
 2.0

julia> g(2)
2-element Array{Float64,1}:
 1.0
 2.0

julia> Base.return_types(f, (Int,))
1-element Array{Any,1}:
 Array{Float64,1}

julia> Base.return_types(g, (Int,))
1-element Array{Any,1}:
 Array{Float64,N}

[GunnarFarneback]

This is actually a regression from 0.3.

   _       _ _(_)_     |  A fresh approach to technical computing
  (_)     | (_) (_)    |  Documentation: http://docs.julialang.org
   _ _   _| |_  __ _   |  Type "help()" for help.
  | | | | | | |/ _` |  |
  | | |_| | | | (_| |  |  Version 0.3.10
 _/ |\__'_|_|_|\__'_|  |  
|__/                   |  x86_64-linux-gnu

julia> g(n) = Float64[1:n;]
g (generic function with 1 method)

julia> Base.return_types(g, (Int,))
1-element Array{Any,1}:
 Array{Float64,1}

[GunnarFarneback]

Let's widen the scope.

f1() = [1]
f2() = [1;]
f3() = [1;2]
f4() = [1;2.0]
f5() = [1 2]
f6() = [1 2.0]
f7() = Int[1]
f8() = Float64[1]
f9() = Int[1;]
f10() = Float64[1;]
f11() = Int[1;2]
f12() = Float64[1;2]
f13() = Int[1;2.0]
f14() = Int[1 2]
f15() = Float64[1 2]
f16() = Int[1 2.0]
f17() = [1:2;]
f18() = Int[1:2;]
f19() = Float64[1:2;]
f20() = [1:2;1:2]
f21() = Int[1:2;1:2]
f22() = Float64[1:2;1:2]
f23() = [1:2;1.0:2.0]
f24() = Int[1:2;1.0:2.0]
f25() = [1:2 1:2]
f26() = Int[1:2 1:2]
f27() = Float64[1:2 1:2]
f28() = [1:2 1.0:2.0]
f29() = Int[1:2 1.0:2.0]
const A = zeros(Int, 2, 2)
const B = zeros(Float64, 2, 2)
f30() = [A;]
f31() = Int[A;]
f32() = Float64[A;]
f33() = [A;A]
f34() = Int[A;A]
f35() = Float64[A;A]
f36() = [A;B]
f37() = Int[A;B]
f38() = [A A]
f39() = Int[A A]
f40() = Float64[A A]
f41() = [A B]
f42() = Int[A B]

for k = 1:42
    f = eval(symbol(string("f", k)))
    try
        f()
        println(k, " ", Base.return_types(f, ())[1])
    catch
        println(k, " error")
    end
end

Many of the result types are not concrete and some of them are regressions from 0.3.

	0.3.10	0.4.0-rc2+31	Regression
1	Array{Int64,1}	Array{Int64,1}
2	Array{Int64,1}	Array{Int64,1}
3	Array{Int64,1}	Array{Int64,1}
4	Array{T,1}	Array{Float64,1}
5	Array{Int64,2}	Array{Int64,2}
6	Array{T,2}	Array{Float64,2}
7	Array{Int64,1}	Array{Int64,1}
8	Array{Float64,1}	Array{Float64,1}
9	Array{Int64,1}	Any	x
10	Array{Float64,1}	Any	x
11	Array{Int64,1}	Any	x
12	Array{Float64,1}	Any	x
13	Array{Int64,1}	Any	x
14	Array{Int64,2}	Any	x
15	Array{Float64,2}	Any	x
16	Array{Int64,2}	Any	x
17	Array{Int64,1}	Array{Int64,1}
18	Array{Int64,1}	Array{Int64,N}	x
19	Array{Float64,1}	Array{Float64,N}	x
20	Array{Int64,1}	Array{Int64,1}
21	Array{Int64,N}	Array{Int64,N}
22	Array{Float64,N}	Array{Float64,N}
23	Array{Float64,N}	Array{Float64,N}
24	Array{Int64,N}	Array{Int64,N}
25	Array{Int64,2}	Array{Int64,2}
26	error	Array{Int64,N}
27	error	Array{Float64,N}
28	Array{Float64,N}	Array{Float64,N}
29	error	Array{Int64,N}
30	Array{Int64,2}	Array{Int64,2}
31	error	Array{Int64,N}
32	error	Array{Float64,N}
33	Array{Int64,2}	Array{Int64,2}
34	error	Array{Int64,N}
35	error	Array{Float64,N}
36	Array{Float64,N}	Array{Float64,N}
37	error	Array{Int64,N}
38	Array{Int64,2}	Array{Int64,2}
39	error	Array{Int64,N}
40	error	Array{Float64,N}
41	Array{Float64,N}	Array{Float64,N}
42	error	Array{Int64,N}

[JeffBezanson]

Thanks for the thorough analysis here.

[tkelman]

Just a note for posterity (ref #13834 (comment)) that this change was responsible for a roughly 15% slowdown in the rand_mat_stat benchmark.

[KristofferC]

Worth creating a separate issue for?