Poor performance of vcat/hcat/hvcat on mixed input types

[mikmoore]

From this Discourse thread.

On version 1.8.0 (sorry it's what i have handy, but the offending method appears unchanged on current master):

julia> using BenchmarkTools

julia> x = 1; y = 2:2;

julia> @btime vcat($x, $x);
  26.506 ns (1 allocation: 80 bytes)

julia> @btime vcat($y, $y);
  28.744 ns (1 allocation: 80 bytes)

julia> @btime vcat($x, $y);
  2.667 μs (35 allocations: 1.05 KiB)

with similar results for hcat and hvcat.

An initial inspection suggests that the offending methods for vcat/hcat are LinearAlgebra._vcat(A::Union{Number, LinearAlgebra.UniformScaling, AbstractVecOrMat}...; array_type) and the matching _hcat, which are (as of this writing) defined at uniformscaling.jl#L426. A loop over the input vararg tuple appears to cause dynamic dispatch.

hvcat has the same problem, likely arising from a similar loop in uniformscaling.jl#446.

[dkarrasch]

I guess I'm and #41394 are to blame. GitHub says that will only released as v1.9, but I do find the corresponding code in v1.8.x. I think I have a solution to this, will make a PR soon.