large performance regression in obtaining sparse array from a transpose of a sparse array

[ExpandingMan]

using Compat.SparseArrays, Compat.Random
using BenchmarkTools

const N = 10^4

const Nnz = 5*10^6

A = zeros(Float64, N,N)

for i ∈ 1:Nnz
    A[rand(1:size(A,1)),rand(1:size(A,2))] = 1.0
end

A = sparse(A)

On 0.6.4:

julia> @btime sparse(transpose(A));
  292.237 ms (15 allocations: 149.00 MiB)

On 1.0.0

julia> @btime sparse(transpose(A));
  9.729 s (40 allocations: 240.79 MiB)

This scales very badly with the size of the array. This makes life extremely difficult if you want to get the "I, J, V" values of the resulting transpose.

In the meantime any advice on a work around would be hugely appreciated.

[KristofferC]

Using SparseMatrixCSC(A), should be fast (#28049). Probably need to hook up sparse in the same way.

[ExpandingMan]

Ah yes! I realized that literally just 2 minutes before you posted this, thanks!

That's a huge relief, it might have been a lot of work to get around this. Ok, I'll try to make a PR if the fix is straight-forward enough.

[KristofferC]

Fixed by #28954