Matrix difference extremely slow

[biona001]

Hi,

I noticed that a BlockDiagonal minus a dense matrix is very slow with v.0.1.35

bd = BlockDiagonal([rand(5, 5) for _ in 1:100])
dense1 = rand(500, 500)
dense2 = rand(500, 500)

@time bd - dense1
@time dense2 - dense1;

  0.219396 seconds (495.00 k allocations: 134.660 MiB, 16.69% gc time)
  0.000613 seconds (2 allocations: 1.907 MiB)

But interestingly, addition and multiplication is okay (in fact blockdiagonal is faster, as I would expect)

@time bd + dense1
@time dense2 + dense1;

  0.000377 seconds (102 allocations: 1.935 MiB)
  0.000646 seconds (2 allocations: 1.907 MiB)

@time bd * dense1
@time dense2 * dense1;

  0.002117 seconds (800 allocations: 3.850 MiB)
  0.008553 seconds (2 allocations: 1.907 MiB)

Might be related to #101

[mzgubic]

Thanks for opening these performance issues, please keep them coming 🙂