What I'm not sure of is what behavior should we allow, there was a short discussion here of what is allowed in the setindex of an array.
https://julialang.slack.com/archives/C6GHSTN4T/p1641410970153500?thread_ts=1641410970.153500&cid=C6GHSTN4T

But basically this is allowed, even tho it errors on broadcasting operations.

 julia> z = zeros(Int,4,4)

 
 julia> z[3, :] = 10ones(2,2);
 
 julia> z
 4×4 Matrix{Int64}:
   1   1   0   0
   1   1   0   0
  10  10  10  10
   0   0   0   0

I see. That one is allowed here. As you say in the Slack thread, that (w/sh)ould be prohibited by Base.setindex_shape_check, right? I quickly checked, on your branch this example (with z = spzeros(Int, 4, 4)) leads to the original segfault.

That's the thing, when I made my branch I thought it errored and that only (1,n,1...) arrays were allowed, but apparently not.

Since that is the current behaviour of normal arrays I guess sparse arrays should follow it, but IMO that should be considered a bug. In fact if it were up to me setindex should only be allowed if the shapes matched or if RHS is a vector.

Shall we go with this one then @gbaraldi? The fix of the shape issue is independent and has to be made in a different place, such that this setindex! method will benefit automatically.

Yeah it's fine by me, it segfaulted before, so nobody will miss the weird behaviour anyway.

Can you add the backport labels so it gets backported? They were present in my original PR so I imagine they should be here too.