Incorrect derivatives when new arrays are constructed in computation

[jeffreyesun]

After much head-scratching about why my derivatives weren't coming out right, I discovered that weird things happen when I build new arrays by indexing from existing arrays. Sometimes, the derivatives are omitted completely:

gb = param(1,1; atype=Array{Float32})
function loss_test(b)
    loss = [b[1]]
    return sum(loss)
end

∇b = @diff loss_test(gb)
grad(∇b, gb) #Sparse(Matrix{Float32}(1,1)()) # should be [1.0;;]

Sometimes, they overwrite the derivatives of arrays they are added to:

gb = param(2,2; atype=Array{Float32})
function loss_test(bb)
    loss = sum(bb, dims=2)
    b = bb.*1
    loss += [b[i] for i=1:2]
    return sum(loss)
end

∇b = @diff loss_test(gb)
grad(∇b, gb) #[1.0 0.0; 1.0 0.0] # should be [2.0 1.0; 2.0 1.0]

If this kind of construction is known to be disallowed, then trying to do it should raise an error. If it's not known to be disallowed, then this is a bug.