Simplify Removes Arrays Used Only For Indirect Accesses

[ThrudPrimrose]

I have a case of simplify failing.

This SDFG computes:

@dace.program
def scatter_store(A: dace.float64[N], idx: dace.int64[N], B: dace.float64[N], scale: dace.float64):
    for i, in dace.map[0:N:1]:
        B[idx[i]] = A[i] * scale

Then I have a vectorized variant that does:

@dace.program
def scatter_store(A: dace.float64[N], idx: dace.int64[N], B: dace.float64[N], scale: dace.float64):
    for i, in dace.map[0:N:8]:
        B_buf[0:8] = A[i:i+8] * scale
        indices[0:8] = idx[i:i+8]
        B[indices[0:8] = B_buf[0:8]

In this case idx is used only on an interstate edge and it gets removed.
This is the pattern that causes it:

SDFGs are attached:
scatter_store.sdfg.json

scatter_store_vec.sdfg.json

[tbennun]

First of all, this is not a failure if the code is valid. Is it not?
This is because it becomes a view.
You should use np.copy

[ThrudPrimrose]

First of all, this is not a failure if the code is valid. Is it not? This is because it becomes a view. You should use np.copy

I do not generate the vectorized variant from the Python frontend. I do it through transformations. I put the numpy-like code as it is easier to understand. Simplify removes the __tmp_508_10_r array and the corresponding symbols that are assigned from it.

Btw this was just a reminder for me to not forget, I will debug this in more detail after the ICS submission.