Unification results for index expressions are non-deterministic

[SamirDroubi]

I have noticed on multiple occasions that unification results are non-deterministic. For example, unifying some access A[i, j] with reg :[f32][1] might be A[i:i+1][j] or A[i,j:j+1]. Here is an another example where the codegen results is different:

ssyrk_rm_l_stride_1_mk_A_pack(ctxt,-24 * ((23 + N) / (24)) + -(24 * io) + N + 24,512,(struct exo_win_2f32c){ &A.data[(24 * ((23 + N) / 24) + 24 * io + -24) * (A.strides[0]) + 512 * ko], { A.strides[0], 1 } },packed_A);
ssyrk_rm_l_stride_1_mk_A_pack(ctxt,-24 * ((23 + N) / (24)) + N + -(24 * io) + 24,512,(struct exo_win_2f32c){ &A.data[(24 * ((23 + N) / 24) + 24 * io + -24) * (A.strides[0]) + 512 * ko], { A.strides[0], 1 } },packed_A);

Most of the times, we get around these issues by simplifying (which would sort the operands during normalization resulting in a deterministic output), but sometimes we hit some edge cases that our simplification doesn't handle and you end up in a similar case as the above.

The main problem with this is that it makes is hard to write tests that checks the output of scheduling/codegen.