These Shape_i nodes will not always be optimized away otherwise, and
they prevent local_shape_to_shape_i from being applied.
Instead, we manually apply the equivalent of local_shape_to_shape_i,
by using shape_of[...] directly.