iris: Use render_batch/compute_batch locals in memory_barrier