Profiling shows it as a bottleneck on subgraph scans. Find a way to make it faster or not enter it in some cases.