The current implementation performs a deep copy of Renderer::m_data every time there is a loop statement in the template. This copying becomes a significant overhead with large input data and complex templates.
This change introduces Renderer::m_loop_data which is used to track the loop context and avoids the deep copy.
The updated benchmark shows no difference in performance between the small and large input data. If this benchmark is run against the original code, rendering the large data takes 10x longer than the small data. In real-world scenarios I have observed up to 75x performance improvement with this change.