If you're not doing recursion, I prefer using an appropriately sized thread_local buffer in this scenario. Saves you the allocation and does the bookkeeping of having one per thread