Engineering High-Performance Parsers with Data-Oriented Design

What are the downsides of this design compared to the traditional way?

Doesn't deallocation only happen once, when the work is completed, and is it worth optimizing?