Aycock & Horspool came up with a 'practical' method for implementing Earley parsing (conversion to a state-machine) that has pretty humorously good performance delta over "naive" Earley, and is still reasonable to implement. Joop Leo figured out how to get the worst-case of Earley parsing down to either O(n) (left-recursive, non-ambiguous) or O(n^2) (right-recursive, non-ambiguous). That means the Earley algorithm is only O(n^3) on right-recursive, ambiguous grammars; and, if you're doing that, you're holding your language wrong.
A somewhat breathless description of all of this is in the Marpa parser documentation:
https://jeffreykegler.github.io/Marpa-web-site/
In practice, I've found that computers are so fast, that with just the Joop Leo optimizations, 'naive' Earley parsing is Good Enough™:
A somewhat breathless description of all of this is in the Marpa parser documentation:
In practice, I've found that computers are so fast, that with just the Joop Leo optimizations, 'naive' Earley parsing is Good Enough™: