This is probably known somewhere, but something I noticed is that a way to think about Earley parsing is that, when you create an LR(1) shift-reduce parser, rather than having it be an error in the grammar for there to be shift-reduce or reduce-reduce conflicts, instead you nondeterministically reduce (i.e., your state machine has a set of states rather than a requiring there to be a single state). This also ends up solving the left recursion problem in an efficient way.

This is certainly slower than a deterministic LR(1) parser, but maybe if you're clever about the implementation you can efficiently handle local ambiguities in a grammar. Large-scale ambiguity is bad since you don't want small changes to have action-at-a-distance, but you could in principle have a strict LR(1) grammar that contains Earley sub-grammars for expressions.

> LL or LR is almost certainly faster then Earley, since in general more restricted = faster.

Earley's algorithm should be able to be trivially parallelized which might make it a contender for languages which are ambiguous (like most real world languages) where they are hand-rolling parsers. I haven't tried since I have no real need but looking at the steps I can't see a reason it couldn't do its work across multiple threads.

Honestly, other than JavaScript 1.1 I can't think of a popular language which has an unambiguous syntax and I really like playing with language grammars for some odd reason -- probably wrong though...