if we care about performance, we should stop using dynamic dispatch and make the parser generic around the dialect, with that we could make lots of these methods const (or drop the Precedence enum and just have const values on the trait) and probably improve performance significantly in general.

That would of course be a big change to the public API.

This is definitely how would implement Parser if I was starting now, but I think we should see some evidence that parsing SQL is a meaningful chunk of time for anyone before making a change like this.

My guess is that:

• even for quick queries, SQL parsing is <1% of query time
• making Parser generic and therefore dropping the restriction on Dialect that it has to be 'object safe' would actually only save us ~20%

If both those assumptions are right, this doesn't seem worth it unless it makes the code generally easier to reason with and work on.

Originally posted by @samuelcolvin in #1379 (comment)

(separate issue seems worth it for this discussion)