OK, my first concept was that substrings in the regex are identified and the rows filtered in a first step before applying the regexe to them. Kind of Danny's idea.

But ... If the regex is decomposed to a query with simpler sub rexes and the Boolean logic is delegated to the SQL engine building a plan....

Well I have to admit, this could be very efficient for complex regexes

I tried to look into the PQ documentation, unfortunately it doesn't give much insight, except that substrings with at least 3 characters are needed.