It seems that (if I'm taking @TJones' comment at T410758#11410188 correctly) that regex searches are first preprocessed to find sections that can be matched via the index, and in particular that if a regex search has no non-escaped metacharacters (and is thus looking for some exact string) then it will be almost as fast as an index search (one would want to do this sort of "regex" to search for an exact string without applying grayspace filters, for example).
This contradicts the documentation at https://www.mediawiki.org/wiki/Help:CirrusSearch#Regular_expression_searches which implies such things need an indexed precomponent to be efficient. Could the documentation for regex searches please be updated to show how the software actually behaves?