I thought about this multiple times already. The tags table consists only of 3 fields now (id, data table id, tag name). introducing an mm results in a new table with 'foreign' and 'local' id, and the tags table then has 'id' and 'tag name' only. So as a result, this might have benefits if single tags are applied to very many data entries, this would reduce the number of rows in tags table (the mm table would have the rows) and the tags name wouldn't be stored multiple times.
since the tags names are usually not very long, and indexes on the tags table are properly adjusted to the queries run on this table, i doubt an additional normalization gives us much benefit (we actually need more queries to handle them).
since the patch in gerrit is waiting for a long time already without anything happening, it also seems that nobody else sees a real benefit in adjusting this construct. if the issue is ever brought up again, it should be feeded with a definition of different use cases (many data rows with less tags, many data rows with lots of mm relations, ...) and carefully benched.
For now, i'll abandon the patch on gerrit and will close this issue here.
Updated by 
Updated by 
Updated by 
Updated by