When an update query is to be written to the binlog, and it reads unset
user variables (example: INSERT INTO t VALUEs(@A) where @A does not exist),
we create the variable like if it had been set with SET before, and we loop
(i.e. we redo the steps of logging, now that the variable exists).