When using the GTID-aware master-slave protocol, the master skips
transactions that the slave already has.  This caused the slave IO
thread to remember the wrong position.  When using GTIDs, the position
is not normally used for anything, so this should not be a big
problem.  However, as a special case, the position was used in
addition to the GTID when the slave reconnects to the same master. In
this bug, we fixed the problem by making the GTID-aware master-slave
protocol not use positions at all.
In addition, we make the master dump thread completely skip
binary logs that were generated when GTID_MODE!=ON (or by an old
server), if using the GTID protocol. This is necessary because
otherwise such binary logs will be sent again. This does not skip
anything that should have been replicated, because it is a requirement
in the upgrade procedure that the user waits until all GTID-free binary
logs have been replicated before using the GTID protocol.