According to my instrumentation of MessagePumpForUI there is a flaw in SetTimer / KillTimer logic that results in KillTimer being called up to 5 times more often than SetTimer.

Basically KillTimer would be called every time when any work was done - either regular or delayed - and there is no more delayed work. A better implementation would be to keep track of SetTimer calls and only call KillTimer if there is an active one.  

Considering about 2-8% of GPU process hangs (depending on build number) are deadlocks with KillTimer call on the top it would be good to fix this.