there are a few major changes that we should make to GMainLoop:

 - race-free with respect to signals
 - microsecond accurate
 - not O(n)
 - possibility of edge-triggered event sources


Set the edge-triggered event sources thing aside for a moment since it's pretty extremely complicated, and consider the others.  The biggest thing standing in the way is the existing GSource API.  Mostly for two reasons:

 - each source expects to be polled each time (which is O(n))
 - the source prepare function returns a time in milliseconds

It seems that we could easily fix quite a lot of problems with two relatively minor API tweaks:

  1) Add a g_source_set_wakeup_time() to allow setting an absolute monotonic
     time (in microseconds) at which the source should be woken up.

  2) Allow the possibility of having the check/prepare functions given as
     NULL to mean "don't poll this source, but only dispatch it when one
     of its pollfds is ready or when its wakeup time has been reached".


Internally, we would separate sources into two lists: those with the check/prepare functions and those without.  On a mainloop iteration, we'd treat the sources with check/prepare in the same way as we do now, and completely fail to visit the ones without.  Then we're only O(n) in terms of sources that use the old API.