I'm attaching the patch that looks promising (from local instrumentation):

Changes:
- render_block mechanism when doing "inner" run loops
- don't CGContextFlush manually (documentation recommends to leave
  window renderer to handle that)
- some small fixes in releases (toolbar)
- niling Emacs' frame struct on window dealloc - it should be garbagged
  but it seems something is still attached to the ref and cannot be GCed

Known issues:
- Can't get resizing as it's chicken-and-egg problem - drag redraws and
redraw reads on socket (which blocks redraws)

Other remarks:
- Increasing number of buffers in "double buffering" prevents leaks and
  memory allocation, but I didn't include it in the patch (e.g. not
  array of 2, but array of 4 or 6)

Best,
Przemysław Alexander Kamiński