Basically you partition edges into sub-pixel polygons and sort them. Then you calculate visibility and coverage before shading.
(note: this is not how REYES work for 3D, but I think it could be adapted with good results for non-realtime 2D this way.)
