Trac #9872 showed the importance of processing goals in depth-first, so that
we do not build up a huge pool of suspended function calls, waiting for their
children to fire.  There is a detailed explanation in
     Note [The flattening work list]
in TcFlatten

The effect for Trac #9872 (slow1.hs) is dramatic.  We go from too long
to wait down to 28Gbyte allocation.  GHC 7.8.3 did 116Gbyte allocation!