Otherwise, we may replace an ia32_Sub by an ia32_Cmp
if the result proj has a different mode than mode_gp.

Also provide a cleaner, better documented and slightly faster
implementation.

- block_strictly_dominates() is trivially to compute yourself anyway.
- free_dom() and free_postdom() did not free any memory anyway and the
  irg_properties handling is performed by confirm_irg_properties()
  anyway.

There are no middleend jump nodes at this point anymore.

The ILP blockscheduler couldn't create optimal schedules anyway (because
we heuristically inserted loop breakers into the ILP)