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

We only use control flow scc analysis.

== and != comparisons are always lowered with the "or trick" so no need
to have the other lowering code around.

This avoids constant testing whether an output buffer was specified or
the global accumulation buffer should be used. Also remove pointless
return values.

Projs are not scheduled anymore for a long time.