> > > Could you explain a bit more about how this can happen?
> > 
> > Sure.  Consider a really simple block from MachineOutput:
> ...
> 
> Thanks.  Your example makes sense.  It looks like you have to repeat the
> liveness computations for the new overflow blocks.   Another possible solution
> to the problem is to change the CPS IL so that overflow primapps are transfers
> so that the blocks are explicit at the CPS level.

But, if the liveness information for everything else is correct, then I
won't need to do a fixed point computation.  It will be sufficient to
recognize overflow blocks, trace forward through blocks until I find a
non-overflow block whose liveness is known, then propagate the liveness
back through all the blocks to the original overflow block (updating other
overflow blocks as I go along).  That should be approximately linear in
the size of the overflow blocks.