Instead of re-generating the allocation nodes at each step, we only do
it once before calling the allocation routines, and then reuse
them. While this provides just a small speedup (~3%), it's worth doing
as it also reduces the GC cost.

Signed-off-by: Iustin Pop <iustin@google.com>
Reviewed-by: Guido Trotter <ultrotter@google.com>