We present a rewriting method for Datalog-programs
which simulates SLD-resolution
more closely than the ordinary ``magic set'' method does.
This is especially advantageous in the case of tail-recursive programs,
but already in non-recursive programs
we can often save a number of joins.
In contrast to the method of {Ross} [Ros91],
we do not only solve the problem of tail-recursion,
but try to simulate SLD-resolution as fully as possible.

Based on an idea of {Bry} [Bry90],
our method can be described in an elegant way
by means of a meta-interpreter.
An especially nice feature of our approach
is that we get many other known optimizations ``for free''
when we simulate SLD-resolution in this way.

In contrast to our earlier method [Bra95],
we can now handle the full range
from a direct simulation of SLD-resolution
to the standard ``magic set'' technique
(selectable by annotations within the program).
We compare the efficiency of SLD-resolution and magic sets
within a common framework,
explain their main differences,
and give hints for choosing between them.