Reading Prolog Programs

While you may write a Prolog program only once, and may even rewrite it a few times, you and others will typically read it many, many, many times. Consequently, it is important to know how we can best read Prolog programs: What exactly do they mean?

There are several ways to read pure Prolog programs, and we explain some of them in this text.

Declarative reading

Declaratively, Prolog programs state what holds. A Prolog program consists of clauses, and each clause is either a fact or a rule. Facts state what is always true. Rules state what is true under certain conditions.

Declaratively, a rule of the form:
Head :- Body.
is read as: "If Body is true, then Head is true." In logical terms, we say that Body implies Head, written as BodyHead, or equivalently as:
Head ← Body.
Note that :- is in fact meant to represent the arrow ←. Since Body defines the conditions under which Head holds, it can also be regarded as a constraint on the set of solutions.

This way to read Prolog programs is also called concluding reading.

A major advantage of this approach is that it is easy to explain, understand and use. You state what holds under what conditions, and the Prolog engine finds solutions for you. A disadvantage of this approach is that it does not explain why logically equivalent program variants may exhibit different performance or termination characteristics.

Example: list_list_together/3, describing the concatenation of two lists:
list_list_together([], Bs, Bs).
list_list_together([A|As], Bs, [A|Cs]) :-
        list_list_together(As, Bs, Cs).
Let us read this definition declaratively, that is in terms of relations that hold between arguments, as described by the two clauses:
  1. The concatenation of the empty list [] and any list Bs is just that list Bs.
  2. If the concatenation of As and Bs is Cs, then the concatenation of [A|As] and Bs is [A|Cs], for any element A.
Note how general this reading is: It is applicable if any arguments are instantiated, and also if they are not.

Procedural reading

To complement the declarative approach of reading Prolog programs, we can also read them procedurally. This means that we take into account the actual computation strategy of the Prolog engine.

Operationally, invocation of a Prolog predicate is similar to a procedure or function call in other languages. However, two critical differences remain: First, Prolog variables are truly logical variables and may be unbound or only partially instantiated. This cannot happen with variables in most other languages. Second, Prolog provides backtracking as a built-in feature, and will exhaustively try alternatives.

For these reasons, understanding a Prolog program procedurally is significantly harder than understanding the control flow of many other programming languages. In particular, when tracing Prolog, you need to take into account: The need to keep track of these complexities and their interactions is a major drawback of this approach. An advantage of the approach is its potential to explain different performance characteristics and termination properties of program variants.

Note also that a procedural reading almost invariably implies a particular direction of use, and therefore typically does not do justice to the full generality of logical relations.

Example: Let us read list_list_together/3 (shown above) procedurally for the query ?- list_list_together([x,y], [z], Cs).
  1. Does the first clause apply? No, because [x,y] does not unify with [].
  2. Does the second clause apply? Note that due to the way resolution works, we must introduce fresh variables when considering a clause. So, let us use A', As', Bs', and Cs' for the variables A, As, Bs and Cs that appear in the clause head. The answer is: Yes, the second clause applies with the bindings A'=x, As'=[y], Bs'=[z], Cs=[A'|Cs']. This is already rather cumbersome and error-prone, and it gets even harder to keep track of all bindings as we proceed further.
  3. Carrying on, we consider the goal list_list_together([y], [z], Cs'), repeating the same questions for this goal.
  4. Does the first clause apply? No, because [y] does not unify with [].
  5. Does the second clause apply? We need to rename the variables again. Let us use A'', As'', Bs'' and Cs''. Yes, the clause applies with A''=y, As''=[], Bs''=[z] and Cs'=[A''|Cs''].
  6. Now the whole ordeal once more, as we consider the goal list_list_together([], [z], Cs'').
  7. Does the first clause apply? Yes, at last! Note that we again need to introduce fresh variables of course. Let us use Bs''' to denote the single variable of the first clause at this step of the computation. So the first clause applies with Bs'''=[z] and Cs''=Bs'''. This means that at last a solution is found and reported as a binding for the original variable Cs which is the only variable that appears in the query. If you have carefully followed this trace (as I am sure you have, since it is so enjoyable to read), you know that Cs was unified with [A'|Cs']. Since A' was unified with x, and Cs' was unified with [A''|Cs''], and further A'' was unified with y, this makes Cs the same as [x,y|Cs'']. As we just mentioned, Cs'' was unified with Bs''', and Bs''' is [z]. Thus, the solution we found is Cs=[x,y,z], and this solution is reported by the toplevel.
  8. We still need to consider the second clause too though: No, it does not apply, because [] does not unify with [_|_].
And this was only one particular case! Accurately covering all possible modes of invocation with a procedural reading is extremely complex, and typically would take an extremely elaborate explanation. In general, you will not be able to carry this approach through, because there are too many cases to consider.

Program slicing

Program slicing is a simple and powerful technique that uses very general properties of pure Prolog to study the effects of generalizations and specializations of a program.

Examples of such properties are: In a very precise sense, program slices are explanations that answer why we observe certain phenomena.

We illustrate this with a simple example. Suppose a programmer has written list_length/2, relating a list to its length as follows:
list_length([_|Ls], N) :-
        list_length(Ls, N0),
        N #> 0,
        N #= N0 + 1.
list_length([], 0).
The predicate works exactly as intended if the list is sufficiently instantiated. For example:
?- list_length([], L).
L = 0.

?- list_length([_,_,_], L).
L = 3.
However, the predicate does not generate a single answer for the most general query:
?- list_length(Ls, L).
Program slicing helps us to see the reason. Strikeout text is used for parts that are not relevant for the behaviour we observe:
list_length([_|Ls], N) :-
        list_length(Ls, N0),
        N #> 0,
        N #= N0 + 1.
list_length([], 0) :- false.
The remaining fragment is by itself already responsible for the nontermination. No change in the strikeout parts can prevent it.

Program slices can be generated automatically and are a powerful way to locate the causes of nontermination and other unintended properties in Prolog programs. See for example Stefan Kral et al., Slicing zur Fehlersuche in Logikprogrammen, WLP 2000.

More about Prolog

Main page