Did you know ... Search Documentation:
SWI-Prolog future directions

The introduction of new features and new incompatibilities with the ISO standard in SWI-Prolog version 7 has raised considerable concerns about the future of SWI-Prolog. This page addresses some of these issues. The page is organised as questions and answers.

What guides the development of SWI-Prolog?

With SWI-Prolog, we want to provide a language that satisfies the needs of academic and industrial application programmers. SWI-Prolog is a dialect of Prolog because we believe that the logic foundation provides a good basis for many relatively simple reasoning tasks seen in applications. Its reflective capabilities as well as its `program is data' view makes it an ideal platform for domain specific languages (DSLs) or micro languages, which allows for concise description of application knowledge and separation of this knowledge from how it is applied. And, of course, Prolog provides a safe environment, free of crashes and memory leaks.

SWI-Prolog does not want to be an implementation that solves the N-queens problems elegantly and in splendid isolation. Instead, we want it to be a system that can operate as a component in modern IT architectures. That is why it concentrates on multi threading, network communication (sockets, HTTP, TIPC) and exchange of commonly used structured documents (XML, JSON, RDF, CSV, etc.). That is also why it supports relevant data types: Unicode text, big integers, rational numbers and strings, as well as extended runtime detection of Prolog data types that allows for a natural representation of dynamic data while being able to distinguish a string from a list and the empty list from an identifier without additional type information in the form of declarations or wrapping data into terms.

With SWI-Prolog we want to maintain a living language. That means that we will try to make the language evolve with insights and trends in the IT world. Together with our aim to support application programming, this leads to the following priorities:

Robustness and scalability
These should be obvious. The first aim here is to ensure that properly debugged programs can run 24x7 reliably and without memory leaks. This is more or less satisfied. The second is to ensure that broken programs and development interaction (debugging, reloading, etc.) does not crash the system. There is still work to be done here.
(Backward) compatibility
We try to make as few as possible changes that break backward compatibility and stay as close as possible to the ISO standard (more about this below) and other Prolog systems (notably YAP). If incompatible changes are needed to enhance the compatibility or accommodate new extensions deemed necessary, we try to do this in such a way that (1) upgrading is relatively easy and (2) it is not hard to program in such a way that the code still runs on older versions or other Prolog implementations.

Did SWI-Prolog give up on ISO compliance?

Not (much) more than it used to. If you are looking for a Prolog system that restricts you to the ISO standard, SWI-Prolog should not be the first thing to look at. ISO compliant programs that do not explore corner cases such as relying on specific behaviour on basically invalid programs (e.g., expecting length(42,X) to fail silently) should run fine.

The ISO standard has done a great job in synchronising and cleaning the syntax and core semantics of the language. However, the standardised core is too small to accommodate real applications, the process to enlarge is is too slow (while some vendors do not even want to enlarge it) and there are no mechanisms that allow us to make even tiny incompatible changes needed to accommodate new features in a clean way. Especially this last restriction turns it into a practically dead language.

Below are a number of specific issues with the ISO standard that we experience as counterproductive.

Definition of error handling

ISO Prolog defines the behaviour of all built-in predicates both when operating on arguments within the meaningful domain for the predicate and when faced with illegal input, such as passing a non-list to the first argument of length/2. It often even defines the precise error that must be raised if the call is wrong for multiple reasons. We consider this counterproductive for the following reasons:

  • It puts an enormous stress on the ISO process itself. For example, defining that length(List, Len) is true when Len is the number of elements in the list List is easy. Defining that this predicate is non-deterministic if List is a partial list and Len is unbound is necessary. Defining what happens on length(List, -1) is also necessary, because this this can be a goal resulting from a sensible program.

    However, it makes little sense to define what happens on length(42, Len). The ISO committee decided this must fail because it wanted implementations to allow rewriting e.g., length(List,5) into List = [_,_,_,_,_] and 42 = [_,_,_,_,_] fails silently. These debates are involved, trying to balance between usefulness of an exception, performance costs and implementation effort to do the required checking, possibilities for rewriting (as with length/2), consistency, etc. There is no single truth here.

  • Defining the precise error or failure for basically invalid goals makes it hard for systems to comply because it assumes a certain implementation technique and order of execution. For example, consider meta-calling, where ISO dictates that call((fail,1)) must raise a type_error. Here, SWI-Prolog complies because call/1 compiles the argument before execution. Systems that prefer an opportunistic approach however will execute fail/0 and never try to execute the invalid 1.

    The precise error and failure conditions make it hard to perform program rewrites that now needs to maintain the exact behaviour on invalid input. It also does not allow for introducing exceptions in places where failure was prescribed because the costs of generating an exception was deemed to be too high by the ISO committee. It even disallows systems to reject goals like Len is A+B, length(Len, List) based on static analysis.

In the long run, we expect that SWI-Prolog will become a partially typed language. Type systems have to choose between decidability and expressiveness. Here, we plan to go for expressiveness by defining a type, mode and determinism annotation that can capture the richness of practical Prolog programming as we see it now. Based on that, we expect analysis tools that proof errors rather than correctness. Some of this is likely to be based on the Ciao assertion language.

Arithmetic

Being a logic based and dynamically typed language, Prolog should offer precise arithmetic results whenever possible. It should have unbound integers and rational numbers at its core. Unbound integers are not prescribed by the ISO standard and rational numbers are not even mentioned. Arithmetic is defined almost as the C language defines it, except that all overflow and evaluation errors must be mapped to exceptions (a good idea). Typed languages have no choice but defining that a specific operation returns a value of a specific type. Untyped languages however can define that X**Y evaluates to an integer if this represents the exact result and a rational number or floating point number otherwise. ISO decided for two exponentiation operators (** and ^), where ^ evaluates to to an integer and raises an exception if the result is not integral. We find this confusing. If you want ** to do floating point arithmetic, you can cast one of the arguments to escape from the world of integers:

?- Exp is 2**float(2).
Exp = 4.0

In the long run we expect a tighter integration of rational numbers. This will involve integral division to be mapped to rationals and might involve syntactical extensions to accomodate rationals.

Representing text

ISO Prolog provides no sensible way to represent a string of characters. In general, such data cannot be represented using atoms because systems pose limits on the length of atoms, the characters that can be inside atoms or the number of atoms or do not provide atom garbage collection. There are two list representations for strings, one as a list of character codes and one as a list of characters (atoms containing exactly one character). Both representations are expensive and neither can be distinguished at runtime from either a list of integers or a list of atoms or the empty list. Without runtime type information on strings, debugging becomes hard (should the debugger print "ab" as is or as [97,98]?) and dynamic data structures cannot be created. The two string representations suggest a choice, but in reality this choice needs to be made for the whole application and is therefore not a real choice.

The SWI-Prolog extensions fix some of these problems by reviving a string type as there was in the BSI Prolog standard and which survived in several implementations (e.g., ECLiPSe, Amzi!). We expect that YAP will follow. The primitives will be synchronised with ECLiPSe.

In the long run, we might do something about the _chars and _codes predicates, possibly by introducing a new data type char.

Syntax

The ISO Prolog syntax has several flaws.

Quoted atoms and strings
Here, we see several issues. Long strings with good looking source layout is not supported because there is no syntax do concatenate strings (as in C, where "hello " "world" is identical to "hello world") and there is no escape sequence that ignores a newline and leading white space. In addition, there is the idiosyncratic `\XXX\` notation for numeric character codes embedded in quoted atoms and strings, which is not compatible with older practice in Prolog, not with anything else and means nothing because it does not define how XXX must be interpreted.

Unicode is there now long enough for SWI-Prolog to support the widely accepted \uXXXX and \UXXXXXXXX. In addition, SWI-Prolog supports quasi quotations, which can support pretty looking long strings as well as safely interprolate Prolog variables into source code fragments of external languages.

Lack of support for commonly seen syntactical primitives
Expressions such as array[index], X.member, function() or function(Arg) { Body } cannot be expressed in ISO Prolog. This results in needlessly verbose and unnatural notations (see e.g., library(record)) and makes it hard to define DSLs with a natural syntax.

Should SWI-Prolog still be called Prolog?

When introducing the extensions in version 7, several people have claimed that SWI-Prolog should choose a new name that does not refer to Prolog, such as Picat or Mercury did. Both languages share concepts with Prolog, but both differ so much that it is practically impossible to run programs unmodified on both a Prolog processor and either Picat or Mercury.

This is quite different for SWI-Prolog. Most `reasonable' programs that satisfy the ISO standard or where designed for (especially YAP or SICStus Prolog) run unmodified on SWI-Prolog or can be changed easily to run on multiple systems.

What about vendor lock in?

With SWI-Prolog we wish to maintain as good as possible compatibility with ISO and other Prolog implementations in the `close family' (notably YAP and SICStus and at somewhat larger distance Ciao, ECLiPSe and XSB), while adding extensions to the system that supports our guiding principles). In practice, this means that an application programmer who experiences problems running the same source on another Prolog system while this is not a priori impossible (for example because of completely different feature sets, such as the (un)availability of attributed variables) and there is no sensible work-around will be taken seriously.

YAP and SWI-Prolog have a similar drive, where YAP concentrates on performance and SWI-Prolog concentrates on development and stability. YAP uses many of SWI-Prolog's packages and generally copies features that are required to support these packages.

Within the ISO core, it is fairly cheap to switch or maintain a portable application to just about any Prolog. If ISO doesn't satisfy your requirements and you want to be able to switch, you should carefully examine the language features you need and which systems are capable to support these.

And, of course, SWI-Prolog is open source, so you are free to fork it under the conditions of the license.

How about SWI-Prolog and education?

SWI-Prolog has always been used extensively in education. The changes introduced in version 7 do not make it significantly less suitable for this purpose. There are two issues that might require some attention.

Double quoted strings
Courses that depend on the mapping of double quoted strings to lists of character codes must either switch the double_quotes flag, run using --traditional= or use back quoted strings as illustrated in the calls below:
?- phrase("hello", "hello world", R).
false.

?- phrase(`hello`, `hello world`, R).
R = [32, 119, 111, 114, 108, 100].

?- phrase("hello", `hello world`, R).
R = [32, 119, 111, 114, 108, 100].
Lists can no longer be displayed using .(A, .(B, []))
The list functor was changed to '[|]'. This can be made visible using e.g. =../2, but no longer using display/1 or write_canonical/1 which always use the list syntax.
?- [H|T] =.. L.
L = ['[|]', H, T].

In the long run we would like to establish comprehensive tutorial material for SWI-Prolog's extensions.

Acknowledgements

I would like to thank all people who constructively helped shaping SWI-Prolog's recent extensions and expressed their concerns about the directions taken.

See also
- The SWI-Prolog extensions