UBC Theses and Dissertations

UBC Theses Logo

UBC Theses and Dissertations

xpProlog : high performance extended pure prolog Lüdemann, Peter Gerald 1988

Your browser doesn't seem to have a PDF viewer, please download the PDF to view this item.

Item Metadata


831-UBC_1988_A6_7 L82_2.pdf [ 11.46MB ]
JSON: 831-1.0051961.json
JSON-LD: 831-1.0051961-ld.json
RDF/XML (Pretty): 831-1.0051961-rdf.xml
RDF/JSON: 831-1.0051961-rdf.json
Turtle: 831-1.0051961-turtle.txt
N-Triples: 831-1.0051961-rdf-ntriples.txt
Original Record: 831-1.0051961-source.json
Full Text

Full Text

xpProlog: High Performance Extended Pure Prolog by PETER G E R A L D L U D E M A N N B. Sc., The University of British Columbia, 1975  A THESIS S U B M I T T E D IN P A R T I A L F U L F I L L M E N T O F T H E R E Q U I R E M E N T S FOR T H E D E G R E E O F M A S T E R OF S C I E N C E in T H E F A C U L T Y O F G R A D U A T E STUDIES D E P A R T M E N T OF C O M P U T E R SCIENCE  We accept this thesis as conforming to the required standard  T H E U N I V E R S I T Y OF BRITISH C O L U M B I A February 1988 © P E T E R L U D E M A N N , 1988  In  presenting  degree freely  at  this  the  available  copying  of  department publication  of  in  partial  fulfilment  University  of  British  Columbia,  for  this or  thesis  reference  thesis by  this  for  his thesis  and  scholarly  or for  her  of  The University of British Columbia Vancouver, Canada  DE-6  (2/88)  I  I further  purposes  gain  the  shall  requirements  agree  that  agree  may  representatives.  financial  permission.  Department  study.  of  be  It not  that  the  advanced  Library shall  by  understood be  an  permission for  granted  is  for  allowed  the that  without  make  it  extensive  head  of  copying my  my or  written  Abstract Adhering to the principles of logic programming results in greater expressiveness than is obtained by using the many non-logical features which have been grafted onto current logic programming languages such as Prolog. This report describes an alternative approach to high performance logic programming in which the language and its implementation were designed together. Prolog's non-logical features are discarded and new logical ones are added. (xpProlog)  Extended pure  Prolog  is a superset of conventional Prolog; it is sufficient in itself, without  any need for "impure" non-logical predicates. This gives both greater expressiveness and better performance than conventional Prologs.  XpProlog  programs have the following advantages over conventional Prolog  programs: •  They are often easier to understand because their meaning does not rely on the underlying computational mechanism. ii  •  Coroutining, automatic delaying and sound negation are available.  •  As technology improves, better implementations and optimization techniques can be used without affecting existing programs.  This report covers:  •  The proper use of logic programming.  •  How  Prolog must be changed to become a good logic programming language  (xpProlog).  •  Sound negation and coroutining.  •  An efficient abstract machine (xpPAM) which can be efficiently emulated on conventional machines, translated to conventional machine code, or implemented in special purpose hardware.  •  How  to compile extended Prolog and functional (applicative) languages to  the abstract machine or to conventional machine code. •  Discussion of alternative Prolog abstract machine designs.  The x p P r o l o g Abstract Machine's design allows: •  Performance similar to the Warren Abstract Machine (WAM)  for sequential  programs. •  Tail recursion optimization (TRO).  •  Parallelism and coroutining with full backtracking.  •  Dynamic optimization of clause order.  •  Efficient  •  Simple, regular instruction set for easily optimized compilation.  •  Efficient memory utilization.  if-then-else  ("shallow" backtracking).  iii  •  Integrated object-oriented virtual memory.  •  Predicates as first-class objects.  •  Simple extension to functional programming.  C.R. categories: 1.2.5: Prolog; D.1.3: concurrent programming; D.3.2: very high level languages; D.3.3: language constructs: coroutines, backtracking; D.3.4: interpreters.; 1.2.3: logic programming.  iv  1  Contents  Pure logic programming  1  1.1 What is wrong with impure logic programming  1  1.2 On good design of a logic programming language  3  1.3 Principles of logic programming and pure Prolog  4  1.4 Background  7  1.5 Delay notation  8  1.6 Organization of the Report  12  The abstract machine  13  2.0 Fast append on a conventional machine  15  2.1 Data structures  17  2.2 C code for append function  19  2.3 C code for deterministic append predicate  21  2.4 Machine code for deterministic append predicate  25  2.5 Abstract machine code for deterministic append predicate  26  3.0 The basic sequential inference engine  28  3.1 Objects  29  3.2 General registers  32 V  3.3 Allocating and freeing object cells  33  3.4 Status registers  33  3.5 Execution stack  34  3.6 Code segments  36  3.7 Machine instruction format  36  3.8 Net effect of calls  39  3.9 Tail recursion optimization (TRO)  39  3.10 Unification  40  3.11 Instructions  41  3.11.1 Unification and testing instructions  44  3.11.2 Control instructions  47  3.11.3 Call / return instructions  50  3.11.4 Allocation / deallocation instructions  54  3.11.5 Miscellaneous instructions  54  3.12 Code example  55  3.13 User defined unification  58  3.14 Other object data types  59  3.15 Virtual memory  61  3.16 Object paged virtual memory  62  4.0 Backtracking and delaying  66  4.1 Stacks for backtracking  66  4.2 Backtracking instructions  70  4.3 Cost of backtracking  74  4.4 Delays  75  4.5 Cost of delaying  80  vi  4.6 Weak delays: dynamic reordering of clauses  81  4.7 Delay instructions  83  5.0 Compiling Prolog  85  5.1 Unoptimized compiling of a predicate  88  5.2 Basic compiling of a clause  90  5.3 Optimized compiling of a clause  95  5.4 Spilling registers within a clause  97  5.5 Optimized compiling of a predicate  98  5.6 Shallow backtracking  105  5.7 Global optimizations  106  5.8 Compiling a predicate with delays  108  5.9 Optimized compiling to conventional machine code  110  5.10 Speeding up deterministic predicates - modes and types  112  6.0 Compiling a functional language to the logic engine  116  6.1 Introduction  116  6.2 Functional Programming background  117  6.2.1 Normal and applicative order evaluation  119  6.2.2 Lazy and eager evaluation  120  6.2.3 Lexical and dynamic scoping (deep and shallow binding)  120  6.2.4 Mechanical evaluation of functional programming constructs 6.2.5 Functional programming via logic programming 6.3 Sample code  ... 121 122 123  6.4 Thunks, lazy evaluation and higher order functions 6.4.1 Lazy thunks vs. delayed predicates  vii  127 132  6.5 Equality: "is" and " = "  133  6.6 Combinators  134  7.0 Design decisions  138  7.1 Comparison with the Warren Abstract Machine instructions  138  7.2 Environments on the execution stack  141  7.3 Global and local stacks  142  7.4 Saving environments (WAM)  144  7.5 Allocating from a list or from a stack  145  7.6 Cut  146  7.7 Code indexing  147  7.8 RISCs, CISCs and in between  148  7.9 Instruction format  150  7.10 Functor and list storage  151  7.11 Object cell size  152  7.12 Reference counts and garbage collection  153  Extended pure Prolog  156  8.0 Execution order  158  8.1 Language issues  158  8.2 Conventional Prolog's execution order  159  8.3 Example of conventional execution order  163  8.4 A more flexible execution strategy  165  8.5 Negation  168  8.6 Single solutions  170  viii  8.7 Cut  173  8.8 Clause order  176  8.9 All solutions predicates  178  8.10 Non-strict execution order: delays  181  8.11 Input/Output  183  8.12 Efficiency  184  9.0 Coroutining, pseudo-parallelism and parallelism  185  9.1 Notation  186  9.2 Coroutining example  187  9.3 Test and generate  191  9.4 Pseudo-parallelism  194  9.5 Correctness and completeness  196  9.6 Parallelism  198  9.7 Comparison with other designs for delaying  200  10.0 Extensions to Horn logic  206  10.1 Negation  207  10.2 Closed predicates  210  10.3 Setof, bagof  211  10.4 Meta-variables  ~.  212  10.5 Constraints vs. delays  215  10.6 The occurs check  217  10.7 Assert and retract  219  11.0 Arrays and I/O done logically and efficiently  ix  222  11.1 Introduction  223  11.2 Array Operations  225  11.3 Transformation to allow destructive assignment  230  11.4 Efficient implementation of array operations  232  11.5 An example: Quick-sort  236  11.6 Direct access I/O  239  11.7 Databases  240  11.8 Sequential Input/Output  241  11.9 Implementation of many object types  244  11.10 Expressiveness  247  Conclusion  252  Glossary and Index  255  References  272  Appendix A. Sample machine code  280  A.l Machine code for deterministic append predicate  280  A. 2 Machine code for append function  282  Appendix B. Details of xpProlog syntax  285  B. l Functors and lists  285  B.2 Lexical and syntactic details  286  B.3 Critique of Prolog's syntax  291  X  B.4 Debugging extensions  293  Appendix C. Implementation status  295  Appendix D. Built-in predicates  300  D.l Built-in opcodes  311  xi  Preface He had explained this to Pooh and Christopher  Robin once  before, and had been waiting ever since, for a chance to do it. again, because it is a thing which you can easily explain before anybody knows what you are talking — A. A.  Je n'aifait  Milne,  twice about.  Winnie the Pooh  celle-ci. plus tongue que parce que je n'ai pas eu Ie loisir de la (aire plus courte.  (I have made this letter longer than usual, only because I have. not. had the time to make it  — Blaise Pascal,  Lettres Provinciates  shorter.) (1656-1657)  This report is the result of several year's investigation into logic programming and its most popular realization: Prolog. I started by wanting to design a typesetting system, implemented in Prolog but the Prologs that were available xii  then were too slow and lacked some necessary features such as coroutining and sound negation. I therefore decided to investigate logic programming and implement an interpreter.  Implementing a compiler and interpreter is not a task to be undertaken lightly. I have worked part time on this for three years and I have not produced a production quality implementation. However, I have succeeded in demonstrating the possibilities of logic programming. Three or four people should be able to produce a production quality implementation of this in about a year.  ... when you are a Bear of Very Little Brain, and you think of things, you find that sometimes that a Thing which seemed very Thingish  inside you is quite  different when it gets into the open and has other people looking at it.  — A. A. Milne,  Winnie the  Pooh  This report contains two themes: pure logic programming and an abstract machine for implementing pure logic programming. These themes affect each other because I want to show that pure logic programming can have an efficient implementation. I wish to put forward my abstract machine design as a viable alternative to the popular Warren Abstract Machine (WAM) [Warren 1983], so I have described the implementation first and the language second.  This work describes practical issues which have been sufficiently implemented to give an existence proof of their efficacy. Some ideas in this report have come xiii  from analysis of the implementation and have not yet found their way into the implementation. Most of the abstract machine ideas have been implemented but some of the language ideas have not. There are no benchmark data in this report. The usual benchmark (naive reverse) is probably not a very good predictor of over-all execution speed. My implementation has achieved competitive speed for this simple benchmark, compared against published figures for the fastest commercial Prolog implementations. However, proper speed evaluation would take many months of hard work and is beyond the scope of this report.  1  Acknowledgments To spend too much time in studies is sloth.  — Francis Bacon,  Of  Studies  This research has been partially funded by a Shared University Grant from IBM Canada Ltd. I thank Harvey Abramson and Douglas Teeple for suggesting that I become involved in logic programming (Harvey also let me use his extensive collection of papers and provided the introductory material for the chapter on compiling a functional language); Block Bros. Data Centre which let allowed me the  The interested reader might obtain the E C R C benchmarks, published on Usenet. Other benchmarks exist, for example, Evan Tick's translation of the LISP benchmarks described in [Gabriel 1985].  xiv  necessary flexibility in my work schedule; Lee Naish, Tom Rushworth, Paul Voda and Marc Gillet for listening, criticizing and suggesting; Romas Aleliunas for extensive editing suggestions; and my wife, Kikuyo, for letting me leave many household chores undone while I worked on this project.  XV  Pure logic programming  Art, it seems to me, should simplify.  That, indeed, is very  nearly the whole of the higher artistic process; finding  what  conventions of form and what detail one can do without and yet preserve the spirit of the whole — so that all that one has suppressed and cut away is there to the reader's  consciousness  as much as if it were in type on the page.  — Willa Gather, On  1.1  the Art of Fiction  What is wrong with impure logic programming  A typical attitude to non-logical features in logic programming is [Cohen and Feigenbaum 1982, p. 123]: To a certain extent, the development of logic programming has followed the pattern of LISP. Both languages are founded on clear, mathematically 1  motivated foundations. Both languages have a side-effect-free kernel and a procedural interpretation that can be defined in a simple and elegant fashion. Yet both language families have yielded to the practical needs of their user communities and have incorporated numerous features that detract from their underlying elegance in favor of improved convenience and efficiency. In a sense, the fact that logic programming has progressed to the point of incorporating such features attests to its practicality and growing popularity.  This author admits that non-logical features are wrong, yet he defends them because of their "improved convenience and efficiency." This brings to mind the "go to considered harmful" [Dijkstra 1968] controversy. Many letters were written by people, claiming that go tos were needed because of their convenience and efficiency. We now know that programs written in a disciplined style, without go tos, are usually clearer and just as efficient (if not more so) than go fo-filled spaghetti code. Similar results are now being published for cut [Debray 1986] [O'Keefe 1985].  This report describes the implementation of a pure logic programming language. The language's features remove any excuses for writing bad, non-logical programs. The vision expressed in "Algorithm = Logic + Control" [Kowalski 1979b] is possible. And programs written with such a pure logic programming language will run as fast as programs written in conventional Prolog sometimes faster.  2  1.2  On good design of a logic programming language  Simplicity requires much thought; complexity merely requires much work. But many computer language and machine designers refuse to see the advantages of simplicity — they correct perceived defects by adding new "features," thereby burying a simple design under a mass of graceless and cumbersome accretions. Eventually, the whole ungainly mess must be replaced by something new.  A programmer's major tool is her programming language. She uses it every day; its quality continually affects her. No programming language can stop a bad programmer from writing bad programs; but a bad programming language can prevent a good programmer from producing good programs. Logic programming is a simple idea which is already in danger of being buried under a myriad of new "illogical" features. Prolog is already encumbered this way — the result is less powerful than a simple design based strictly on the original concepts. Programming language designers must strive to give the highest quality product possible, with every part carefully considered as to how it will help good programmers to produce good programs. The language designer must never succumb to doing "what is 'reasonable' even when it isn't any good" [Pirsig 1974] - his motto should be: "Only the good; never the expedient." A simple design is not necessarily easy to implement. The small number of basic principles mustfittogether well without duplication. Any careless side effects will prevent some pieces Fitting together. The designer needs much 3  discipline to not deviate from the basic principles by introducing special cases and restrictions. This discipline results in a product which is easy to use and easy to extend.  1.3  Principles of logic programming and pure Prolog Louis XVI: Le Rochefoucauld-Liancourt:  C'est une grande  Non, Sire, c'est une. grande  revoke.  revolution.  The central ideas of logic programming are:  •  The most interesting thing about a program is what it does, not the exact sequence of computations.  •  The best way of describing a program is by writing specifications using a subset of mathematical logic.  •  When the logical specifications are stated using Horn clauses, they can be executed efficiently.  •  There is no concept of time or sequence in the meaning of a logic program (execution may, of course, be sequential).  The programmer does not need to write a program, only the specifications. This is "declarative programming." Logic programming offers a solution to one of the problems of software engineering: precise problem specification. Control information may be added to a logic program, as suggested by "Algorithm Logic + Control"  =  [Kowalski 1979b]. This additional control information does  not affect a program's meaning; it just improves execution speed.  4  Logic programming is new to many computer scientists and programmers. A new way of thinking is required to write good logic programs. Without a new way of thinking, programmers produce programs which are full of non-logical features such as cut ("!") or var — these are really disguised conventional programs which have been hacked to look like logic programs. These non-logical features are not necessary. See [Walker, McCord, Sowa and Wilson 1987, Chapter 3] for methods of avoiding or localizing non-logical features.  Early implementers of Prolog "solved"its space-and time-inefficiencies, by adding poorly conceived "non-logical" features. For a discussion of non-logical constructs and their inconsistent implementations, see [Moss 1986]. Instead of patching existing implementations, extended  pure, Prolog (xpProlog)  generalizes  the original concept. Syntactically, it is similar to Prolog, but:  •  Extended:  with control rules which add flexibility to Prolog's strict  left-to-right depth-first computation rule, allowing natural specification of a larger set of programs than ordinary Prolog. Control can be specified separately from predicate definitions. •  Pure:  without (and not needing) non-logical predicates such as  cut ("!"), var,  etc..  Although xpProlog's syntax is similar to Prolog's syntax, some x p P r o l o g programs will not work on a conventional Prolog implementation - they will either go into infinite loops or run very slowly. X p P r o l o g programs can be given an obvious declarative reading in first-order logic — this is not true of many conventional Prolog programs. X p P r o l o g programs' behaviour do not depend on execution order, although efficiency will depend on execution order. XpProlog's 5  greater expressive power costs almost nothing in efficiency; sometimes it greatly increases efficiency.  Prolog is based on the Horn clause subset of first order logic. Negation, first and second order logic and set theory can be added [Lloyd and Topor 1984] [Voda 1986]. These are implemented unsoundly in conventional Prolog, using non-logical predicates such as cut ("!") and var. Sound implementations are no more difficult than unsound implementations, as will be shown.  Warren's pioneering design for implementing logic programming (the "Warren Abstract Machine" or WAM)  [Warren 1977] [Warren 1983] [Gabriel,  Linkholm, Lusk and Overbeek 1985] proves that logic programming languages can be executed as efficiently as other symbol-oriented languages (see [Tick 1986] for a comparison with LISP). WAM  allows efficient execution on both  conventional and special purpose hardware (for example [Dobry, Patt and Despain 1984] [Tick and Warren 1984] [Dobry 1987]). The x p P r o l o g Abstract Machine (xpPAM) is a significant modification of WAM,  to support the flexible  execution order needed for coroutining and sound negation. Prolog code can be compiled to xpPAM more easily than to WAM;  execution performance is similar  for both machines.  This paper, then, explores some extensions to Prolog —  Extended pure Prolog  —  which remove restrictions from conventional designs. X p P r o l o g often allows more compact, understandable and efficient programs than conventional Prolog allows.  6  I will start by discussing the implementation xpPAM.  xpProlog  and its abstract machine  The abstract machine can support conventional Prolog,  xpProlog  and  functional languages - I will focus mainly on its support of x p P r o l o g . But the reader should be careful do distinguish between what is imposed by the abstract machine, what is imposed by the compiler and what is part of the language design.  Efficient implementation of x p P r o l o g , like efficient implementation of conventional Prolog, requires a good compiler. The techniques for such a compiler will also be discussed.  1.4  Background If I have seen further it is by standing on the shoulders of giants.  — Sir Isaac Newton,  Letter to Robert  Hooke  I assume that the reader has some basic knowledge of mathematical logic, logic programming and conventional Prolog. [Kowalski 1979] is probably the best place to start because he concentrates on logic and avoids being constrained to one particular implementation. Texts on conventional Prolog include [Walker, McCord, Sowa and Wilson 1987], [Sterling and Shapiro 1986], [Bratko 1986], [Kluzniak and Szpowiez 1985], [Clark and McCabe 1984] and [Clocksin and Mellish 1981]. Mathematical logic is covered in [Quine 1941], [Kleene 1967] and [Hodges 1977].  A glossary is provided at the end of this report.  7  I do not assume any knowledge of the Warren Abstract Machine (WAM). The papers which describe the WAM  often gloss over some of the more subtle design  details which contribute crucially to its speed. Many of these details also exist in my design and I will try to explain them fully.  My ideas about the control rules for unrestricted pure Prolog derive from [Naish 1985b]. The control rules are sound [Lloyd 1984, pp. 45-47]. The implementation techniques derive from the Warren Abstract Machine [Warren 1983]. I have made some small changes and extensions to Naish's ideas and significant changes to Warren's. I recommend reading [Naish 1985b] although I will present many of his ideas, but in a slightly different form.  1.5  Delay notation  X p P r o l o g is pure Prolog, extended for delays. In x p P r o l o g , any variable may be  suffixed by a question mark ("?"). This will cause the predicate to delay until the variable becomes instantiated. This instantiation is required only at the "top level"; for example, if the variable became instantiated to a list element ("cons cell"), the head and tail would not necessarily need to be instantiated for execution to resume. The "?" is propagated outward from a compound term. "X?.Y" is the same as " ( X ? . Y ) ? " - the parameter must be instantiated to a list element, the head  must also be instantiated, but the tail of the list element need not be instantiated. XpProlog's "?" differs from the "?" in Concurrent Prolog [Shapiro 1983]. For example: p r e d ( X ? . R e s t , Y) :- t e s t l ( X ) , t e s t 2 ( Y ? ) , t e s t 3 ( X , Y ) . 8  will delay until the first parameter is instantiated to a list element with its head (X)  also instantiated. Once these have become instantiated, t e s t l will be tried.  If t e s t l succeeds, p r e d will delay until Y becomes instantiated, after which test2  and t e s t 3 are tried.  A delayed predicate acts as if it had succeeded. When the required variables become instantiated, the delayed predicate's execution is resumed. Using the above example of p r e d with the query ?- pred(A, B ) , A = (H.T), H = a, Y = b.  execution will proceed: delays on A  pred(A, B)  pred  A = (H.T)  instantiates A  pred(H.T, B)  pred  H = a  instantiates H  pred(a.T, B)  p r e d resumes  testl(a)  (succeeds)  test2(B)  pred  Y =b  instantiates Y  test2(b)  pred  t e s t 3 ( a , b)  (succeeds)  resumes and delays on H  delays on B; t e s t 2 is not tried resumes; tries t e s t (succeeds)  Instead of marking predicates with "?"s, separate proceed declarations may be used. Thus,  2  2  Here, the "x," " r e s t " and " r e s u l t " are comments. predicate which tries to match question marks.  9  " P r o c e e d " acts like an ordinary  It is an error to call p r o c e e d with any  ?- p r o c e e d p r e d x ( x ? . r e s t , r e s u l t ) . /* "x", " r e s t " ,  " r e s u l t " are comments */  predx([], [ ] ) .  p r e d x ( X . R e s t , R e s u l t ) :- ...  is the same as predx([]? [ ] ) . predx(X? .Rest, R e s u l t ) :- ... 5  The main advantage of  proceed  declarations is that they are separate from the  clauses. They can also be used when a predicate must delay until any one of several variables become instantiated. Multiple proceed declarations are or-ed. An example of this is append when used to implement  append3  for combining  (or splitting) three lists: ?- p r o c e e d ?- p r o c e e d append([], append(X.A, append3(A,  Append  append(a?, b, c ) . append(a, b, c ? ) . X, X ) . B, X.C) :- append(A, B, C ) . B, C, D) :- append(A, B, Z ) , append(Z, C, D).  will proceed (not delay) only if the first argument or if the third  argument is instantiated; otherwise it will delay.  Proceed  declarations are not  needed for append3 because all the delaying is done within  append.  Without the proceed declarations, this would go into an infinite loop for the query  uninstantiated variables because these will incorrectly unify with question marks. The declaration could be given ?-proceed p r e d x ( ? . - , - ) .  10  ?- append3(l.W, X, Y, 2.Z).  and there is no way to re-order the clauses of append to prevent the infinite loop.  The effect of the proceed declarations for  append  is given b y the following  pseudo-code (which is not supported b y xpProlog): a p p e n d ( P l , P2, P3) :i f v a r P I then i f v a r P3 t h e n o r D e l a y ( P l , P3) e l s e a ( P l , P2, P3) e l s e a ( P l , P2, P 3 ) . a ( [ ] , X, X ) . a(X.A, B, X.C) :- append(A, B, C ) . /* Note: c a l l s "append", n o t " a " */  The pseudo-control structure  ifvar  tests the variable for being instantiated. The  pseudo-control predicate delay suspends the predicate until any one of its arguments becomes instantiated - execution then resumes at the beginning of the predicate. The ifvar pseudo-control structure is similar to conventional Prolog's var predicate. It can be used correctly only if it is used only with other var  predicates or with  and safer than  ifvar  delay.  XpProlog's proceed declarations are more readable  and delay, so only proceed declarations are provided in the  language.  Examples of using delays are given in sections 9.2, "Coroutining example" on page 187, section 9.3, "Test and generate" on page 191 and section 9.4, "Pseudo-parallelism" on page 194.  11  1.6  Organization of the Report  The rest of this report consists of two sections:  •  the abstract machine (xpPAM) and its implementation  •  the extended Prolog language x p P r o l o g and examples of its use.  The two sections are almost independent of each other. The common theme is delaying which allows predicates to be tried in a different order from conventional Prolog's strict left-to-right top-down order. The "?" operator and proceed  declarations (which mark delays) are explained in section 1.5, "Delay  notation" on page 8.  12  The abstract machine The extended Prolog abstract machine (xpPAM) can be used to implement conventional Prolog, extended pure Prolog (xpProlog) or pure functional programming languages. This report will concentrate on x p P r o l o g implementation. The extended Prolog language (xpProlog) is similar to conventional Prolog, except for: •  No "impure" non-logical predicates.  •  Delay notation (described in section 1.5, "Delay notation" on page 8).  •  More control structures such as if-then-else.  •  Meta-variables for all-solutions predicates.  13  XpPAM  has superficial similarities with the Warren Abstract Machine (WAM)  [Warren 1983]. Programs can run on xpPAM about as quickly as they can run on WAM.  XpPAM  However, xpPAM is much simpler than  WAM.  can be thought of either as the design for a logic programming engine  (implemented as an interpreter or in hardware) or as an intermediate code for producing machine code on conventional hardware.  The description is divided into sections:  •  Section 2.0, "Fast append on a conventional machine" on page 15 uses the append  predicate as an example of how Prolog can be compiled to very  efficient machine code on a conventional machine using C and IBM/370 assembler. •  Section 3.0, "The basic sequential inference engine" on page 28 describes the basic machine and lists all its instructions, skipping over the non-deterministic features.  •  Section 4.0, "Backtracking and delaying" on page 66 describes how the basic sequential engine is extended to allow backtracking and delaying.  •  Section 5.0, "Compiling Prolog" on page 85 describes a compiler for producing good abstract machine code.  •  Section 6.0, "Compiling a functional language to the logic engine" on page 116 describes how functional languages can be compiled to efficient xpPAM code.  •  Section 7.0, "Design decisions" on page 138 discusses some of the design trade-offs in xpPAM.  14  2.0 Fast append on a conventional machine You must lie upon the daisies and discourse in novel phrases of your  complicated  state of mind, The meaning doesn't matter if it's only idle chatter of a transcendental And  kind.  everyone will say,  As you walk your mystic way, "If this young man expresses himself in terms too deep for  me,  Why, what a very singularly deep young man this deep young man must be!"  — Sir William S. Gilbert, Patience,  I do loathe  — J. M. Barrie, My  explanations. Lady  To illustrate what a Prolog implementation should do on a conventional machine, I will translate deterministic append into C. The xpPAM abstract machine and backtracking will be introduced in a later chapter. a p p e n d ( [ ] , X, X ) . append(X.A, B, X.C) :- append(A, B, C ) .  15  act I  Nicotine  A good Prolog optimizing compiler should detect common patterns of code and translate them into special sequences. The  append  predicate is typical of a more  general sequence: P(U> [ ] ) • /* t e r m i n a t e at n i l */ p(A.X, A2.X2) :- q(A, A 2 ) , /* p e r f o r m some o p e r a t i o n on A, g i v i n g A2 */ p(X, X2). /* c o n t i n u e w i t h t h e r e s t o f t h e l i s t */  This is often deterministic and can therefore be handled well by conventional machines. On an IBM/370, the  append  inner loop can be reduced to 8 machine  instructions, compared to 2 instructions for xpPAM abstract machine (7 for WAM ). For mostly deterministic predicates, a conventional machine can 3  therefore be as fast as special purpose hardware.  The xpPAM code is given later in this chapter. The W A M  code is:  append/3: s w i t c h o n t e r m ... []=>L1, l i s t = > L 2 , var=>L0 LI: get_nil 1 g e t _ v a l u e 2,3 proceed L2: getlist 1 unify_var 4 unify_var 1 getlist 3 unify_val 4 unify_var 3 execute append/3 LO: try LI trust L2 This has 8 instructions in the inner loop, It can be reduced to 7 by replacing the last execute by the first switchonterm. 16  I will describe the Prolog equivalent of the LISP function ( d e f u n append(A B) (cond ( ( n i l A) B) (T (cons ( c a r A) (append ( c d r A) B ) ) ) ) )  However, there is an important difference: the Prolog version is tail-recursive but the LISP version is not (the recursive call to LISP's append is inside a call to cons).  To produce the iterative code given below, the LISP code requires a more  complex transformation than does the Prolog code.  2.1  Data structures  First, the data types to define a value cell (slightly simplified from the actual implementation):  4  Many implementations save space by using variable size value cells and by coding information directly when possible. Variable size cells do not change the accessing code but they do complicate the heap manager. pointers.  However, more savings can be had by avoiding  For example, the list [1,2] could be fit into one list element if a list element  contained two value cells instead of two pointers: the values 1 and 2 could be of type "short integer" which fits inside a word. Such a scheme complicates accessing elements; I have chosen implementation simplicity and speed over memory usage (although this scheme could be worked into my design). Other list compaction schemes such as "cdr-coding" could also be used but they also complicate the implementation, particularly unification.  17  0  nil  //////(///////  1  ptr  2  value  nun  integer  3  ptr  nun  indirect  4  age  mm  uninstantiated  ptr  list  elem  reference  variable  In C, this would be expressed: /* „  value  enura t g E  cell =  tags:  */  {tgNil=0,  tgLlstElem=l,  tgVariable=4, tgUnalloc=5 union  cellUnion  int  };  tglnteger=2,  tgReference=3,  /*  uninstantiated  /*  unallocated:  on f r e e  variable  "age"  variable  */  list  */  { asInteger;  int  varAge;  struct  cell  struct  {  /*  of  the  */  *asReference;  struct  cell  *head,  *tail;  }  asListElem;  }; struct  cell  enum t g E union  [ tag;  /*  cellUnion  tgNil,  tgListElem,  tgVariable  */  u;  };  The contents of a value cell depends on the t a g . A list is made up of list elements (tag t g L i s t E l e m ) each of which contains a head and a tail. The last element is nil (tag  tgNil).  A "reference" cell contains a pointer to another value 18  cell - it is invisible to the user because it is always dereferenced whenever it is used. An uninstantiated variable (tag t g V a r i a b l e ) is a cell which has not yet been assigned a value. When the value is determined, the tag is changed and the new value filled in. The "age" of a variable is used to determine whether or not information must be saved when the variable becomes instantiated, to allow backtracking (this corresponds to the Warren Abstract Machine method of determining the variable's age by a comparing stack positions).  2.2  C code for append function  Here is the most efficient version of append possible with the above value cell definitions. It assumes that there are no "reference" cells and that the third argument is always the output (that is, append is not used to split a list) — these extra complications will be dealt with later.  The  append  function takes two value cell pointers  (pi, p2)  and a pointer to a  value cell pointer (p3) to return the new list (in a language like Pascal, p3 would be a var parameter; C requires declaring it as a pointer). The algorithm is: while pi points at a list element cell { allocate a new list element cell; assign it to what p3 points at make pi point to the next input list cell (or nil if at end of list) make p3 point to the tail of the new list cell } if pi points at nil, assign p2 to what p3 points at otherwise raise error condition  19  Note that for speed, the tail of each new list element is notfilledin; however, p3 points at it and the next iteration fills it in. Here is the C code. The aliocListElem(Head,Tail) dummyCell  procedure is used to allocate new list cells. The 5  is strictly a place holder for allocListElem;  its contents do not matter  at all. v o i d a p p e n d ( p l , p2, p3) s t r u c t c e l l * p l , *p2, **p3; /* p3: v a r c e l l P t r */ { while (pl->tag = tgListElem) [ *p3 = a l l o c L i s t E l e m ( p l - > u . a s L i s t E l e m . h e a d , &dummyCell); pi = pl->u.asListElem.tail; p3 = & ( ( * p 3 ) - > u . a s L i s t E l e m . t a i l ) ; }  i f (pl->tag =  t g N i l ) { * 3 = p2; } else { error(); } P  }  The query ?- a p p e n d ( [ l , 2 ] , [ 3 , 4 ] , L 3 ) .  is coded: s t r u c t c e l l * p l , *p2, *p3; pi = allocListElem(allocInteger(1), allocListElem(allocInteger(2), SmilCell))); p2 = a l l o c L i s t E l e m ( a l l o c I n t e g e r ( 3 ) , allocListElem(allodnteger(4), &nilCell)); /* p3 does n o t need t o be i n i t i a l i z e d */ a p p e n d ( p l , p2, &p3);  and results in (note that p3 is shown between pi and p2):  In the actual program, this statement is written slightly differently, using a macro.  20  — >  1—>  pi  V 2  p3  1  ///  >  2  2  ///  -> 1  /// / / / p2  2  3  ///  2  4  ///  Optimization to this level is only possible if the compiler knows that there are no "reference" cells. If there are reference cells, they must be dereferenced each time they are used by: while (pl->tag =  tgReference) { p i = pl->u.asReference; }  This will never go into an infinite loop because the unification algorithm guarantees that a reference cell will never point to itself.  2.3  C code for deterministic append predicate  The above C code in effect implements a function which returns a value in p3. If p3 is allowed to be any kind of value, the following code will do the job. For brevity, only code for pi as (a reference to) nil or a list cell and for p3 as a 21  reference cell or uninstantiated variable is shown; the other cases generate calls to error. The full implementation has code for handling backtracking when  pi  is an uninstantiated variable - this is discussed in section 4.0, "Backtracking and delaying" on page 66. v o i d a p p e n d ( p l , p2, p3) /* append u s i n g t g V a r i a b l e , t g R e f e r e n c e */ s t r u c t c e l l * p l , *p2, *p3; 1 start: switch (pl->tag) [ case t g R e f e r e n c e : p i = pl->u.asReference; goto s t a r t ; case t g N i l : w h i l e (p3->tag = t g R e f e r e n c e ) { p3 = p3->u.asReference; } /* d e r e f p3 */ i f (p3->tag = t g V a r i a b l e ) { pushOnResetStack(p3); p3->tag = t g R e f e r e n c e ; p3->u.asReference = p2; } else ( error(); } break; case t g L i s t E l e m : w h i l e (p3->tag == t g R e f e r e n c e ) { p3 = p3->u.asReference; } /* d e r e f p3 */ i f (p3->tag = t g V a r i a b l e ) { push0nResetStack(p3); p3->tag = t g R e f e r e n c e ; p3->u.asReference = a l l o c L i s t E l e m ( p l - > u . a s L i s t E l e m . h e a d , allocVariable()); /* t a i l r e c u r s i v e c a l l . . . */ pi = pl->u.asListElem.tail; p3 = p 3 - > u . a s R e f e r e n c e - > u . a s L i s t E l e m . t a i l ; goto s t a r t ; /* t a i l r e c u r s i o n o p t i m i s a t i o n */ } else { error(); ) break; default: error();  22  3  Explanation.  The first case of append dereferences  pi  by guaranteeing that it is  not a reference cell. Once pi is fully dereferenced, one of the other cases is executed. If pi is nil, p2 and p3 are unified (the above code does only the situation where p3 is an uninstantiated variable). If pi is a list element, p3 is instantiated to be a list cell and the code iterates on the tail of the list element. The procedure call to push0nResetStack(p3) is actually a macro which expands into a test which decides whether or not the instantiation of p3 should be recorded so that it can be undone on backtracking (described in section 4.0, "Backtracking and delaying" on page 66).  This code implements a deterministic Prolog predicate which can either return a result (if the third argument is fully or partially uninstantiated) or test that the third argument is the correct answer, so the query must be slightly different: p3 is initialized here to be an uninstantiated variable. /* ?- a p p e n d ( [ l , 2 ] , [ 4 , 5 ] , L 3 ) . */ s t r u c t c e l l * p l , *p2, *p3; pi = allocLlstElem(allocInteger(1) , allocListElem(allocInteger(2), &nilCell); p2 = a l l o c L i s t E l e m ( a l l o c I n t e g e r ( 3 ) , allocListElem(allocInteger(4), &nilCell); p3 = a l l o c V a r i a b l e O ; a p p e n d ( p l , p2, p 3 ) ;  However, the result is different:  23  1  pi  ///  p3 — > 3  P  2  1 i  1  ///  3  1 i  1  *>  3  1 i  ///  1  2  3  ///  2  4  ///  ///  Not only is this second version slower, but it also produces extra "reference" cells. Therefore, the compiler should recognize situations similar to append 6  (which occur frequently) and optimize them as described above (this can beconsidered as a kind of peephole optimization on the abstract machine).  In fairness to the Warren Abstract Machine (WAM), this only happens if my scheme if individual value cells is used; if uninstantiated variables were stored within the heads and tails of list cells (instead of requiring separate cells), then no extra "reference" cells would be required, in this case. However, execution would still be slower because a check still must be made on each iteration for whether the third parameter is an uninstantiated variable, something which is already known (this avoids general unification and testing for whether the instantiation must be recorded on the reset stack).  24  2.4  Machine code for deterministic append predicate  Appendix A, "Sample machine code" on page 280 has sample machine code which translates the above C code to IBM/370 code. These transformations cannot easily be done by a C compiler, because certain values must be kept globally in registers for ultimate performance.  The best performance can be attained if the abstract machine status registers and the argument registers can be mapped into the target machine's registers. The xpPAM registers contain pointers to objects, so they translate in a natural way to IBM/370 registers (see section 3.0, "The basic sequential inference engine" on page 28 for the xpPAM status registers). Additional speed is possible, if the target machine allows some kind of object access, dereferencing and tag handling in parallel.  The inner loop of the second (slower) version is 19 or 23 machine instructions, depending on whether heap overflow is detected in-line or by an exception. The inner loop of the first (faster) version is 8 or 10 machine instructions, depending on whether heap overflow is detected in-line or by an exception. On a "1 MIPS" machine, the above loop will run at over 100 KLIPS (thousands of Logical Inferences Per Second).  25  The inner loop has the same instruction count, for either allocating from a heap or for allocating from a stack. For reference counting, three extra instructions are needed (load, add one, store).  7  2.5  Abstract machine code for deterministic append predicate  Skipping ahead a bit, here is the code used b y the abstract machine (xpPAM) for a purely deterministic predicate. This abstract machine code can be used to directly generate the C or assembler code given above (a slightly different version of this, with delays, is in section 4.4, "Delays" on page 75): swXVNL fO, n3, nO builtin 3 goto v a r goto n i l 1st: e q l s t f 2 , f 3 , n2  % LI = X . A % on f a i l u r e : 7 var % [ ] % _._  error  0  % % % % %  L3 = X . C argO: A ( a l r e a d y t h e r e ) argl: B (already there) arg2: c ( a l r e a d y t h e r e ) append(A, L2, C)  ICallSelf nil: eq f l , f 2 % L2 = L3 return var: % from h e r e on, s e t up f o r b a c k t r a c k i n g pushB vO pushB v l pushB v2  For this particular example, it appears as if reference counting is significantly slower than a marking garbage collector.  However, a marking garbage collector must eventually scan the  list and it probably will take more than three instructions to mark a cell. See section 7.12,  "Reference counts and garbage collection" on page  26  153.  mkCh eqlst goto mkch2: popB popB popB  mkch2 fO, n3, nO 1st  % LI = X . A  n2 nl nO  eq fO, [ ] goto n i l  % L I = []  The inner loop is three abstract instructions (swXVNL, e q l s t , I C a l l S e l f ) which can be reduced to two if the I C a l l S e l f is replaced by the swXVNL.  27  3.0 The basic sequential inference engine  Total grandeur of a total edifice, Chosen by an inquisitor of structures For  himself.  He stopsupon  this threshold  As if the design of all his words takes form And  frame from  — Wallace Stevens, To  thinking and. is realized.  an Old Philosopher  in Rome.  Brevis esse laborio, Obscurus  fio.  (I strive to be. brief, and I become obscure.)  — Horace, Ars  XpPAM  Poetica  consists of a basic sequential part which can be described almost  completely independently of the features which allow backtracking and delaying. The machine has 32 general registers, three stacks (execution, reset and backtrack) and a heap. The general registers are used for passing arguments. The execution stack is used to save status and registers across calls. The reset and backtrack stacks are used for backtracking and are described in section  28  4.0, "Backtracking and delaying" on page 66. The heap is used for allocating all objects.  3.1  Objects  All objects are stored the heap: the object table (oTab) and the extension area (xArea).  All objects are first class citizens and are tagged. For simplicity,  arrays and I/O objects (stream or direct) are not described here. Their implementation can easily be inferred from section 11.0, "Arrays and I/O done logically and efficiently" on page 2 2 2 ) . Any number of types can be added to the machine with no loss of efficiency (they are just extra cases in branch tables). The following are discussed: type  description  uninstantiated (or not ground) logical variable. reference  (pointer) to another object which is automatically dereferenced whenever it is accessed,  number  integer or floating point. The implementation has only floating point numbers.  8  string  commonly called an  nil  denoted"!]".  atom.  list element containing pointers to two objects, denoted " H e a d . T a i l " or "[Head|Tail]".  This is based on two considerations: it makes implementation easier and some modern machines can do floating point arithmetic nearly as fast as integer arithmetic.  29  paged out  object with pointer to backing store. Data in backing store never point to data in primary store. See section 3.15, "Virtual memory" on page 61 for details. Paged out objects are automatically paged in when accessed.  code segment contains abstract machine code (section 3.6, "Code segments" on  page 36). thunk  code pointer with environment [Ingerman 1961] (section 6.4, "Thunks, lazy evaluation and higher order functions" on page 127).  cut point  information for handling a cut.  unallocated on the free list (only used for heap overflow testing).  XpPAM uses structure copying because its implementation is simpler than structure sharing and it has similar space and time efficiency [Mcllish 1982]. There is, however, nothing in xpPAM's design which prevents using structure sharing instead of structure copying.  Each object is identified solely by its address within the object table  (oTab).  The address remains constant throughout the object's lifetime. Each object fits in afixedsize cell (about 12 bytes, depending on the base machine) containing:  •  tag (determines the object's type)  •  flags (meaning depends on the object's type)  •  reference count (or garbage collector marking bits)  •  data, consisting of one of: - one pointer - two pointers 30  - a numeric value - pointer into the extension area  (xArea)  plus additional information (such  as a hash bucket pointer for strings) - "age" and delay information for uninstantiated variables  All objects are kept in one heap (oTab) with an extension area  (xArea)  for large  objects (strings and code). All objects are subject to garbage collection — cells within  oTab  are simply added to the free list and corresponding space within  xArea  is compacted as needed. I implemented a reference counting garbage  collector because it allows reclaiming memory as soon as possible. A marking garbage collector could easily have been used instead (see section 7.12, "Reference counts and garbage collection" on page 153).  Numbers, nils and list elements fit entirely within these cells. Strings, code segment cells and thunks contain pointers into a separate area which is divided into segments and is compacted like Smalltalk-80's LOOM (see section 3.15, "Virtual memory" on page 61).  Although new strings can be created by the  concatenate  or substring operations,  most strings are constants which are known when the code is loaded. The loader ensures that only one copy is kept of each such constant string (fast lookup at load time uses hash values). Two constant strings are equal if and only if they are at the same address. Dynamically created strings require full character by character comparison. Constant strings and dynamically created strings are distinguished by a flag bit.  31  For simplicity, xpPAM treats functors as lists (as in micro-Prolog [Clark and McCabe 1984]) so that, for example, f ( a , b )  ~  [f,a,b].  Functors and lists are  distinguished by a flag in the head element so that a list element cannot unify with a functor element (however,  F(A)=f(l)  results in F=f,  A=l).  See section  7.10, "Functor and list storage" on page 151 for a fuller discussion of alternative representations.  3.2  General registers  The machine has 32 general registers. The number of registers is somewhat arbitrary and does not affect the overall design. [Auslander and Hopkins 1982] note that on a RISC with 16 registers, about half the programs need register spill code; with 32, fewer than 5% need spill code. This statistic, however, may not apply to an inference engine. Each of the registers contains either an object address or is flagged as being empty. A hardware implementation could keep a shadow cache of the objects referenced by the registers (that is, a register would contain the object's address and the shadow cache associated with the register would contain the 12 or so bytes of the object's value cell). The general registers are used primarily for passing arguments to predicates. Anrc-placepredicate will receive the addresses of its arguments in registers 0 through n - 1.  If a predicate has more than 30 parameters, it must be transformed by the compiler to code the last n-30 into a structure. The limit is 30 rather than 32 32  because some instructions require up to two extra registers (for example, splitting a list element into its head and tail may require two additional registers to hold the head and tail). p(..., P30, P31, P32,  Pn)  is transformed to p(... , (P30 . P31 . P32 . ... . P n ) )  3.3  Allocating and freeing object cells  New cells are allocated from a single free list. As all objects are the same size, this list never needs compacting (see section 7.3, "Global and local stacks" on page 142 for comparison with other methods). "Freeing a register" simply means flagging the register as empty. If reference counting is used, the reference count is decremented and, if it becomes zero, the cell is returned to the free list, possibly causing other cells to be freed. If a marking garbage collector is used, the register must still be flagged as empty so that the marking algorithm knows where to start and so that fail and delay work properly.  3.4  Status registers  In addition to the 32 "general purpose" registers which are mainly used for passing arguments, special registers keep the machine's status. Some of these are needed only for backtracking or delaying and are described more fully in the next chapter. 33  pc  program counter: contains the code segment and offset of the next instruction.  cpc  continuation program counter: contains the code segment and offset of the next instruction to be executed after a  toes  return  instruction,  top of execution stack.  ptoes protection for (top of) execution stack: toes value in the top frame on the  backtrack stack. linkr  link register: used by l i n k , c a l l and u n l i n k instructions to save the  toes  (needed for backtracking), tobs  top of backtrack stack,  tors  top of reset stack.  popbkr pop backtrack register: a flag and pointer into the choice stack which is set by a popBKeep instruction (it is set off whenever failure occurs). If this flag is on, a mkCh or mkChAt instruction will restore the choice stack point to what it was when the failure happened; that is, effectively pushing everything back onto the choice stack. orwaitr or-wait register: keeps the last "or wait" entry. It is null if no w a i t O r  instruction has been done since the last w a i t instruction.  3.5  Execution stack  The execution stack looks like this (growing from the top to the bottom of the page). Section 4.1, "Stacks for backtracking" on page 66 has a diagram showing all the stacks.  34  execution (call) stack xxxl  bottom o f s t a c k s  linkr cpc xxx2 xxx3 linkr cpc xxx4 xxx5  top o f s t a c k s  toes-  The "frames" contain information put there by the push (shown as x x x l through xxx5  above) and c a l l instructions (shown as l i n k r and cpc above). Each entry  on the execution stack is one of:  •  a value put there by a push instruction; or  •  a call stack frame: -  cpc in two slots (code segmentfirst,followed by offset),  - value used by the u n l i n k instruction to reset the stack frame after returning from the call; or •  a preemption entry (described in section 4.4, "Delays" on page 75).  35  The back pointers are not needed if strictly deterministic computation is done. They are needed to allow backtracking to "protect" parts of the stack.  3.6  Code segments  XpPAM instructions are kept within  code  segments.  Each code segment  corresponds to a single compiled predicate. It contains:  •  object address of the predicate's name.  •  number of parameters.  •  size of the constants' vector.  •  size of the code vector.  •  debugging information.  •  pointer to source (for debugging).  •  constants and code. (The constants' vector contains object pointers for all the constants used within the code. The code follows immediately after.)  The constants' vector contains pointers to objects. The loader always handles pointers to code segments by inserting an indirection via a reference cell; this allows replacing a code segment dynamically.  3.7  Machine instruction format  A xpPAM machine instruction fits within 4 bytes. The three operands take 18 bits (6 bits each) and the opcode plus flags can be encoded within the remaining 14 bits (the exact encoding is not important and will not be discussed here). In 36  the emulator, a looser packing is used for efficient extraction: the opcode is 1 byte, the flags are in 2 bytes and the three operands take 3 bytes (1 byte each). Some instructions could be packed tighter but such packing is not worth the extra time needed to extract fields.  The two formats are: •  three operands: Opl, op2 and op3 are each a register number or a constant number; the meaning is determined by the flag values. For some instructions, only one or two of the operands are used. opcode  •  flags  opl  op2  op 3  one operand and offset: Opl is a register number or a constant number; the meaning is determined by the flag values. For some instructions, opl is not used. opcode  flags  opl  1 offset i  A "register number" is a 1-byte quantity with a value from 0 through 31. A "constant number" is an index into the constants' vector for the code segment with a value from 0 through 63. In instructions, each register is annotated: v contains a value. 37  n empty, possibly requiring the allocation of a new object, f contains a value which is emptied after use. x  is empty and the value is unneeded  (n+f).  This could be used, for example,  when a head or tail value is not needed (the anonymous variable "_" in "_.Tail").  c constant (index into the code segment's constants' vector), s  (for the e q l s t or swXVNL instructions' first operands only) means structure head rather than list element.  V, n and c are mutually exclusive. The flags can be encoded by combining them with the opcodes — most instructions have only one or two possible flag combinations, so this would be the fastest possibility, at the expense of having many opcodes (the software emulator uses a slower method, by separately decoding the flags). F annotations mark where reference counts are decremented (if a marking garbage collector is not used). Whether or not reference counting is used, registers must still be flagged as empty to minimize the amount of information stored when choice points or "thunks" are created.  Each code segment has a vector of up to 64 constant objects' addresses. Most instructions allow c annotations to indicate that the operand is the index of an entry in the constants' vector.  38  3.8  Net effect of calls  When an «-ary predicate is called, the caller must save its registers on the execution stack and put the arguments in registers 0 through n-\.  The called  predicate must ensure that on return all registers are empty — the caller then pops the stack to restore the saved registers. That is, on entry to an /2-ary predicate only the first n registers are non-empty. On return, all the registers must be empty - the called predicate must explicitly free them.  3.9  Tail recursion optimization (TRO)  The pc and cpc are used to allow tail recursion optimization. The cpc contains a return address. A call must •  Push the active registers onto the stack.  •  Push cpc onto the stack.  •  Copy pc into cpc.  •  Set pc to be the first instruction of the called predicate.  A return must do the reverse: •  Copy cpc into pc.  • •  Pop the stack into cpc. Pop the saved registers from the stack.  39  In a call immediately followed by a return there is no need to do this push/pop sequence. A  last call  instruction acts just like a go  to.  The cpc is left alone so  that the called predicate will return to where the calling predicate would have returned.  Tail recursion optimization works only because parameters are passed in registers and not on the execution stack. If the last call is to the same predicate, the TRO works like iteration, with all the values' addresses in registers.  3.10  Unification  Unification is one of the fundamental concepts in logic programming. The abstract machine has several instructions which implement this. The essence of unification is:  •  If either operand is a reference cell, it is repeatedly dereferenced.  •  If both operands are atomic (numbers, strings or nil), they are compared for equality (which may fail). If the strings were known at load time (not created dynamically), they are compared by comparing their addresses (the loader ensures uniqueness).  •  If both operands are uninstantiated variables, the newest one is changed into a reference cell which points to the other operand (and possibly recorded on the reset stack). The "age" information is used for this (see section 4.1, "Stacks for backtracking" on page 66 for more details).  •  If one operand is an uninstantiated variable, it is changed into a reference cell which points to the other operand (this instantiation is recorded on the reset stack if the variable is older than the latest choice point). 40  •  If both operands are list elements or functor elements, unification is done recursively on the heads and tails of the operands.  •  Otherwise, unification fails.  Note that the above definition guarantees that a reference cell is always newer than the cell it points at, so there can never be any dangling pointers.  Unification does not do the occurs check (see section 10.6, "The occurs check" on page 217). There is no separate push-down stack for unification and deallocation because the Deutsch-Schorr-Waite algorithm is used [Knuth 1973] to traverse lists by reversing pointers using special tags which are not normally visible. These special tags are also used to terminate an infinite unification.  3.11  Instructions  Some instructions may fail or delay. These concepts are described in the next chapter. The assembler notation is a symbolic opcode followed by up to three operands. Each operand is a single letter (its annotation: v, n,f, x, c, s) followed by the register number or constant number.  9  Just to confuse you, the assembler has the single letter after the register number so that " e q n l , f 2 " is input " e q ( 1. n , 2. f ) " 41  The instructions are given with pseudo-code for interpreting them. There are five steps in interpreting: •  Decode the opcode.  •  Pre-process the operands into internal registers rl, r2 and r3.  •  Increment the program counter (pc = pc + l).  •  Execute the instruction.  •  Post-process the operands in the actual registers.  As mentioned above, every operand is one of: •  A register number  • •  A constant number An offset value  For the first two cases, a value is developed in one of the internal registers rl, r2 or r3, corresponding to operands 1, 2 or 3, according to the annotation: Annotation Action  v, f, s  Internal register receives the appropriate register's contents, then dereferences it if necessary,  n, x  Internal register receives the address of a new uninstantiated variable cell.  c  Internal register receives the address of the appropriate constant (by indexing into the code segment's constants' vector).  42  For most instructions, the internal register values are repeatedly dereferenced until a non-reference object is reached (the unification algorithm guarantees that there is never an infinite cycle of references). Exceptions include the push instructions.  After the instruction, the operands are processed according their annotation: Annotation Action  v, n, s  Do nothing (the results are already in the general register),  f, x  Decrement the object's reference count and mark the general register as being empty (this is not done for some cases of some instructions, such as  c  swXVNL).  Do nothing (it's not a register).  The internal registers are necessary to allow certain "overlapping" register combinations. For example, it is often convenient to have a list element (object pointer) in register 0 and then replace register 0 by the list element's head (the tail goes to another register). This can be accomplished by a single instruction (e.g.,  eqlst  f0,n0,n5).  The internal registers can not be blindly copied back to the operand registers because the internal registers may have been dereferenced - a future unification may need to point to the un-dereferenced object. Therefore, the instructions must operate directly on the general registers, using the values in the internal registers for input. Instructions must also increment reference counts as necessary.  43  The annotations c, v and n are mutually exclusive. The pseudo-code does hot show any test for this error - the assembler assures against invalid annotations.  The pseudo-code is not necessarily optimal. For example, the pseudo code for eq nl,v2 has the sequence:  •  Allocate a new cell for an uninstantiated variable and put its address in register 1.  •  Unify the objects pointed by registers 1 and 2.  •  Decrement the reference count of the object pointed by register 2 and mark register 2 as empty.  But the implementation simply copies the contents of register 2 into register 1.  Failure handling is described in section 4.2, "Backtracking instructions" on page 70. In the pseudo-code, failure is handled by calling d o F a i l ( ) .  Pseudo-code does not show pushing instantiations onto the reset stack (discussed in the next chapter) nor does it show adjusting reference counts. 3.11.1 Unification and testing instructions  eq op I, op2  Unifies the two operands (which may fail). The operands can be constants or registers. Eq can be used to unify values or to move or copy registers. Examples:  44  eq n l , f 2  moves register 2 into register  eq n 3 , v 4  puts a copy of register 4 into register 3.  eq n 5 , c 3  loads register 5 with the address of the 3rd constant in the code  1,  segment. eq v l , f 2  unifies register the object pointed by register  1  with the object  pointed by register 2, then frees register 2.  Pseudo-code: if  (!  unify(rl,  r2))  doFail()  eqskip opl, op2  Like eq except that it skips the following instruction if the unification (usually an equality test) succeeds (eq fails to the most recent choice point described in the next chapter). If the equality test fails, any / annotations are ignored. Failure does not undo instantiations caused by unification so there will usually be some tests before  eqskip  to ensure that the operands are not  variables. This instruction always appears to succeed (never "fails").  Pseudo-code: if  (unify(rl, pc = pc +  r2)) 1 /*  pc  already  points  at  next  instruction  */  testskip opl, op2  Tests if the two operands are equal.  Testskip  is the same as  eqskip  that it will not instantiate anything; if instantiation would occur,  45  except  testskip  does a delay for the uninstantiated variable (section 4.4, "Delays" on page 75).  eqlst opl, opHd,  opTl  Unifies thefirstoperands to a list element composed of the second and third operands. May fail. Examples:  E q l s t f I,n2,n3  tests for register 1 containing (the address of) a list element  (or, if it is a logical variable, instantiates it to a list element) with the head being put in register 2 and the tail in register 3. Register 1 is then emptied. E q l s t fO,nO,nl  replaces register 0 by the list element's head (the tail goes  into register 1). Pseudo-code: i f ( r l p o i n t s t o a l i s t element) { i f (! un.ify(rl->head, r 2 ) ) { doFail() j e l s e i f (! u n i f y ( r l - > t a i l , r 3 ) ) ( doFail() }  } e l s e i f ( r l p o i n t s t o an u n i n s t a n t i a t e d v a r i a b l e ) { r x = new l i s t element c e l l w i t h head = r 2 , t a i l = r 3 rl->tag = reference /* r l p o i n t s a t */ rl->ptr = rx /* new l i s t element */ } else [ doFail() }  46  3.11.2 Control instructions swXVNL  opl, opHd,  opTl  Jumps to one of the following four instructions that immediately follow it, according to the type of the first operand:  Atom (none of the following). The second and third operands are completely ignored  (including/annotations).  Variable. The first operand is not freed even if it has/annotation. The second and third operands are completely ignored (including/ annotations). Nil. The second and third operands are completely ignored (including/ annotations). List element. The second and third operands are unified with the head and tail of the list element. If during unification, some uninstantiated variables become instantiated but the entire unification fails, the variables are not reset to uninstantiated. Normally, this instruction is used to split a list element, the second and third operands having  n  annotation. The four instructions following the swXVNL  instruction are usually gotos or b u i l t i n 3 ("error"), except for the last one. Other switch instructions are possible for multi-way branching on strings or numbers; they can be emulated by sequences of eqskips. The first operand must have the annotation v or /(n is not allowed). N annotations for the second and third operands are treated like v annotations if the first operand is a variable or atom. 47  The first operand must have the annotation s if it is to handle a structure element. Pseudo-code: i f ( r l points to a uninstantiated variable) { pc = pc + 2; /* case v a r : don't f r e e */ } e l s e i f ( r l p o i n t s t o a l i s t element) [ i f ( u n i f y ( r l - > h e a d , r 2 ) && u n i f y ( r l - > t a i l , r 3 ) ) t h e n pc = pc + 1; /* " f a i l " w i t h o u t undoing any i n s t a n t i a t i o n s e l s e { pc = pc + 4; copy r 2 , r 3 i n t o r e g i s t e r s ; f r e e r e g i s t e r s as needed  3 } else pc } else pc  i f ( r l points to n i l ) { = pc + 3; /* case [ ] : don't f r e e r 2 , r 3 */ { = pc + 1; /* case d e f a u l t : don't f r e e */  }  caseGoto op/,  number  This is a generalization of swXVNL. The  number  is the number of cases w h i c h  follow (it is encoded like an offset). CaseGoto dereferences the operand. In the next  instruction space are two offsets:  1. where to go if the operand is an uninstantiated variable. 2. where to go if none of the other cases match (default case).  Following are the cases. Each consists of an object pointer and an offset. If the  constant i n a case matches the operand given by the caseGoto, execution  continues at the indicated offset within the current code segment. If more  48  than one case matches, it is unspecified which one is taken (this allows hashing or other non-sequential selection).  goto  offset  Unconditionally goes to the offset within the code segment if the first operand is an uninstantiated variable. Pseudo-code: pc =  offset  varGoto  opl,  offset  Goes to the offset within the current code segment if the first operand is an uninstantiated variable. Pseudo-code: if  ( r l points pc =  nonvarGoto  t o an u n i n s t a n t i a t e d  variable)  offset  opl,  offset  Goes to the offset within the current code segment if the first operand is not an uninstantiated variable. Pseudo-code: if  ( r l does pc =  n o t p o i n t t o an u n i n s t a n t i a t e d  offset  49  variable)  3.11.3 Call / return instructions  push opl  Pushes the operand onto the execution stack. This instruction is normally after a  link  and before a  call.  The operand is not dereferenced before  pushing.  Pseudo-code: *toes++ =  rl  /*  copy  rl  to  top of  stack;  increment stack p o i n t e r  */  pop opl  Pops the operand from the execution stack. This instruction is normally after a  call  and before an  unlink.  The operand must be a register with the n  annotation. Pseudo-code:  r  l  = *--toes  /*  copy top o f  stack  to  r l ;  decrement s t a c k p o i n t e r  */  gete opl,offset  Gets the operand from a specified offset within the execution stack. This instruction is normally used for register spill code. The operand must be a register with the n annotation. Pseudo-code: rl  =  toes[0-offset]  50  link Saves toes in linkr. If the ptoes (protection execution stack) value is above the  toes  value, toes is set to the protection value.  Pseudo-code: linkr if  =  toes  (toes  <  ptoes)  toes  =  ptoes  unlink Restores  toes  from  linkr.  Pseudo-code: toes  =  linkr;  return Restores the machine registers to their state when the c a l l was executed (see description under  call).  Pseudo-code: x = *toes--;  / * pop t o p  if  to  (x  points  pc =  }  else do  element  of  execution  stack  into "x"  */  {  cpc;  pc = * t o e s - - ; linkr  code) /*  = *toes--;  (pop 2 elements) / * pop l i n k  */  register  */  { s p e c i a l preempt  or  delay  return  }  51  (described  i n next  chapter)  call opl  calls (tries) a predicate. There are two varieties:  •  call  cC: call a known code segment whose address is given by the  constant C. The argument registers are assumed to be already loaded. Note that the constant is usually a reference to a code segment (this is what the assembler generates). c a l l fR; dynamic call - R points to either  •  — The argument list of which the name is the first element. The name is looked up and the registers are loaded with the elements of the list. - A thunk. The thunk specifies the first m of an rc-ary predicate. The first (n - m) registers are shifted right in places and the first m registers are loaded from the thunk. The register should have/ annotation to ensure that it is freed before the registers are loaded with the arguments.  Cpc pc  and  linkr  are pushed onto the execution stack,  pc  is copied into  cpc  and  is set to the first instruction in the code segment. R e t u r n does the inverse  by copying  cpc  into pc and popping the execution stack into  the registers must be empty when a return is done). Using  cpc  cpc  and  linkr  this way  allows the last call ( I C a l l ) instruction for tail recursion optimization.  Pseudo-code:  52  (all  *++toes = l i n k r ; /* push l i n k r */ *++toes = cpc; /* push cpc (two elements) */ cpc = pc; /* cpc = next i n s t r u c t i o n */ i f ( r l does n o t p o i n t t o code segment) { i f ( r l p o i n t s t o an atom) { name = r l ; n = 0; } e l s e i f ( r l p o i n t s t o a l i s t element o r s t r u c t u r e element) { name = r l - > h e a d ; n = number o f elements i n r l ' s l i s t ; } else error; s e a r c h f o r code f o r name/n p r e d i c a t e and f i l l t h e r e g i s t e r s 0 t o n-1 w i t h t h e arguments pc = f i r s t i n s t r u c t i o n i n thunk's code segment; }  pc = f i r s t  i n s t r u c t i o n i n code segment;  When a predicate is searched for, and it has not been created, a new predicate is created which calls "<undef>." The <undef> predicate can be tailored to the user's needs, for example, to just fail (as in conventional Prolog), to output an error message ("undefined predicate") or to prompt for a definition (a "query the user" facility). The operand for a dynamic call may be a functor or a list. X p P r o l o g allows call([Name,Arg])  instead of Name(Arg) (avoiding the need for univ ("=..")  for constructing a functor from a list. callSelf Like c a l l except that there is no need to switch code segments (a slight optimization). ICall opl Last (tail recursive) call. This is the same as c a l l except that it pushes nothing onto the stack. 53  ICallSelf Last (tail recursive) call to self. This is the same as g o t o  0.  It is included to  help in debugging. Pseudo-code: pc = 0  3.11.4 Allocation / deallocation instructions  free opl  Frees the operand. The operand must be a register with /annotation. The reference count of the object pointed at by the operand is decremented and the register is marked as free. This instruction is seldom needed because/ annotation in other instructions can free registers. new  opl  Allocates an uninstantiated variable in the operand (the register must be empty). The operand must be a register with/annotation. This instruction is used mainly for setting up arguments which return values from predicates; however, the p u s h is usually used instead with n annotation.  3.11.5 Miscellaneous instructions  builtin number  The  number  is coded as an offset operand which identifies the built-in. This  is used to extend the machine's instruction set with "foreign" instructions 54  such as arithmetic, string manipulation, i/o, etc. The various built-in instructions are somewhat idiosyncratic in how they use registers. They are described in an appendix. A built-in may fail or delay.  stop  Stops the top level interpreter, label  a pseudo-instruction for the assembler, comment  a pseudo-instruction for the assembler - ignored. The (not very useful) example:  3.12 [])• Code example p(N, p ( a . R s t , x.OutRst) :- p ( R s t , O u t R s t ) , q ( a , R s t ) , p ( b . R s t , y.OutRst) :- p ( R s t , O u t R s t ) .  is compiled to:  55  swXVNL  f0,n2,n0  fail  % s w i t c h o n p a r m O : Hd=>reg2, % invalid  parm  goto  var  % variable  goto  n i l  % parmO^f]  1st:  Rst=>regO  % parmO=Hd.Rst % %  eqskip  f2,V  goto  else  eqlst  f 1,  % %  Hd  i s now i n r e g 2 ( f r o m  swXVNL)  R s t i s now i n r e g 0 test  Hd =  fail  'a'  test  ..  go t o  "else"  X ,nl % 'x'.OutRst  link  % % %  vO  push call  P/2  pop  nl  save Rst p(Rst,  OutRst)  restore Rst into  required reg  unlink eq  q/2  % % %  f2,'b*  Jo t e s t  V  nO,  ICall  first  arg for  second q(a,  q/2  arg already  i n reg 1  OutRst)  se: eqskip fail  Hd =  Jo e l s e :  eqlst ICallSelf  fi> .  y  , n l Jo  'b'  invalid  parm  V.OutRst  Jo p ( R s t ,  OutRst)  nil: eq  f l , [] 1  1  Jo r e s u l t  :=  []  return var: ... c o d e  Note:  left  o u t f o r now  The notation  eqskip f2, ' a '  is not supported by the assembler.  10  Instead, the program would have to explicitly list out. the constants: a, x, b, y, [  1 0  T o allow the assembler to easily handle labels, a slightly different notation is used: the register number is given first, followed by the annotation. eq(l.n,2.c). 56  F o r example, e q n l , c2 would be coded  and  call(p,2).  n  Using this, e q s k i p  f2, 'a'  assembler as e q s k i p ( 2 . f ,o.c) (the constant "call  p/2"  would be presented as  would be presented to the "a"  is the Oth constant) and  call(5.c).  The code for handling a variable for the first parameter has been left out. The above code implements a deterministic predicate - if backtracking code were added, it would not affect the efficiency of the deterministic code. Backtracking code is described in section 4.0, "Backtracking and delaying" on page 66 and section 5.1, "Unoptimized compiling of a predicate" on page 88.  If this were marked as a "never fail" predicate, failure should give an error indication. To do this, the  fail  instructions could be replaced by  builtin  3  ("error") instructions. The  ICallSelf  instruction has the same meaning as g o t o  0  (the different opcode  helps in debugging). This is a tail recursive call - recursion has been turned into iteration.  1 1  c a l l ( p , 2) is used by the assembler and loader to look up the 2-arity " p " predicate.  A  reference to the predicate's code segment is created. If the predicate p / 2 has not yet been defined, it is created, calling the built-in undef which can use information from the registers to print out an error message or prompt for more input, before failing.  57  3.13  User defined unification  The current implementation implements simple syntactic unification. It does not require a local stack because pointer reversing (the Deutsch-Schorr-Waite algorithm) can encode the stack in the pointers (this algorithm has another advantage: because it changes tags on its way "down," infinite loops can be detected). Another possible implementation is to have the unification instructions handle only atomic values and to invoke a separate eqxx predicate for compound values (that is, list elements). This is defined: eqXX(X, Y) :- a t o m i c ( X ) ,  X=Y.  e q X X ( H l . T l , H2.T2) :- eqXX(Hl,H2), e q X X ( T l , T 2 ) .  which is compiled into swXVNL goto goto goto 1st: eqlst link push push  f0,n0,n2 atomic atomic atomic  % % % %  atomic var [] _._  fl,nl,n3 f2 f3  callSelf pop nl pop nO unlink ICallSelf atomic: fO,fl eq return  % % % %  save T l save T2 1 s t a r g a l r e a d y i n rO 2nd a r g a l r e a d y i n r l  % T2 % Tl  % g u a r a n t e e d a t o m i c / v a r 1st operand  58  A similar predicate would be needed for the t e s t s k i p instruction, being careful to not cause any unifications. This simplifies the abstract machine because unification is much simpler, possibly speeding up execution. Most unifications are for very simple cases, so this eqXX is seldom called. The intriguing thing about this design is that allows easily changing the unification to implement any kind of equality. In this way, functional programming could be easily added to a logic programming language. Also, new user data types can be added and integrated into the standard unification mechanism.  3.14  Other object data types  [Mukai and Yasukawa 1985] propose complex  indeterminates.  These have  many similarities with arrays, records and "frames" except that they are of indeterminate size. Here is a sample complex indeterminate: <fl/X, f2/abc,  f3/[a,b]>'  This has three fields:fil,f 2 and f 3 with associated data (X, abc and [ a , b ] , respectively). Field names must be strings; they may be given in any order. Here are some sample unifications: < f l / X , f2/abc> = <f2/abc, f 1/Y> < f l / X , f2/abc> = <f2/X> A=<fl/X>, B=<f2/abc>, A=B  = > = > = >  X = Y X = abc A = B = < f l / X , f2/abc>  A new field can be "added" to a complex indeterminate, as shown by the last example. Complex indeterminates are very useful for keeping associative information, such as symbol tables, syntax parts of a sentence or frames (for Al problem solving). 59  Complex indeterminates have not been implemented in xpPAM. I will probably be implement them as binary trees. Of course, they do not need to be built into xpProlog;  the user could explicitly unify them. But it is nicer to have them  built-in. If they are not built-in, user-defined unification can achieve the same affect.  New built-in object data types such as complex indeterminates can easily be added to xpPAM. Three changes must be made:  •  The storage allocator/deallocator must be modified. This is very simple, consisting mainly of code to compute the size needed in the extension area xArea.  •  The unification algorithm must be modified. This consists mainly of adding new cases. Thus, new object data types do not slow anything down.  •  New unification instructions (for example, similar to e q l s t or new switch instructions) to handle the extra data types.  To alleviate the work required to add new object data types, xpPAM could be modified to have more general forms for some instructions. For example, instead of swXVNL, there could be a more general switch statement which is simply followed by a list of tags and branch offsets.  60  3.15  Virtual memory  The virtual memory system is similar to Smalltalk's LOOM (Large Object-Oriented Memory) [Goldberg 1983]. As in most Smalltalk implementations, I have used reference counting (see section 7.12, "Reference counts and garbage collection" on page 153).  The heap uses an object table called  oTab  which is a vector of fixed size entries,  each containing the type flag, reference count and a variant part depending on the type. Numbers, nil, list elements, uninstantiated variables and references can all be contained entirely within a single oTab entry's variant part. For strings and code, the variant part is a pointer to a variable size second part which is allocated in an extension object area  xArea.  Functors could be stored as separate structures in the extension area  (xArea).  See section 7.10, "Functor and list storage" on page 151. Every object is referred to only by its index in oTab which is called an "address." For efficiency, the  xArea  part of an object may be locked for a short  time to prevent its being moved by the memory manager (this is normally only for code segments) but normally the memory manager may reorganize the items in  xArea  at any time so long as it updates the pointers from  oTab  into  xArea.  Because register contain only object addresses (pointers into oTab), the extension data for an object (in  xArea)  can be moved at any time, provided the pointer  from the object is updated.  61  The  xArea  area is divided into fixed size segments. Whenever a new object  cannot be allocated in the current segment, the next segment is compacted and the object is allocated in that segment. Because most objects have short lives, compaction usually moves only a few items - objects with long lives sink to the bottom of the segments and tend not to be moved. There is no noticeable delay when a segment is compacted because segments are relatively small.  To aid in compacting  xArea,  each object in  xArea  has a pointer back to the  object in oTab which needs it. Compaction is done by stepping through the objects in a xArea segment, skipping over any objects with null back pointers (freeing an object sets the back pointer to null, in addition to returning its  oTab  cell to the free list). If there is a locked object in the segment, compaction starts after it (the implementation allows only one locked object at a time).  The unallocated objects are kept in a free list within  oTab  with new objects  being allocated from the front of this list. All value cells except strings and code can be allocated entirely within single oTab entries. Code is fairly static and strings are only allocated and deallocated by explicit string operations like "concatenate" and "substring," so allocating and deallocating in xpPAM's heap is as efficient as pushing and popping a stack because the  xArea  area will seldom  need compacting (see section 7.3, "Global and local stacks" on page 142).  3.16  Object paged virtual memory  Full virtual memory design can be implemented similar to that in [Goldberg 1983]. Objects rather than fixed size pages are moved on demand between main memory and secondary memory. Objects in secondary memory may have 62  a different representation than objects in primary memory - for example, pointers could be larger in secondary memory, or functors could be kept as records instead of as linked elements (similar to "cdr-coding").  Objects in secondary memory never point to objects in primary memory. Objects in primary memory may point to objects in secondary memory. If an object in primary memory has an image in secondary memory, the primary image has a pointer to secondary memory - when an object is paged out, it goes to its original place in secondary memory.  When an object containing an address (for example, a list element or a reference object) is moved into primary memory, the addresses within it are set to point at special "in secondary memory" objects which contain the secondary memory addresses. Whenever such an object is used, xpPAM brings the object in from secondary memory. It is possible that an object in secondary memory is already in primary memory. To aid infindingthis, oTab entries have an additional field which is the secondary memory address (null if the object is only in primary memory). The  oTab  entry for an "in secondary memory" object is  picked by hashing the secondary memory address (this requires that the free list be forward and backward chained so that objects can be allocated from the middle of the list). In this way, an object can be brought in from secondary storage and efficiently transformed into its internal form.  A special "in secondary memory" object is used to reduce the number of entries in  oTab.  xpPAM  There is only one instance of this object. Whenever it is referenced,  cannot directly tell which secondary memory object is needed - it must  fetch a fresh copy of the secondary memory object which points at this object, 63  and use the secondary memory pointers contained there. This process can be optimized - the details are explained in [Krasncr 1983].  Reference counts are slightly different for paged memory. Each object in secondary memory has its own reference count. The reference count in primary memory is the number of primary memory references to the object. In addition, a delta reference count is needed to record the conversions between secondary memory pointers to primary pointers. The actual reference count for an object is the sum of these three numbers (again, for details, see [Krasner 1983]).  To allow paging out unused objects, each  oTab  entry has a "recently used" bit.  Whenever objects must be paged out, those with this bit set off are copied out and all the bits are set off. This is a primitive version of "least recently used." Clearly, hardware assist would improve the efficiency of this operation.  I have gone to the trouble of designing a separate virtual memory system because small machines do not provide virtual memory and I want xpPAM to run well on them, too. [Krasner 1983] points out that implementation which rely on a large heap in a standard paged virtual memory tend to degenerate to poor locality of reference so that eventually almost every memory reference produces a page fault and slows execution. However, there is a cost to this virtual memory scheme:  oTab  entries must be larger to contain the secondary memory  information and there will be about a 10% overall execution overhead.  Even when gigabyte real memories become reality, some form of virtual paged memory is still likely to be needed. The above techniques are merely suggestions about how object paged virtual memory can be added to xpPAM without affecting 64  its design very much. Anyone who wishes to implement large address spaces should review the literature - research is progressing rapidly.  65  4.0 Backtracking and delaying  O! call back yesterday, bid time return.  — William Shakespeare,  Richard  11, act 3  The deterministic inference machine is easily extended to allow backtracking by adding the backtrack (choice point) stack and reset stack ("trail").  4.1  Stacks for backtracking  The three stacks look like this (growing from the top to the bottom of the page): backtrack (choice) stack  reset stack (trail)  execution (call) stack bottom o f s t a c k s <-, < — i  //// ////  tobs—> <  <-,  //// ////  tors ptoes-  V top o f s t a c k s toes  66  >  Execution frames marked "////" are frames which are "protected" by the backtrack stack (using ptoes). In a stack machine for a conventional language like Pascal, these would have been reclaimed on procedure returns; with  xpPAM,  these are reclaimed only if the ptoes doesn't protect them, (described later in this section). The back pointers on the execution stack are used to skip over these protected pieces on return from predicates. Such back pointers are common on conventional machines although they can be avoided - for backtracking, they are essential.  The contents of the reset stack are put there by  •  unification, when a variable becomes instantiated: — object's age (pushed last) — object address  •  delaying and resuming predicates: — special value to indicate a delay (pushed last) — object address  When an uninstantiated variable cell is created (by a new instruction or by n annotation), its "age" is recorded. This age decides whether or not information will be pushed onto the reset stack when the cell becomes instantiated. The "age" is the depth of the backtrack stack when the cell was created. When unification instantiates a value cell which is older than the top choice frame, the cell's address is put on the reset stack. If the cell is newer, backtracking will free it anyway, so there is no need to record its instantiation. Deterministic predicates do not create reset stack entries because such predicates do not create  67  choice points. Any new variables created within a deterministic predicate are newer than the last choice point.  For non-deterministic predicate, choice points must be created on the backtrack stack using the mkCh ("make choice point") instruction. Each choice point frame contains sufficient information to reset the machine to the state it was in when the mkCh was executed:  12  •  where to go on failure (code segment pointer and offset) (pushed last),  •  cpc,  •  depth of reset stack,  •  execution stack protection point,  •  execution stack depth,  •  non-empty registers (put there by pushB instructions).  •  "age" of the previous backtrack entry (so that cut instructions leave the ages of uninstantiated variables correct).  If a choice point or continuation point is pushed onto a stack, the code pointer is always pushed on last. This is so that debugging and tracing can determine what is going on (there is no guarantee that, for example, code offsets and object pointers have disjoint ranges). Choice point frames are of variable size, depending on the number of pushB instructions used to save registers.  12  Actually, the "age" of the previous choice point must also be recorded if cuts are allowed, so that instantiations are properly recorded on the reset stack. Cuts can remove entries on the backtrack stack, so the "virtual choice depth" must be saved in backtrack entries.  68  When failure occurs, by a unification failing or by an explicit fail instruction, all registers are emptied and the top choice point frame is used to fill the registers. The reset and execution stacks are popped to what they were when the choice point was created. As the reset stack is popped, its entries are used to reset objects to uninstantiated (the age is recovered from the reset stack).' Execution 3  then resumes at the failure instruction address.  When backtracking occurs, the execution stack must be restored to what it was when the choice entry was made. This means that a return instruction may not pop the execution stack if a choice entry needs it - the choice entry "protects" the entry in the execution stack [Gabriel et al 1985]. The top of execution stack value in the top choice point frame is used to determine whether the execution stack can be popped. Each frame in the execution stack has a back pointer to the previous frame, skipping frames which are protected by the choice stack. If only deterministic predicates are executed, nothing is put onto the choice stack and the execution stack grows and shrinks just like the execution stack in a conventional machine (Algol, Pascal, etc.).  The backtrack stack could be embedded in either the execution or the reset stack. The WAM  combines the backtrack stack with the execution stack by  threading its entries among the execution entries - I would prefer to combine it with the reset stack because both stacks handle information which is only used  13  Destructive assignments can be handled by keeping a pointer to the previous value instead of just keeping the object's age. This can be useful for implementing compound objects such as arrays (see section 11.0, "Arrays and I/O done logically and efficiently" on page 222) in which changes are recorded on the reset stack.  69  on backtracking. For clarity of explanation and simplicity of implementation, the stack are kept separate. Simple instructions improve performance, even for interpreters. Therefore, a "call" is several instructions: link push call  % t o e s s k i p s over p r o t e c t e d frames % one push f o r each saved v a l u e  pop unlink  % one pop f o r each saved v a l u e % r e s e t t o e s below p r o t e c t e d frames  Similarly, "make choice point" is coded: pushB ... % one pushB f o r each non-empty r e g i s t e r mkCh l a b e l % push t h e f a i l u r e a d d r e s s , t o e s , t o r s label: popB ...  % one popB f o r each saved r e g i s t e r  For implementing  if-then-else,  the mkChAt instruction saves the  fobs  value in the  indicated register. A later c u t A t ("hard cut") pops the choice stack (and reset stack) back to the designated choice point; a rmChAt ("soft cut") changes the failure address to point to a fail instruction (if the choice point is at the top of the backtrack stack, the choice point is removed instead).  4.2  Backtracking instructions Some men a forward  motion love,  But 1 by backward steps would move.  - Henry Vaughan, The 70  Retreat, I. 29  pushB opl Pushes the operand onto the backtrack stack. The operand usually has v annotation (in which case the reference count is incremented) or/annotation. Pseudo-code: *tobs++ = r l /* copy r l t o t o p o f s t a c k ; increment  s t a c k p o i n t e r */  popB opl Pops the operand from the backtrack stack. The operand must be a register with the n annotation. Pseudo-code: r l = * - - t o b s /* copy t o p o f s t a c k t o r l ;  decrement s t a c k p o i n t e r */  popBKeep opl Pops the operand from the backtrack stack and sets a flag so that a subsequent mkCh or mkChAt instruction will reset fobs to the value it had when the last failure occurred (failure causes this value to be saved). The operand must be a register with the n annotation. Pseudo-code: r l = * - - t o b s ; /* pop b a c k t r a c k s t a c k */ p o p b k r - > f l a g = t r u e ; /* s e t popBKeep happened f l a g on */  71  mkCh  offset  Makes a choice point so that backtracking will continue at the instruction specified by the offset. Pushes toes, ptoes, backtrack stack. If ptoes is below  toes,  tors, cont  and  offset  onto the  it gets set to toes.  Pseudo-code: i f ( p o p b k r - > f l a g ) [ /* was t h e r e a popBKeep i n s t r ? */ tobs popbkr->ptr; /* r e s t o r e v a l u e p u t t h e r e by f a i l */ p o p b k r - > f l a g = f a l s e ; /* t u r n o f f f l a g */ , =  }  *tobs++ = *tobs++ = *tobs++ = *tobs++ = *tobs++ = i f (ptoes toes =  mkChAt  opl,  toes; ptoes; tors; cpc; /* two s l o t s */ pc; /* two s l o t s */ < toes) ptoes;  offset  Makes a choice point and saves information for a subsequent rmChAt or cut At. tobs  rmChAt  Same as mkCh but additionally creates a cut-point entry containing the  value which is put into  opl,  opl  (which must be a register with  n  annotation).  offset  Removes a specified choice point. Opl must be a register containing a cut-point entry (from mkChAt); the stack frame on the backtrack stack has its continuation point changed to that given by offset - usually, this is a series of popB instructions followed by a f a i l instruction. This instruction is used for "soft cuts."  72  cut At  opl  Cuts the choice stack back to a specified choice point. Opl must be a register containing a cut-point entry; the backtrack stack is popped to this point (as if a series falls were done). This instruction is used for "hard cuts."  chopBack Removes all unnecessary backtrack points from the top of the backtrack stack. Each backtrack "frame" whose reset stack pointer is above the top of the reset stack can be removed because the choice point did not instantiate anything, so no new information will be produced on backtracking.  fail Causes an unconditional failure. All registers are freed and the backtrack point is popped — the reset stack and execution stacks are popped to the points recorded on the backtrack stack (entries on the reset stack are used to un-instantiate variables or to put delay entries back onto the delay queue), the registers are loaded from the information in the backtrack stack, and execution proceeds at the place stored in the backtrack frame. Pseudo-code:  73  free p  all  registers;  = *--tobs;  /*  two s l o t s  */  cpc  = *--tobs;  /*  two s l o t s  */  tors  -  C  *--tobs;  ptoes  =  *--tobs;  toes  =  *--tobs;  uninstantiate (2 free  slots entries  or  4.3  one  to  for  on e x e c u t i o n  popbkr->flag = popbkr->ptr  return  each:  d e l a y queue object, stack  one  from r e s e t for  back to  stack  object's  back t o  "tors"  age);  "toes";  false;  = tobs;  /*  remember  backtrack  stack  position */  Cost of backtracking  Deterministic predicates run slightly slower on the full backtracking machine than on a purely deterministic machine. There are four overheads:  •  making choice points rather than just branching to a failure address for if-then-else.  •  link  and  unlink  instructions, which are not needed for deterministic  execution (nor are the execution frame back pointers). •  testing whether an instantiation should push an entry onto the reset stack (for deterministic execution, nothing will ever be pushed); similarly for recording delay information on the reset stack.  The first two items can be avoided by a smart compiler, using the switch instructions or by using knowledge of such predicates as " = ,"  " <," not,  etc. When backtracking is needed for if-then-else, some optimizations of and  pushB  push  instructions are possible. The c h o p B a c k can also be used before  74  return  or I C a l l instructions to remove unnecessary choice points (those which  did not instantiate anything, so they do not need to be retried) [Sahlin 1986].  Another possibility is to be able to switch the machine to deterministic mode. This would require using a different c a l l instruction (which does not assume link and unlink). In this way, xpPAM would function much like a conventional machine.  4.4  Delays  A delay is handled by using an instruction such as swXVNL or varGoto to detect that a value is uninstantiated - a d e l a y  r,offset  instruction then suspends the  predicate by saving a thunk (with all the non-empty registers) on the delay list associated with the variable and executes a  return.  The calling predicate continues execution until it instantiates a variable which caused a delay. The executing predicate is suspended and the delayed predicate is resumed. The resumed predicate will eventually return, so information about the newly suspended predicate is pushed onto the call stack - the r e t u r n instruction will use that information to resume it. The net effect is as if the instruction which woke up the delayed predicate had been replaced by link, c a l l delayed predicate, u n l i n k  and the delayed predicate's r e t u r n had returned  to the u n l i n k instruction.  There may be more than one predicate delayed on a variable, so the oldest such predicate is resumed and all others are pushed onto the execution stack after the suspended predicate. When the delayed predicate returns, the next delayed 75  predicate is resumed and so on until the suspended predicate is reached and resumed. This all happens automatically with the r e t u r n instruction.  When a predicate delays it must also be recorded on the reset stack so that it can be removed from the delay list on backtracking - when a delayed predicate is woken, it is recorded a second time on the reset stack so that backtracking can put it back on the delay list. A n optimization similar to tail recursion optimization is performed: if no choice point has been created since the original delay entry was created, both entries are removed from the reset stack (shuffling the stack down if necessary).  Putting all this together, let us trace f(b). f(c). ?- X * a , X * b , f ( X ) , X + d.  which is compiled to (X is in register 0):  1 2 3 4 5 6 7 8 9 10 11 12 13  eq nl, a link push vO call V / 2 pop nO unlink eq nl,'b' link push vO call V / 2 pop nO unlink link  % reg 0 = X % reg 1 = a  % save X % c a l l not-equal(X, Jo  restore X  % reg 0 = X % r e g 1 = 'b' % save X % c a l l not-equal(X, % restore X  76  14  p u s h vO  % save  15  call  f/2  % call  nO  % restore X  16  pop  17  unlink  18  ICall  % last  V / 2  X f(X)  c a l l not-equal(X,  'd')  V / 2 must be a separate predicate. If it were expanded in-line, it would cause a delay in the entire predicate, rather than just in the inequality.  14  21  testskip  22  fail  23  return  f/2 is  fO,fl  % delays  if  % reverse  necessary  success/failure  1 5  31  p u s h B vO  7  32  mkCh  % make c h o i c e  33  eq  34  return  35  else:  0  else  save  1st  arg point  fO,'b'  popB  nO  36  eq  f0, c'  37  return  1  When the first 'V is executed, it delays because its first argument (X in register 0) is uninstantiated. A delay object is built, containing the values of the registers (X and this object. A  1 4  'a')  return  and set to redo the  testskip  instruction  (#21). X  points to  is simulated. Execution continues to X?tb which also delays  T h e current implementation has a built-in for '  / 2 which delays as necessary.  Experience  with this built-in lead to the definition of testskip which has not equals as a special case.  1 5  This is not very optimized; a test should be first done for the parameter being instantiated, thereby avoiding the creation of choice point.  - the delay object is added to the list pointed from x. Finally, f (X) is executed which creates a choice point and instantiates X=b. At this point, we have the following situation for the stacks and x. r e s e t <-  backtrack  =>35  X  V >(=>21, (X,B))  execution  =>16  V >(=>21, ( X , c ) )  Note that the simulated returns have popped the execution stack from the two calls to ±. We haven't reached the r e t u r n instruction, so the return information hasn't been popped (even if the r e t u r n had been reached, it still would not be popped because the choice point "protects" it). The notation =>2l means that there is a code pointer to instruction 21.  At this point, the first delay entry pointed from x is tried (it is removed from the list pointed from x but left on the reset stack, marked as having been re-started). It succeeds. The next delay entry is tried but it fails. Backtracking rebuilds the delay entries for X and then jumps to instruction #35. The popB restores the saved register environment (just one register in this case) so that the machine is now in the same state as when the choice point was created (including delay entries). From here on, execution is deterministic because there is no choice point left on the backtrack stack (the reset stack entries will also disappear as the ' V ' s are resumed).  78  To allow or-delays, one or more d e l a y O r instructions may precede a d e l a y instruction. They add information to the delay list entry which is finished by the d e l a y instruction. When the d e l a y instruction is executed, it places the delayed predicate on the delay lists for all the variables which can cause it to resume.  As another example, consider append The code is: % % % %  ?- proceed append(A?, B, C ) . ?- proceed append(A, B, C ? ) . append(X.A, B, X.C) :- append(A, B, C ) . a p p e n d ( [ ] , X, X ) .  code(append, 3, [ /* 0 */ c o n s t ( [ ] ) ] , [swXVNL(0.f, 3.n, O.n), % L I = X . A builtin(3), /* on f a i l u r e : e r r o r / 3 */ goto(HVar), /* v a r */ goto(HNil), /* [ ] */ label(HLstEl), /* _._ */ e q l s t ( 2 . f , 3 . f , 2.n), % L3 = X . C % argO: A ( a l r e a d y t h e r e ) % argl: B (already there) % arg2: c ( a l r e a d y t h e r e ) ICallSelf, % append(A, L2, C) label(llNil), eq(l.f, 2.f), % L2 = L3 return, label(HVar), nonvarGoto(2, m k c h l ) , delayOr(0, 0 ) , % delay(L0, delay(2, 0 ) , % L2) label(mkchl), pushB(O.v), p u s h B ( l . v ) , pushB(2.v), mkCh(mkch2), e q l s t ( 0 . f , 3.n, O.n), % L I = X . A goto(llLstEl), label(mkch2), 79  p o p B ( 2 . n ) , p o p B ( l . n ) , popB(O.n), e q ( 0 . f , O.c), % L I = [] goto(llNil)]).  The code from l a b e l ( H V a r ) on handles delays. The delay restarts at the beginning so that the swXVNL instruction will be re-tried. The code from label(mkchi)  on sets up for backtracking. Each choice point instantiates the  first argument to one of the possibilities (a list element or nil) then jumps to the deterministic code for that case.  4.5  Cost of delaying  The current implementation has a slight cost when instantiating a variable: it checks to see if the variable has caused a delay. This can be changed by making a new object type: "uninstantiated variable which caused a delay." As mentioned earlier, adding new object types causes no slowdown because it results in just adding extra cases to a branch table. Currently, when unification instantiates a variable, a flag is set to indicate that a delayed predicate is eligible for waking up. This means that every xpPAM instruction must check to see if it should be preempted. This is potentially expensive. There are three possibilities:  •  Only check for woken predicates at expensive instructions, such as call or return.  •  Allow instructions to be halted in the middle and later resumed. This adds significantly to the abstract machine's complexity. 80  •  Do complex unifications by a user-defined "eq" predicate (see section 3.13, "User defined unification" on page 58.). This is probably the best solution because it turns unification into a non-atomic operations which can be delayed at any time by already existing mechanisms.  The overhead for providing delays is quite low: about 1%. Because delays can speed up some predicates by an order of magnitude or so, the cost is worthwhile.  4.6  Weak delays: dynamic reordering of clauses  The ancestor predicate is inefficient if the first argument is uninstantiated: a n c e s t o r ( A n c e s t o r , Descendent) :p a r e n t ( A n c e s t o r , Descendent). a n c e s t o r ( A n c e s t o r , Descendent) :parent(Ancestor, Z), a n c e s t o r ( Z , Descendent). ?- a n c e s t o r ( X , g e o r g e ) .  In the second clause, p a r e n t with two uninstantiated variables will repeatedly generate all p a r e n t relations by backtracking. Changing p a r e n t to delay until both parameters are instantiated would prevent this and give efficient execution, but would also cause a n c e s t o r to delay permanently if it is called with two uninstantiated variables, for example when computing ancestors beyond grandparent. To handle this, xpPAM allows an associating a "cost," proportionate to the size of the predicate's solution space, with each delay instruction. For example, if there are 100 parent-child relationships, an average of 2 parents per child and 3 children per parent, the code would be: 81  varGoto rO, L I varGoto r l , L2 LO: ... code f o r b o t h rO and r l i n s t a n t i a t e d . return L I : varGoto r l , L3 delayCost 2 % d e l a y parmO, cost=2 delay rO, L l a % resume a t next i n s t r L l a : n o t v a r G o t o rO, LO % parmO p o s s i b l y v a r ... code f o r rO u n i n s t a n t i a t e d and r l i n s t a n t i a t e d . L2: d e l a y C o s t 3 % d e l a y p a r m l , cost=3 delay r l , L2a L2a: n o t v a r G o t o r l , LO ... code f o r rO i n s t a n t i a t e d and r l u n i n s t a n t i a t e d . L3: d e l a y C o s t 100 % delay parml, delayOr r l % o r parm2, cost=100 delay rO, L3a L3a: varGoto rO, L2 n o t v a r G o t o r l , LO ... code f o r r l u n i n s t a n t i a t e d and r l u n i n s t a n t i a t e d .  As before, the machine resumes clauses when their arguments are sufficiently instantiated. If all predicates are blocked, the least expensive one is resumed. The machine thereby dynamically decides the least expensive way to continue a non-deterministic computation.  Ordinary delays could also be treated something like weak delays with extremely high costs. Usually, when computation halts with some predicates still delayed, the answer is not "yes" or "no" but "don't know." Instead, the delayed predicates could be allowed to resume, to try to get to a definite answer. Unfortunately, this will often lead to infinite loops.  82  4.7 Delay instructions delay opl, offset  The object pointed by opl is flagged as having caused a delay (the object must be an uninstantiated variable). A delay object is created (pointed to by the  opl  object) with sufficient information to restart the predicate at  offset.  If a d e l a y O r instruction had been executed, its delay object is linked in. Finally, a r e t u r n is simulated.  delayOr opl, offset  The object pointed by opl is flagged as having caused a delay. A delay object is created (pointed to by the opl object) with sufficient information to restart the predicate at offset. A flag is set so that a subsequent d e l a y instruction will link in this delay object. Execution continues at the next instruction.  delayRec opl, offset  Like d e l a y except that a delay is made if an uninstantiated variable is found anywhere within the operand. The delay test is recursive on compound objects (list elements, etc.). delay Cost opl, cost  Like d e l a y but with an associated cost. See section 4.6, "Weak delays: dynamic reordering of clauses" on page 81)  83  mkThunk  opl, opl, ftregs  Creates a thunk in opl from opl The operand may be either a constant (a code segment) or a thunk. The thunk is constructed from thefirstjf-regs registers plus any registers specified by the operand (if it is a thunk). For more details, see section 6.4, "Thunks, lazy evaluation and higher order functions" on page 127.  84  5.0 Compiling Prolog  There are  only two  qualities  efficiency  — George Bernard Shaw,  and  in the  world:  inefficiency  ...  John Bull's Other Island, act  4  XpPAM  is designed to make compilation easy. My simple compiler from Prolog to  xpPAM  is about 600 lines of Prolog code - a full compiler from x p P r o l o g to  xpPAM  is about twice that, to take care of delays, if-then-else, etc. Optimization  requires considerably more analysis which is essentially independent of the target machine (5000 or so lines of Prolog ). 16  Compiling from Prolog to xpPAM is somewhat different from compiling to  WAM.  In particular, allocating "local" and "global" variables - is a problem which does not exist with xpPAM. Some approaches to compiling Prolog to WAM  are  given in [Debray 1985], [Debray and Warren 1986] and [Van Roy 1984].  1 6  This number is derived from the compiler which is described later in this section.  About  2000 lines of Prolog did most of the optimizations described here; I estimate that the full compiler would be about twice as large, for the same reason that the full x p P r o l o g compiler is about twice as large as the simple Prolog compiler.  85  Three types of optimization are possible for logic programs: Clause optimizations include common expression elimination and removing unnecessary register copying. These are fairly straightforward — they are similar to "peephole" optimizations in conventional compilers. Predicate optimizations mainly consist of detecting determinism and common sequences of goals among clauses; in effect, turning clauses into  if-then-else  form where possible. To do a good job, these must have knowledge of predicates such as  equals  and  not  equals.  Global optimization include flow analysis, mode inferencing, delay inferencing and combining predicates (to avoid the overhead of calling "helper" predicates). These can become quite involved, although they are considerably simpler than similar techniques for conventional programming languages because of the lack of destructive assignment. The programmer can help the optimizer. As predicate optimization consists partly of turning clauses into if-then-else form, the programmer can, if he wishes, use  if-then-elses  from the beginning. However, the programmer's style  may use conventional clause form without penalty.  Although some Prolog predicates can be run "backwards" (such as  append  for  splitting lists), many cannot. The programmer may indicate this to the compiler. In the absence of full global type inferencing, the compiler must still generate code for dynamically detecting uninstantiated variables but it can save considerably on the amount of code generated and help debugging by generating calls to error instead of backtracking and failing.  86  There is almost no limit to the amount of optimization that can be done. For example, l i n k and u n l i n k instructions can be left out if global analysis determines that the predicate is never called in the context of a choice point. However, such a level of optimization is probably not desirable because these instructions are not very expensive and because such optimizations require substantial checking whenever a new predicate is added.  In the following discussion, clause, predicate and global optimizations are described as separate processes. However, they may often be combined.  The compiler is written entirely in Prolog. The raw code for a predicate is stored in an internal form which is either compiled right away or compiled on first use. For example,  pred(n,  A, B.C) :- z o t ( A ) . pred(a, [ ] , [ ] ) .  is stored internally as: [ [ [ c o n s t . p r e d , c o n s t . [ ] , name.'A', l i s t . ( n a m e . ' B ' ) . ( n a m e . ' C ' ) ] , [ c o n s t . z o t , name.'A']], [ [ c o n s t . p r e d , c o n s t , a, c o n s t . [ ] , c o n s t . [ ] ] ] ]  The entire predicate is turned into something like the following, where predClauses  is the internal form of the clauses (described in the previous  paragraph):  87  pred(Pl,P2,...,Pn)  :-  compilefpredClauses. /*  compile  Name,  instantiates  NArgs,  Name,  replacePred(pred(Pl,P2,...,Pn), replacePred(Pred,  Code)  addCode(Name,  NArgs,  call(Pred).  % calls  ConstList,  Name,  Code  */  NArgs, C o n s t L i s t ,  Code)  :ConstList,  Code),  compiled version  of  Pred  Note that the call to replacePred is tail recursive. pred  Code),  NArgs,ConstList,  with the newly compiled code.  ReplacePred  AddCode  replaces the body of  then tail recursively calls that  code. This extra step is necessary because otherwise the machine would be attempting to execute code while it was being replaced.  5.1  Unoptimized compiling of a predicate  For simplicity, I will consider only clauses without any imbedded "or"s or Predicates containing "or"s can always be transformed by  if-then-elses.  producing auxiliary predicates. For example: p(A,  B)  :-  q(A,  Z),  (r(A,  B)  ;  s(Z)),  t(A).  can be transformed to p(A,  B)  p2(A, p2(_A,  :-  B,  q(A,Z),  _Z)  _B,  If-then-elses  Z)  p2(A,  :-  r(A,  :-  s(Z).  B).  B,  Z),  t(A).  /*  _Z is  /*  _A, _B,  unused are  */ unused  */  are generated as an optimization and are treated later.  The general form of a compiled predicate is:  88  make choice point first  clause  modify choice point second clause  modify choice point third clause  remove choice point last clause  If there is only one clause, choice points are not required. If there are two clauses, there are only "make choice point" and "remove choice point" instructions (no "modify choice poinf's). The general compiled form of a clause is unify arguments in head build arguments for first goal call first goal build arguments for second goal call second goal build arguments for last goal call last goal tail recursively  89  If there are no goals, a " r e t u r n " , instruction is generated after the arguments in the head are unified. If the last goal is to itself, a last-call-self (in effect, goto) instruction is used.  5.2  Basic compiling of a clause  At the beginning of a clause, the first n registers are occupied by the n arguments to the predicate. All other registers are assumed to be empty. As described in section 3.2, "General registers" on page 32, if there are more arguments than machine registers, the extra arguments must be combined into a structure in the last argument.  Each argument which is a simple variable name is assigned to the appropriate register, except when the name is used elsewhere in the head (this prevents some pathological cases for trace analysis; see section 5.5, "Optimized compiling of a predicate" on page 98). All other names are assigned sequentially from the unused registers. For example, in p ( a b c , A, Z, B, X.Y, B, X) :- q ( x y z , X, X, A.C), r ( Z , C ) .  the following register assignments are produced from the head (_rn is a temporary name for register n): Register Variable  Comment  0  _r0  abc is treated as an expression  1 2  A  3  _r3  Z  B is also used in another argument  90  Register Variable 4  _r4  5  _r5  6  _r6  7  B  8  X  9  Y  Comment  These register assignments will change as the goals are processed. C is not used in the head and therefore has no initial register assignment.  The clause is transformed into a series of equalities, for A=(B.c). p ( _ r O , _ r l , _ r 2 , _ r 3 , _ r 4 , _ r 5 , _ r 6 ) :_ r 0 = abc _rl = _rl /* A */ _r2 = _r2 /* Z */ _r3 = _r7 /* B */ _r4 = _r8 . _r9 /* X.Y */ _r5 = _r7 /* B */ _r6 = _r8 /* X */ /* r e g i s t e r assignments from here have not y e t been determined */ startCall(4) /* q has 4 parameters */ _r0 = xyz _rl = X _r2 = X _r3 = A . C /* _ r c = A.C */ c a l l q/4 startLastCall(2) _r0 = Z _rl = C lastCall(r/2)  91  More complex expressions are broken down into individual unification steps (this is sometimes called "partial evaluation." For example, [A, (B.C),D] is handled by _templ _temp2 _temp3 _temp4  = = = =  (A._temp2) (_temp3 . _temp4) (B.C) (D.[])  where the temporaries are also registers. For calls, the order is reversed to build up an argument: _templ _temp2 _temp3 _temp4  = (B.C) =.(D.[]) = (_templ . _temp2) = A._temp3  Once everything has been assigned, the list is scanned from beginning to end, adding n (new) every time a variable name is met for thefirsttime; it is then scanned from end to beginning, adding/(free) every time a variable is met for the last time. Each  startCall  or  startLastCall  clears all the "new" indicators but  not the "free" indicators — it gets a list of all the variables needed after it (created during the backward scan and pruned during the forward scan). It is possible to have both "new" and "free" for the same variable; typically this happens when an anonymous variable ("_") is used. (This step is actually combined with the following step, for efficiency.) Variables with neither/nor n already have a value which is needed later. p(_rO, _ r l f._rO f._rl f._r2 f._r3  , = = = =  _r2, _r3, abc _ r l _r2 n._r7  _ r 4 , _r5,  _ r 6 ) :-  /* A */ /* Z */ /* B */  92  f . _ r 4 = ( n . _ r 8 ) . ( n . _ r 9 ) /* X.Y */ f._r5 = f._r7 /* B */ f._r6 = _ r 8 /* X */ s t a r t C a l l ( 4 , [ f . Z , n.C])/* w i t h l i s t o f v a r i a b l e s t o save */ n._rO = x y z n._rl = X n._r2 = f.X n._r3 = ( f . A ) . ( f . C ) /* A.C (C was "new" a t s t a r t C a l l ) */ call(q/4) startLastCall(2) n._rO = f.Z n._rO = f.C lastCall(r/2).  The variables used within calls can now be assigned registers. A call is transformed into a series of pushs, followed by the c a l l , followed by pops into the appropriate registers for the following call. Sometimes, registers must be moved to make room for the parameters for a call. Some equalities can be eliminated or turned into f r e e or new instructions. The example now becomes: p ( _ r O , _ r l , _ r 2 , _ r 3 , _ r 4 , _ r 5 , _ r 6 ) :f._rO = abc /* _ r l = _ r l */ /* A */ /* _ r 2 = _ r 2 */ /* Z *./ f . _ r 3 = n._r7 /* B */ f . _ r 4 = ( n . _ r 8 ) . (n. _ r 9 ) /* X.Y */ f._r5 = f._r7 /* B */ f._r6 = _ r 8 /* X */ link /* s t a r t C a l l ( 4 , [ f . Z , n . C ] ) push(f._r2) /* Z: n o t used i n c a l l */ push(n._r4) /* C (new) */ n._rO = x y z n._r5 = f . _ r l /* get A out o f t h e way *•/ n._rl = _r8 /* X */ n._r2 = f . _ r 8 /* X */ n._r3 = ( f . _ r 5 ) . ( f . _ r 4 ) /* A.C */ call(q/4) unlink 93  pop(n._rl)  /* C */  pop(n._rO)  /*  Z */  /*  startLastCall(2)  */  /*  n._rO =  f._rO */  /*  Z:  already  i n reg 0  /*  n._rl =  f._rl */  / * C:  already  i n reg  */  1 */  lastCall(r/2).  Note how the  pops  get the arguments for a call into the correct registers. The  only time that registers must be moved is for the first call in a predicate (for example, A had to moved in the above code). In all cases, p u s h e s and  pops  will  produce the correct effect.  The final stage transforms this into the form used by the built-in by collecting the constants (note that the  code  code  predicate  predicate reverses the order of the  register number and its annotation - there is no good reason for this, beyond making the program for processing code( , P  7,  /*  constant  /* /* /*  / * p/7  [const(abc),  constant  1 */  const(xyz),  constant  2 */  call(q,4),  constant  3 */  call(r,2)],  O.c),  eq(3.f,  7.n),  eqlst(4.f,  8.n,  eq(5.f,  l.f),  eq(6.f,  8.v),  a little simpler):  */  0 */  eq(0.f,  code  9.n),  abc  /* /* /* /* /* /* /*  X  /* /*  C  */  A */ Z  */  B */ X.Y  */  B */ */  link, push(2.f) push(4.n), eq(0.n,  l.c),  eq(5.n,  l.f),  eq(l.n,  8.v),  /*  /* /*  Z:  not used i n c a l l (new)  arg  0 = xyz  get  A out  arg  1  94  */  */ */  of  = X */  t h e way * /  eq(2.n,  8.f),  eqlst(3.n,  5.f,  4.f),  /*  X */  /*  A.C * /  call(2.c),  /*  q/4  pop(l.n)  /*  C */  pop(O.n),  /*  Z */  /*  Z:  already  i n reg  0  /*  C:  already  i n reg  1 .*/  */  unlink, */  /* r/2 */  lCall(3.c)  If any delay annotation exists (either by "?"s or from proceeds), they can be easily incorporated by adding  vargoto, nonvargoto  and  delay  o instructions, as  appropriate. These will be discussed in more detail below.  5.3  Optimized compiling of a clause  The above method of compiling a clause can produce some inefficiencies: •  Values may be moved unnecessarily between registers. Above, there was no need to move B from register 3 to register 7 when unifying the head of the clause. As another example, fo.o(A)  :-  A=B,  C=x,  B=C.  can be optimized to just foo(_rO)  •  : -  foo(x)  (which is actually handled as  _rO=x).  Common expressions are not noticed. For example, ordered(A.B.C)  :-  A =<  B,  o r d e r e d ( B . C)  95  is transformed to  o r d e r e d ( _ r O ) :- _rO = ( A . _ t e m p l ) , _templ = (B.C), A =< B, _temp2 = ( B . C ) , ordered(_temp2).  But _temp2 is the same as _templ and can be eliminated. The compiler code for detecting these common expressions uses uninstantiated variables in an interesting manner. Each expression is replaced by an uninstantiated variable. If that expression is found later, it is replaced by a new temporary and  _temp=expr  is added to the body; if the expression is not found later, the  uninstantiated variable is replaced by its original value. It should be noted that the expressions are always very simple because everything has already been broken down to either simple "equals" or "equals list element" form. •  The sequence of instructions new(_rn), eq(_rn,X)  can have the "new" removed and the "eq" replaced by "eq(_rn.n,x)." •  Some goto or  last call  instructions can be replaced by their target  instruction(s). These optimizations can be done in a straightforward way although they are not all trivial. Sometimes, they can be combined with the other steps, to avoid extra passes over the clause.  96  5.4  Spilling registers within a clause  The previous sections have assumed that there are enough registers in xpPAM to hold all the variables in a clause. Although 30 is a rather large number for predicates written by humans, it is possible that machine-generated predicates could contain many more.  The push and pop instructions can be used to save and restore registers using the execution stack. The code generation scheme above makes the l i n k and push instructions for a call as late as possible. If register spilling is needed, the l i n k must be made earlier (if there is a l i n k ) and pushes must also be made earlier. In this way, some registers may become free and assignable to other variables. If there are still not enough registers, more drastic action must be taken. Consider compiling p(A.B, C D ) :- q ( A . C ) , r(A.B.C.D).  on a machine with only four registers. The code would be:  97  e q l s t f0,n2,n3 link push f 3 e q l s t fI,n3,n0 push v2 push v3 push fO eqlst n0,f2,f3 c a l l q/1 pop n3 pop n2 eqlst n0,f2,f3 pop n2 pop n3 eqlst n0,f0,f3 e q l s t n0,f0,£2 unlink I C a l l r/1  /* a r g 0 = A.B */ /* save B */ /* arg 1 = C D */ /* save A */ /* save C */ /* save D */ /* a r g 0 = A.C */ /* q(A.C) */ /* r e s t o r e D */ /* r e s t o r e C */ /* rO = C D */ /* r e s t o r e A */ /* r e s t o r e B */ /* rO = B . C D */ /* rO = A.B.CD */  Clearly, the pushing and popping on the execution stack can become quite complicated. To simplify this, an arbitrary element on the stack can be accessed by the g e t e instruction. Using gete the execution stack can be treated much like a stack frame with local variables in a conventional stack machine. Unneeded entries on the stack can be removed by pop x 3 l which pops the value and discards it.  5.5  Optimized compiling of a predicate  One of the most important optimizations is to detect potentially deterministic situations and transforming them into  if-then-elses.  If this is not done,  unnecessary choice points are created. Not only are these expensive to create and delete but they also result in unnecessary computations on backtracking. Optimization should produce code without any choice points when a predicate is 98  called with all arguments sufficiently instantiated - in such situations, no new information is produced by re-trying the predicate (it would simply succeed again, or fail if there were no other way to succeed).  Detecting  if-then-else  situations requires first putting the clauses into "standard  form" and then computing "traces." A trace is a "most general form" for the parameters of a clause - it is used for finding potentially deterministic parts. For example, consider (1)  tn([],  (2)  m(M,  M,  (3)  m(X.A,  Y.B,  X.M)  :-  X=<  Y,  m(A,  (4)  m(X.A,  Y.B,  Y.M)  :-  X >  Y,  m(X.A,  •[],  M). M). Y.B, B,  M). M)  The traces are: (1)  [],  (2)  _,  _ [],  []  (3)  _.  (4)  _.  The traces can be partitioned on the first argument: []  (l)  _._  (3),  (2) (4)  Clause 2 can match anything, including [ ] and  so it must be placed in an  alternate  which is a non-deterministic choice. The other clauses can be placed in  a  which is deterministic if the switch value is instantiated. This results in  switch  99  alternate s w i t c h _rO case [ ] : clause 1 case X.A: alternate clause 3 else clause4 end-alternate end-switch else clause 2 end-alternate  This can only be done if the compiler can re-arrange the order of the clauses (valid in a pure logic programming language). We can make m deterministic by changing clause 2: (1) (2) (3) (4)  m ( [ ] , M, M). m(X.A, [ ] , X.A). m(X.A, Y.B, X.M) :- X=< Y, m(A, Y.B, M). m(X.A, Y.B, Y.M) :- X > Y, m(X.A, B, M)  The partitioning on the first argument is: []  (l) (2) , ( 3 ) , ( 4 )  The second set can be further partitioned by the second parameter: []  (1) (3) , ( 4 )  100  Clauses 3 and 4 can be partitioned because "=<" and ">" are disjoint. This results in switch _r0 case [ ] : case X.A: switch _ r l case [ ] : case Y.B: i f X =< Y clause 3 else clause 4 end-switch end-switch  /* CD */  /* ( 2 ) , ( 3 ) , ( 4 ) */ /* ( 2 ) */  /* O), ( 4 ) */  Some care must be taken in analyzing traces. A case may not contain a variable which was "bound" by a earlier case. For example, na'ive analysis of foo(X.A, X.B) :- g o a l l ( X ) . foo(X.A, Y.A) :- g o a l 2 ( X , Y ) .  would produce switch _ r 0 case X.A: switch _ r l case X.B: alternate goall(X) else g o a l 2 ( X , X ) /* wrong! */ end-alternate end-switch end-switch  instead of 101  switch _rO case X . A : switch _ r l case  _tl.B:  alternate _tl  = X,  goall(_tl)  else goal2(X,  _tl)  end-alternate end-switch end-switch  This could be taken care of by an analysis of "bound" variables when generating the cases, but the analysis is rather complicated. Instead, the compiler relies on the clause compiler guaranteeing that each variable appears in at most one functor. This can result in some unnecessary register moving. In practice,, the unnecessary moves (which are cheap) do not happen very often; they can be removed by a second pass of the peephole optimizer.  The traces are used to turn the predicate into a tree of  alternates  (non-deterministic) and switches (deterministic). This is done repetitively, one layer of the parameter at a time (for example, A . B . C has three layers, so the process could be repeated up to three times). At each stage, the parameter which produces the largest number of partitions is chosen — if there are any delay ("?") notations then only the parameters so marked are considered because all other parameters are assumed to be called with uninstantiated variables.  Once the predicate is transformed entirely into a tree of alternates and it is optimized so that alternates with only one choice (no their single choice. 102  else)  switches,  are replaced by  The final stage transforms the tree of alternates, machine code.  Alternate  mkCh  label1:  and  tests  into xpPAM  is transformed to  label1  pushB  0  pushB  n  ... f i r s t  switches  c h o i c e ...  popBKeep n popBKeep 0 mkCh  label2: labeln:  popB  n  popB  0  ... l a s t  A  label2  ... s e c o n d  switch  c h o i c e ...  c h o i c e ...  is a bit more complex than an alternate because it must provide for  both a deterministic and a nondeterministic case. It uses the swXVNL and instruction. For example switch  _r0  case  []:  case  'a :  case case  clause  1  clause  2  b ' : clause  3  1  1  'c':  clause  4  end-switch  becomes  103  eqskip  swXVNL 0 goto atom /* i t ' s n o t a l i s t element, n i l o r v a r i a b l e */ goto v a r /* i t ' s a v a r i a b l e */ goto n i l /* i t ' s [ ] */ fail /* i t ' s a _._ */ nil: clause 1 /* t h i s w i l l c o n t a i n a " r e t u r n " o r " l a s t c a l l " */ atom: caseGoto 0 c a s e ( f a i l ) /* n o t [ ] , V , V o r 'c' */ case 'a',c2 case 'b',c3 case 'c',c4 case-end c l : clause2 c2: c l a u s e 3 c3: c l a u s e 4  /* each c l a u s e ends w i t h a " r e t u r n " o r " l a s t c a l l " */  /* n o n - d e t e r m i n i s t i c case ... */ var: make c h o i c e a l t l eq 0,[] goto n i l altl: modify c h o i c e a l t 3 eq 0,'b* goto c2 alt2: modify c h o i c e a l t 3 eq 0,'b' goto c3 alt3: remove c h o i c e eq 0,'c' goto c 4 '  where make  choice, modify choice  and  remove choice  are the code sequences as  given for alternate above. An alternative to the above code would be  104  caseGoto 0 case(var) c a s e ( f a i l ) /* n o t [ ] , V , case [ ] , n i l case a',c2 case ' b , c 3 case 'c',c4 case-end  ' b ' or V  */  1  1  The first form would be preferred if "[ ]" were more common; the second form if less common. (Incidentally, the caseGoto could have been synthesized from eqskipS.)  5.6 Shallow backtracking The above compilation of if-then-else will execute slower than similar code on a conventional machine. Better if-then-else code is:  •  Test for uninstantiated variables.  •  Push all active registers onto the backtrack stack.  •  Push machine state and "else" address onto the backtrack stack.  •  Create choice point.  •  Call the test predicate  •  On success: remove the backtrack entries and cut the backtrack stack using rmChCut.  •  On failure: pop the backtrack stack entries.  105  Note that no registers are pushed onto the execution stack: the backtrack stack is used for them. Thus, the only overhead in an if-then-else is in testing for uninstantiated variables and in creating the choice point.  The test for uninstantiated variables can be removed if we know that the test will not instantiate anything (this can be detected by a method similar to that which automatically generates  proceed  annotations). The test predicate itself  can also be transformed to have its own delays.  5.7 Global optimizations Very often a number of predicates are used together as a module. For example, quick-sort  typically has a partition predicate which is used only by  quick-sort.  When the compiler knows exactly how a predicate is called, further optimizations are possible: •  If the helper predicate is not recursive, it can be substituted into the body of the calling predicate (this is especially useful for built-in predicates such as arithmetic and comparison). For example: p(A, q(Xl,  B)  :-  X2)  q(B, X), :-  s(Xl),  r(X,  A).  t(X2).  can be replaced by: p(A,  B)  :-  s(B),  t(X),  r(X,  A).  106  However, some care must be taken. If q has any delays in it, this substitution can not be done because the delay would take effect for predicate p instead of just q.  •  The assignment of registers to the helper predicates can be changed to be more efficient for the calling predicate. For example, in: p(A,  B, C, D) :- q ( a , A, B, C, D).  the arguments to q can be passed in registers 4, 0, 1, 2, 3 so that there is no need to move A from register 0 to register 1, B from register 1 to register 2, etc.  •  Delays can be propagated outwards. For example, in: p(X, Y, Z) :- X <  Y, q ( Z ) .  p(X, Y, Z) :- X >= Y, r ( Z ) .  Delay tests could be done at the beginning of p (indicated by ?-proceed p(?>?>-)) so that a special non-delaying version of the inequality tests could be used. In some cases, this also allows removing tests for uninstantiated variables from  if-then-elses.  However, care must be taken so that possible  solutions are not eliminated - in this example, the optimization could only be done if predicates q and r do not cause their argument to become instantiated. •  Some cases of garbage collection can be done explicitly. See [Kluzniak 1987] and [Bruynooghe 1986].  107  5.8  Compiling a predicate with delays  Predicates with delays only slightly complicate the compiler. As mentioned above, the delay annotations are used to restrict which parameters are used for computing traces. The delay annotations can be carried through to the final code generation stage. At this point, they can occur in the following situations:  •  As part of an delay  or c a s e G o t o instruction.  instructions are added. For example  labell: label2:  •  eq, e q s k i p , e q l s t  nonvarGoto vO,  label1  delay  vO,  labell  nonvarGoto v l ,  label2  delay  vl,  label2  eqlst  fO, n l ,  As part of a  swXVNL  eqlst  NonvarGoto  v0?,vl?,n2  instruction. If it is the first operand, the uninstantiated f0?,ni,n2  become swXVNL goto  atom  goto  var  fO, n l ,  n2  goto n i l goto 1st var:  delay  vO,  switch  Or-delays make things a bit more complicated. Consider ?-  proceed  append(a?,  ?-  proceed  append(a,  append([], append(X.A,  X, B,  b, b,  c). c?).  X). X.C)  would become  n2  variable situation is already detected. Thus, swXVNL  switch:  and  :-  append(A,  B,  C).  108  append:  would  If only the first proceed declaration were given, this would be s w i t c h _r0? case [ ] : _ r l = _r2 case X.A: _ r 2 = X.C, append(A, _ r l , C) end-switch  However, if the first argument is uninstantiated, execution may still proceed if the third argument is instantiated. To show this, switch may have a var case: switch _r0 case [ ] : _rl = _ r l case X.A: _r2 = X.C, append(A, _ r l , C) var: a l t e r n a t e _r2? : d e l a y - o r _ r 0 , _ r 2 _r0 = [ ] , _ r l = _r2 else _r0 = X.A, _ r 2 = X.C, append(A, _ r l , C) end-alternate end-switch  The d e l a y - o r means that if none of the designated values are uninstantiated, delayOrs  must be generated. The code is  /* r e g i s t e r 2 i s a l r e a d y known t o p o i n t a t an u n i n s t a n t i a t e d v a r i a b l e . */ nonvarGoto v2, l a b e l l delayOr vO, 0 delay v2, 0 l a b e l l : ...  109  This causes append to delay until either what is in register 0 or what is in register 2 becomes instantiated. Execution will re-start at the beginning (offset 0)  because when append is resumed, there is no way of knowing which variable  became instantiated and caused append to resume.  5.9 The  Optimized compiling to conventional machine code  xpPAM  code can be interpreted or it can be translated to conventional  machine code. The simplest translation would simply use the separate cases in the interpreter as code templates, removing the interpreter loop. This typically gives an execution speed-up of 2 —3 times at the cost of increasing the code size by a factor of 10 or more. Significant execution speed-ups are possible by using machine code (assembler) instead of a higher level language such as C or Pascal. This is because assembler provides more possibilities of exploiting the machine's registers. The machine registers can be used for: •  pointer to top of execution stack.  •  pointer to next free object.  •  the first n registers of  xpPAM.  This can easily double the speed of the generated code. Note that registers are not used to hold information that is used only on backtracking; x p P A M is optimized for deterministic computations.  110  Assigning the xpPAM registers to machine registers has an additional advantage. When xpPAM is interpreted, the operands must be used as indexes into the vector of "registers." Machine code references these directly, even if they have to be kept in slower main storage because of insufficient hardware registers. To give a feel for the speed-ups possible, I changed the implementation to store an offset into the register vector instead of an index. This avoided some shift instructions, giving about 10% speed-up on an MC68000.  Some other optimizations are possible when compiling to machine code. These are typical of optimizers for conventional languages, such as "peephole" optimizations for removing redundant load/stores. Some additional operand annotation, such as r to indicate that de-referencing is not needed, can help optimize the code. Careful tuning of the machine code can produce dramatic speed-ups. In an experiment, the following were observed for naive reverse: 1.4 KLIPS with a simple clause interpreter. 5 KLIPS with a simple abstract machine interpreter. 10-20 KLIPS with more sophisticated abstract machine interpreters. 20-40 KLIPS with translation to C. over 100 KLIPS with translation to native code. The generated machine code must be position independent because it resides in the xArea extension area and may be moved around by the compactor. Some machines have separate instruction and data spaces; for these, separate  111  xAreas  would have to be provided (these areas could still require compaction, because of incremental re-compilation). The design of special purpose computers for logic programming is discussed elsewhere, for example [Tick and Warren 1984] and [Mills 1986]. Such machines could probably directly execute x p P A M code significantly faster than could conventional machines, using the same technology. Major speed-ups are:  •  Parallel execution of parts of instructions (e.g., tag decoding).  •  Maintaining shadow registers and caches which are better oriented to  xpPAM's  referencing style.  5.10  Speeding up deterministic predicates -  modes and types  A type can be considered to be a predicate which is evaluated at compile time and whose failure causes a compilation error rather than a run-time error (or, worse, an unexpected failure in a predicate which should always succeed). Such a predicate either checks an already instantiated variable or imposes a constraint on an uninstantiated variable. This can be implemented in one of three ways: •  The type predicate simply delays until the argument becomes sufficiently instantiated.  •  A type flag is set for the variable so that unification will fail if an attempt is made to unify the variable with something of the wrong type.  •  Compile-time analysis can propagate the type information outward and eliminate tests. For example, in the following, type propagation could 112  discover that both the argument for q and Hd in p must be integers, so no test is needed if p is called with a list of integers: P(U). p(Hd.Tl) q(X)  :-  :-  q(Hd),  p(Tl).  integer(X),  ...  Typing predicates are useful debugging tools. Instead of just failing with the wrong type argument, they cause a run-time or compile-time error. Typing predicates can be extended to work for more complex types. For example, lists can be handled by:  17  /*  general  listen). list(Hd.Tl)  /*  list  of  list(Type, list(Type,  lists:  :-  */  list(Tl).  a specific  type:  */  []). Hd.Tl)  :-  Type(Hd),  list(Type,  Tl).  These predicates are included in the program just like ordinary predicates. The compiler can be told which predicates are for typing - they are executed at compile time and used for type inferencing. The basic built-in types include integer, number and  1  7  string.  Note the use of x p P r o l o g ' s notation in Type(Hd). Test  =..  [Type,  (where c a l l ( T e s t )  Hd],  call(Test)  could be replaced by T e s t ) . 113  In most Prologs, this would be written:  As an example, here is intMember which is like  member  but which only works  on lists of integers: ?- mode intMember(X, L i s t ) :- i n t e g e r ( X ) , intMember(X, X . _ ) . intMember(X, _ . R e s t ) :- i n t M e m b e r ( R e s t ) .  listOf(integer, List).  This can be considered as an abbreviation of intMember(X, L i s t ) :- i n t e g e r ( X ) , l i s t O f ( i n t e g e r , L i s t ) , intMember2(X, L i s t ) . intMember2(X, X . R e s t ) .  intMember2(X, Y.Rest) :- intMember(X,  Rest).  The optimizer could transform this to intMember(X, L i s t ) :- i n t e g e r ( X ) , intMember2(X, L i s t ) . intMember2(X, Y.Rest) :- /* i n t e g e r ( Y ) , */ X=Y, l i s t O f ( i n t e g e r , R e s t ) . intMember2(X, Y.Rest) :- /* i n t e g e r ( Y ) , */ intMember2(X, R e s t ) .  Here, using type predicates is less efficient than leaving them out. Inferring the types of many predicates is not very difficult - the algorithm is similar to the algorithm for detecting where delays should be added to code (as given in [Naish 1985b]). Thus, the mode of member can be inferred as ?- mode member(X,List) :-  list(List).  There is no way to infer that it should be ?- mode member(X,List) :- t y p e O f ( X , T y p e ) , l i s t O f ( T y p e , L i s t )  In fact, this would be over-specifying the type because it would disallow the query member(l, [ l , a ] ) . Nevertheless, there may be advantages in specifying the stronger type declaration of member because more efficient machine code might be generated if a more complex abstract machine model were used which allowed both tagged and untagged objects - in xpPAM, when testing for equality, 114  the tags are first compared, then the values; comparing only the values would be more efficient.  Increases in efficiency could also be produced by input-output mode declarations and inferences. These are in many ways similar to type predicates. Significant speed-ups are possible, as shown by the example in section 2.0, "Fast append on a conventional machine" on page 15.  Besides increasing efficiency, mode and type declarations and inferences help to show the correctness of x p P r o l o g programs. Conventional Prolog is much like the C language where type mismatches produce wrong results (Prolog produces unexpected failures); x p P r o l o g could be more like Pascal where such errors are detected at compile time.  115  6.0 Compiling a functional language to the logic engine  ... the Form remains, the Function  — William Wordsworth,  never dies.  The River  Duddon  XpPAM can be easily adapted as an efficient target for a functional programming language. The delay mechanism, which allows coroutining and sound negation, is sufficient for all common functional programming constructs such as lazy evaluation and handling functions asfirstclass objects. The resulting machine is about as efficient as a machine designed solely for functional programming.  6.1  Introduction  Much has been written about combining functional programming and logic programming. I will only briefly discuss how the two paradigms of logic programming and functional programming should be combined; the reader should see [DeGroot and Lindstrom 1986] which also has extensive bibliographies. Instead, I will concentrate on how xpPAM can efficiently implement functional programming constructs.'  The proposals for combining the two paradigms have been roughly classified in [Bosco and Giovanetti 1986]: 116  functional plus logic: invertibility, nondeterminism and unification are added to a higher order functional language logic plus functional:first-orderfunctions are added to afirst-orderlogic with an equality theory Which is best is a matter of debate. Since a logic machine already deals with unification, nondeterminism and invertibility, I have taken the logic plus functional approach. XpPAM can be used to answer one of the open problems cited by [Bosco and Giovanetti 1986]: the kind of computational mechanism necessary to execute functional and logical programming constructs.  6.2  Functional Programming background  For illustrative purposes, I will use some of the syntax of SASL and HASL [Turner, 1979] [Abramson 1984]. In this discussion, I will assume functional programming with no side affects. That is, when I use the term "LISP," I really mean "pure LISP," not using features such as RPLACD. The m clauses of an n-ary function /are defined: / «n f  a  fli2  a  21  22  •••  )n = P i  a  ex  ... a  2n  f m\ m2 — mn a  a  a  =  =  ex  r  expr  2  Pm r  117  Expressions may be written as: expr where { definitions  }  The definitions may include function definitions, as well as definitions of the form: /* [1, 1, 1, ...] - ":" is the "cons" operator */  x=l:x  [fl,M = [1,2,3] etc. A definition such as: sf g x = / x (g x) can be thought of as a "syntactically sugared" version of the lambda expression: 5 = Xf. Xg. Xx.f x(gx) .  Some functional programming languages have only single argument functions. The single argument and single result, however, may themselves be single argument functions. That is, if/is defined as an n-ary function, it is implemented in such a way that in an evaluation of /  argi arg  ... arg  2  n  the implicit evaluation is (...(/ arg ) arg ) 2  t  ... arg„)  where the result of {f  arg ) x  is a function of a single argument, which can then be  applied to arg , etc. SASL and HASL treat /?-ary functions this way; LISP and 2  Scheme do not.  118  6.2.1  Normal and applicative order evaluation  Normal order applies the leftmost reduction (evaluation) repeatedly until the normal form (or sometimes just head normal form) of the expression is produced. This means that an argument will never be evaluated if it is not needed within a function body. For example, if JC+1 is passed as an argument, it will not be evaluated until it occurs in the context of an operator which requires its value, such as an equality test. On Turner's combinator machine, the J C + 1 is replaced by its value so that any subsequent use of the argument obtains the value directly.  Applicative order evaluates the arguments before evaluating the function body. This is the usual way with LISP and Scheme, although they may delay evaluation by having the caller package arguments in lambda expressions (and, in the case of LISP, the called function must be prepared for such arguments). Both applicative and normal order evaluation produce the same answers if they terminate. However, it is possible that applicative order may do unnecessary computations which may go into an infinite loop. A compromise between the two is "evaluation by need" where arguments are not evaluated until their values are needed - that is, a conventional machine architecture is used but argument evaluation is delayed as late as possible. In evaluation by need, the value of an argument is saved so that it is not re-evaluated if it is used later in the function.  119  6.2.2  Lazy and eager evaluation.  Lazy evaluation delays evaluations until they are needed, whereas eager evaluation proceeds even though values might not be needed. One effect of lazy evaluation is to permit programming with "infinite" structures such as: x =  1 : x  (an infinite list of l's: [1, 1, 1, ...]). With lazy evaluation, an expression such as "hd x" yields the value 1, the tail of the list not being evaluated. SASL and HASL provide lazy evaluation; LISP, generally, does not, but permits the construction of mechanisms for lazy evaluation. 6.2.3  Lexical and dynamic scoping (deep and shallow binding)  In a language where functions are first class objects, a function must be applied in the correct environment, binding free variables in the function body. The data structure consisting of the function body and an environment is called a closure or a thunk [Ingerman 1961]. In pure functional programming languages, the closure is formed at definition time: this is called lexical scoping or deep binding. Other less pure functional languages form a closure at evaluation time: this is known as dynamic scoping or shallow binding. Pure functional programming requires lexical scoping and is used in such languages as SASL, HASL and Scheme. Many varieties of LISP, however, use dynamic scoping which leads to serious semantic problems, chiefly a kind of destructive assignment.  120  6.2.4  Mechanical evaluation of functional programming constructs  The earliest programming language with a strongly functional flavour was McCarthy's LISP. An evaluator (interpreter) for a functional subset of LISP was defined in LISP. LISP, however, was not purely functional, allowing destructive assignment, go tos, etc. Several years after LISP was introduced, [Landin 1966] described an interpreter for Church's lambda notation. The SECD machine adopts the strategy of applying functions to evaluated arguments and is suitable for programming languages using applicative order. Some lambda expressions, however, can be evaluated only with normal order evaluation; applicative order would result in non-terminating computations.  The "procrastinating" SECD machine, a modification of the original SECD machine, allows the postponement of evaluation of arguments until their values become necessary. In a purely functional setting, this is equivalent to normal order evaluation. The evaluation of an expression can be carried out once, and this value saved if the value of the expression is needed again - the above-mentioned evaluation by need (see [Burge 1975]). Another major technique for evaluating functional programming constructs was introduced by Turner in an implementation of SASL. The primitive operations of SASL are application of a function to a single argument, and the pairing or construction of lists. In this technique, variables are removed from lambda expressions yielding expressions containing only combinators and global names referring either to library or user-defined functions (from which variables of course have been removed). Evaluation of an expression then is accomplished by a combinator machine: a machine whose instructions correspond to the three 121  fundamental combinators S, K and / and additional combinators which are not strictly necessary but are introduced to reduce the size of the generated combinator machine code. The leftmost possible reduction is performed to yield an expression's head normal form. Normal order evaluation and "lazy" evaluation of lists fall out from this and the definition of the combinator expressions. Furthermore, SASL introduced a limited kind of unification, extended later in Abramson's HASL. 6.2.5  Functional programming via logic programming.  Functional programming languages may be evaluated by logic programs either by interpretation of functional expressions or by compilation of functional programs to logic program goals to be solved. The former technique was used by [Abramson 1984] to specify HASL (an extension of Turner's SASL) in Prolog; the latter technique was used by [Bosco and Giovanetti 1986] to show how various functional programming constructs could be represented as Prolog goals to be solved. Other aspects of the implementation of functional programming extensions of logic programming may be found in [DeGroot and Lindstrom 1986].  Since functional programming constructs can be compiled to Prolog, and since Prolog itself may be compiled, it is possible to stop at this stage and consider the problem of implementing functional programming constructs in a logic programming language as being solved. However, more efficient handling of functional programming constructs is still possible if one compiles them not into Prolog but directly into code for a logic engine. In the remainder of this section we discuss this, in the process gaining about the same efficiency as would be 122  possible by compiling functional programming constructs to an abstract machine designed specifically for functional program evaluations. The xpPAM's design is a better target for a functional language than is the Warren Abstract Machine (WAM). The principal differences between xpPAM and WAM  are:  •  Registers contain only pointers to objects, not the objects themselves.  •  All objects are kept on the heap (like WAM's "global" stack, but not popped on backtracking).  •  The unification instructions are simpler (for example, there is no distinction between "local" and "global" variables).  •  The execution stack (WAM's "local" stack) contains only call/return information.  •  There is no "environment" for a predicate - push and pop instructions are used to save values across calls.  6.3 Sample code Here is how a function p which produces a new list by applying the function q to each element of a list may be written in SASL or HASL:  PU = U p (Hd:Tl)  = (q Hd)  : (p  Tl)  123  In Prolog, this is: P([],  H) :-  p ( H d . T l , HdX.TlX) :q(Hd, HdX), !, p ( T l , T1X).  The  cuts  ("!") are necessary to make this deterministic (xpProlog's  if-then-else  could be used to eliminate these). In general, anrc-aryfunction can be turned into an (rc+l)-ary predicate by adding one parameter to hold the result. This extra parameter must always be initialized as an uninstantiated logical variable before a call and the called predicate must always instantiate it before returning.  Here is xpPAM code (comparable WAM  code is in section 7.1, "Comparison with  the Warren Abstract Machine instructions" on page 138): swXVNL builtIn builtin goto it: eqlst link push push call  f0,n0,n2 % s w i t c h on parmO IT % i n v a l i d parm e r r o r II II % can't be v a r i a b l e error nil % parmO=[] % parmO=Hd.Tl f l , n l , n 3 % parml:=HdX.TlX ( c a n ' t tt  f2 f3 q/2  nl pop nO pop unlink ICallSelf nil: eq return  fl,U  % % % %  save T l save T1X q(Hd,HdX): regs 0, 1 are a l r e a d y s e t a r g l : = r e s t o r e T1X % argO:=restore T l P(TI.TIX)  % result:=[]  124  fail)  The I C a l l S e l f instruction has the same meaning as goto o (the different opcode helps in debugging). This is a tail recursive call - recursion has been turned into iteration.  For comparison, here is the pure function version of the function /; which applies q to each element of a list. The result is returned in register 1 which is not pre-initialized to be a logical variable (actually, register 0 should be used to make higher order functions easier to implement - doing this would require more machine instructions for moving values into the correct registers). The link  and u n l i n k instructions are not needed for purely deterministic  computations (if used with a slightly different call instruction). swXVNL builtin builtin goto 1st: link push call pop unlink link push call pop unlink eqlst return nil: eq return  f0,n0,n2 "error" "error" nil  % % % % %  f2  % save T l % argO: a l r e a d y s e t % regl:=q(Hd) % restore T l  q/i  nO  fl P/l n2  s w i t c h on parmO i n v a l i d parm can#t be v a r i a b l e parmO=[] parmO=Hd.Tl  % save q(Hd) % rl:=p(Tl) % r e s t o r e q(Hd)  n 0 , f 2 , f l Jo r e s u l t : = q ( H d ) . p ( T l ) ( c a n ' t  nl, []  Jo r e s u l t : = [ ]  125  fail)  This code is one instruction longer than the code which uses logical variables (because of the return instruction). Also, the functional code is not tail recursive because of the "cons," so much more stack is needed. The functional version builds the return value more efficiently using version does it by e q l s t  fl,nl,n3  eqlst  n0,f2,fl  -  the predicate  and then filling in the head and tail. Thus,  the tail recursion optimization in the predicate has a price: slightly slower construction of the result and one extra level of indirection (using "reference" objects). It also violates the xpPAM convention that all registers are empty when a  return  is executed because it leaves one register containing a value.  Typically, using predicates instead of functions results in about the same number of machine instructions. The slight amount of extra execution time required by using logical variables is more than compensated by the better opportunities for detecting tail recursion optimization (TRO). Without loss of generality, we can use deterministic predicates instead of functions. When we say that a predicate returns a value we mean that the last parameter gets instantiated to the value returned by the equivalent function.  So far, we have treated functions as if they were compiled into Prolog. This can introduce large inefficiencies because of the more general nature of unification and the possibility of backtracking. The e q s k i p instruction is used to avoid creating choice points. For example: />• = [] p ('a' : rst) = 'x' : (p rst) p ('b' : rst)  = '/ : (p rst)  is compiled to:  126  f0,n2,n0 swXVNL it it builtin error builtin error goto nil 1st: f2,V eqskip else goto f1,'x',nl eqlst ICallSelf else: f2,'b' eqskip II II builtin error eqlst fl,'y'.nl ICallSelf nil: eq fl,[] return  % s w i t c h on parmO Jo i n v a l i d  parm % caa^t be v a r i a b l e Jo parmO=[] Jo parmO=hd.rst Jo t e s t hd = 'a' Jo r e s u l t  :=  Jo  x : p(rst)  Jo t e s t hd =  V Jo e l s e : i n v a l i d parm 1,  I I  % r e s u l t := y : p(rst) % % r e s u l t := [ ]  A boolean function returns either true or false, so calling a boolean function and testing the result can be done in a similar fashion — there is no need to create choice points.  6.4  Thunks, lazy evaluation and higher order functions  A thunk contains the address of an arguments  (m<.n).  n-ary  predicate and the values of thefirstm  The thunk can be considered as an (n—  m)-ary  predicate.  When the thunk is called, thefirstn — m registers are moved into registers m through n- 1 and registers 0 through (n-m) — 1 are loaded with the values in the thunk. In applicative order, the code for p l u s ( l , 2 , z ) is  127  eq nO,'l' eq nl,'2' link push n2 call 'plus/3' pop nZ unlink  %Z %Z  In normal order, the code is transformed to compute  z where { *=P1(2). pi =pluss( 1), pluss = Ax Ay {x+y } }  Plus 1 2 is  reduced to  pi 2  (with pl(x)  = Ax{x+1})  On xpPAM, we compile the functional expression z=pl(2)  where pi =pluss( 1)  to the predicate calls ?- p l u s s ( l , P l ) ,  P1(2,Z).  which then becomes:  128  and finally to  3.  eq fO.'l link push n l call 'pluss/2' pop n2 unlink 1  eq n0,'2' link push n l callThunk ' f 2 / 2 ' pop nZ unlink  % argument % PI % PI % argument % Z % call % Z  P1(2,Z)  P l u s s is a 1-ary function (2-ary predicate): p l u s s ( X , Z ) :- t h u n k ( p l u s ( X ) / 2 , Z ) .  which is compiled to: mkThunk f l , p l u s / 2 , 1 return  % Z:=thunk(...)  MkThunk f l , p l u s / 2 , 1 means that register 1 is unified to a thunk pointing to plus/2, the first argument already having been set in register 0.  MkThunk is like  a c a l l , so it frees all the arguments (here, register 0 is marked as empty after its value is saved in the thunk). The third operand to mkThunk may be a register so that we can make a thunk from a thunk. MkThunk turns atomic objects into thunks by using the "=" predicate (defined "X=X").  Whenever an instruction has a thunk as an argument and it requires the value, the thunk is "woken up" and evaluated. The non-empty registers are put into a new thunk (pointing to the current instruction) which is pushed onto the execution stack. This suspends the currently executing predicate (unification is 129  repeatable, so we do not need to store any other information to aid in restarting the suspended instruction). The registers are then loaded from the woken thunk and execution proceeds within it until its last  return  instruction is reached.  Normally, when a return is executed, the top element on the stack is a code segment but in this case, it is a thunk so the registers are restored from the thunk (recall that all the registers must be empty before a return) and execution resumes where it was earlier suspended. This is very similar to how coroutining predicates are suspended and resumed.  When a thunk is evaluated, its result may be another thunk. Therefore, equality instructions may repeatedly wake up thunks until they finally get a value. This is handled automatically because when a thunk returns, it restarts the equality instruction (no state information about the instruction needs to be saved).  Because a thunk saves the registers, pure normal order execution does not need to push and pop registers on the execution stack. But the mkThunk and callThunk  instructions are quite expensive, so using thunks and normal order  evaluation is less efficient than applicative order evaluation.  A simple example of delayed evaluation is:  in tsFrom mi  =  m : (int s From (m+i)  i)  which returns a list of everyrthinteger starting at m. We can sum all the even integers up to n by: 130  sumLim  (x : r) n = if x > n then 0 else x + (sumLim  sumEven  n = sumLim  At each step of sumLim, intsFrom  (intsFrom  r n)  2 2) n  a new list element (x:r) is required. This wakes up the  thunk which makes a list element containing the next number and a  thunk for generating the next list element. Eventually sumLim reaches the limit n and no more elements are needed.  A naive translation to producer-consumer coroutines is: intsFrom(M, I , M.L) :- M2 i s M + I , intsFrom(M2, I , L ) . sumLim(X?.R, N, 0) :- X > N. sumLim(X?.R, N, S) :- sumLim(R, N, S 2 ) , S i s X + S. sumEven(N, S) :- sutnLim(L, N, S ) , i n t s F r o m ( 2 , 2, L ) .  where the "?" notations mark where the predicate must delay until the value is instantiated. A delay is implemented by using a swXVNL or v a r G o t o instruction to detect that a value is uninstantiated - a d e l a y instruction suspends the predicate by saving a thunk (with all the non-empty registers) on the delay list associated with the variable and then executing a return. When the variable becomes instantiated, all associated delayed predicates are made eligible for resumption (the current predicate is suspended and all the delayed predicates are pushed onto the execution stack except for the oldest which is resumed).  131  6.4.1  Lazy thunks vs. delayed predicates  The above program has a subtle error. SumLim is started and immediately delays on L. i n t s F r o m is then entered - it instantiates the first element of L. This wakens sumLim (and suspends intsFrom) which calls itself and then delays on the next element of L. Eventually, sumLim terminates but i n t s F r o m continues generating elements in the list even though they are not needed. Here is a correct version which uses i n t s B e t w e e n to generate a finite list: intsBetween(M, N, I , M.L) :- M < N, M2 i s M + I , intsBetween(M2, N, I , L ) . intsBetween(M, N, I , [ ] ) : - M > N. sum([]?, 0 ) . sum(X?.R, S) :- sum(R S 2 ) , S i s X + S. sumEven(N, S) :- sum(L, S ) , i n t s B e t w e e n ( 2 , N, 2, L ) .  This should not be taken to mean that functionalthunks are more powerful than delayed predicates. Sometimes, delayed predicates are easier to use because.they allow more than one predicate to delay on a single variable. Predicates also can backtrack. The main difference between the two concepts is in how they handle an "infinite" list. For the computation to terminate, the list must be made finite. Thunks do this by eventually leaving the tail uncomputed; delayed predicates eventually instantiate the tail to nil. In both cases, the list need not actually exist (it is, after all, just a communication channel) - reference counting (if used) ensures that only the current element exists, all other elements being deallocated as soon as they are finished with. 132  6.5  Equality: "is" and " = "  Standard Prolog has a simple syntactic equality theory given by the predicate " = " (defined "x=x"). Another kind of equality is provided by the built-in predicate " i s " (":=" in Waterloo and IBM Prologs). This can be considered to be defined: X i s Y :- a t o m i c ( X ) , X=Y. X i s F :- F =.. Fname.FArgs, i s A r g s ( F A r g s , F A x ) , /* e v a l u a t e t h e args */ append(Fname.FAx,[X],FL2), F2  =..  FL2,  call(F2). isArgs([], [ ] ) . : - H 2 i s H, isArgs(T, T 2 ) . +(X,Y,Z) :- {Z:=X+Y}. % b u i l t - i n p r e d i c a t e -(X,Y,Z) :- {Z:=X-Y}. isArgs(H.T,  H2.T2)  The predicates " i s , " "+,"  etc. delay if any of their parameters are not  sufficiently instantiated.  Thefirstclause is a slight extension of the conventional definition (removing the "only numbers" restriction). The second clause expects the right-hand parameter to be of the form F{A A ,... , A„) : it evaluates arguments A through U  2  x  A„ (recursively using " i s " ) to produce 5 , through B„, then computes X by calling F{Bi, B ,..., B„) . If we assume that " c a l l , " "=.." "+" etc. are defined by an 2  infinite number of rules, " i s " is definable in first-order logic.  133  This definition of " i s " gives a kind of semantic equality. The more convenient notation p ( { F j ) means F v i s F, p ( F v ) ({F} is pronounced "evaluate F"). Using this for factorial: f(o, 1 ) . f ( N , (N * f ( { N - l } ) } .  which is an abbreviation for: f(o, l ) . f ( N , NF) :- NF i s N*F, Nsub i s N - l , f(Nsub, F ) .  This definition depends on arithmetic predicates delaying when their arguments are insufficiently instantiated. Although tail recursive, it is much less efficient than the non-tail recursive version because of the overhead of processing the delays.  6.6 Combirtators The implementation described so far corresponds to a lazy SECD machine [Henderson 1980]. But it can also be used for a combinator machine [Turner 1979]. Thunks and code segments, being "first class" objects, can be passed as arguments and be returned as values. All the equality and call opcodes can take thunks as operands, evaluating them when needed as described above. A thunk is evaluated only when required. When a thunk is evaluated, it may return a structure containing another thunk which itself requires evaluation when its value is needed. 134  The traditional combinators can be restated as predicates. Given the definitions: I = Xx.x K = /ix ^y.x S = Af Ag ix.(.f  x ) ( g x)  we get the following predicate definitions (using  thunk/2  as described earlier):  comb_I(F,F) . comb_K(F,Z) :- thunk(comb_kk(F)/2, Z) . comb_S(F,Z) :- t h u n k ( c o m b _ s s ( F ) / 3 , Z ) .  with the auxiliary predicates: comb_kk(X,Y,X). comb_ss(F,G,Z) comb_sss(F,G,X)  :- thunk(comb_sss(F,G)/4,Z). :- F ( X , Z 1 ) , G(X,Z2), Z1(Z2,Z).  Here is a computation using pluss (1-ary addition) and succ (successor). a = 5 plus succ 3  is compiled to the predicate definitions and calls p l u s s ( X , Z ) :- t h u n k ( p l u s ( X ) / 2 , Z ) . s u c c ( X , Z ) :- Z i s X + l .  ?- c o m b _ s ( p l u s s , A l ) , A l ( s u c c , A 2 ) , A2(3,A).  resulting in: Al  = thunk comb_ss {parmO= pluss]  A2 =  thunk combjsss  {parmO = pluss, parml —succ}  A = Z where ( p l u s s ( 3 , Z l ) , s u c c ( 3 , Z 2 ) , Z1(Z2,Z)}  leading to: 135  Zl = Z2  thunk plus {parm0 = 3}  = 4  and finally (when evaluation of Z is forced all the way): z = A = 7 When code is written using the combinators S, K and /, the predicate calls to pred_S, pred_K  and p r e d _ l look, just like normal predicate calls. When they are  executed, their returned values are thunks which can be handled like any other objects. They will be evaluated only when needed and only as much as needed, possibly returning structures containing other thunks.  This definition of combinators requires that thunks be created only for the basic combinators S, K and / and of course for any other combinators (B, C, etc.) which are introduced to control the size of the compiled code (from which variables have been removed). All other definitions are done without reference to thunks.  The combinator machine uses a subset of the inference machine's instructions in a small number of ways, so some optimizations are possible. Call instructions invariably have two arguments (single argument and result), so the "load arguments, push new variable for result, call, pop result" sequence could be optimized into one instruction. By adding these optimizations to the machine, the programmer is given the flexibility of efficiently using either the lambda machine or the combinator machine models of functional computation - or mixing them — as the problem requires.  136  In Turner's combinator machine, combinator reduction occurs in-place. Turner observed that when an expression occurs within a function, it will only be evaluated once, thefirsttime it is needed. Because logical variables share, we get the same effect with our predicate translations of combinators. In the combinator machine, values do not actually get replaced — rather, new values are computed and the old values are abandoned (and eventually garbage collected).  137  7.0 Design decisions  Quieta movere magna merces videbatur. (Just to stir things up seemed a great reward in itself.)  — Sallust, Catiline,  21  Some design decisions are done for a good reason; others are simply arbitrary decisions and some are done just to simplify the implementation. Unfortunately, it is often not obvious to the reader which category a particular decision falls into. The main design decision was to create an alternative abstract machine to the Warren Abstract Machine (WAM). Most implementations of Prolog use  WAM  with a few minor changes, typically adding some new instructions. However, I felt that the WAM  was already too complicated and that it would be further  complicated by adding delaying features.  7.1  Comparison with the Warren Abstract Machine instructions  In Prolog, here is a typical predicate which applies the predicate q to each element of a list:  138  p([],  [])•  p ( H d . T l , HdX.TlX) :q(Hd, HdX), p ( T l , T1X).  Here is xpPAM code: swXVNL  f0,n0,n2  fail goto goto  % none o f t h e 3 below var % parmO i s v a r n i l % parmO=[] Jo c o n t i n u e t o 1 s t case % parmO=Hd.Tl f l . n l , n 3 % partnl=HdX.T1X  1st: eqlst link push push call  f2 f3 q/2  pop nl pop nO unlink ICallSelf nil: eq fl,[] return var: pushB vO pushB vl pushV v2 mkCh else eq fO, [ ] goto n i l else: popB n2 popB nl popB nO eqlst f0,n0,n2 goto 1st  Jo save T l Jo save T1X Jo q(Hd,HdX): regs 0 and Jo 1 a r e a l r e a d y s e t Jo r e s t o r e T1X Jo r e s t o r e T l Jo p ( T l , T l X )  J, make t h e c h o i c e J, p o i n t by s a v i n g J, a l l a c t i v e regs J, parmO=[ ]  Jo r e s t o r e t h e J, a c t i v e regs J, on f a i l u r e % parmO=Hd.Tl  139  And the equivalent WAM  code (courtesy of Saumya Debray):  switch_term v a r , n i l , 1st var: try_me_else else nil: get_nil 1 proceed else: trust_me_else_fail 1st: allocate % c r e a t e an environment. get_list 1 % arg 1, i n r e g 1, i s a l i s t . unify_tvar 1 % Hd i s a temporary, p u t i t i n r e g i s t e r 1; i t ' l l % be needed t h e r e f o r t h e c a l l t o q/2. unify_pvar 2 % T l i s a permanent, save i t a t Jo d i s p l a c e m e n t 2 i n environment. 2 get_list Jo a r g 2, i n r e g 2, i s a l i s t . unify_tvar 2 Jo same comment as f o r Hd. unify_pvar 3 Jo T1X i s a permanent, save i t Jo at d i s p l a c e m e n t 3 i n environment. call Jo n o t i c e args a r e i n p r o p e r p o s i t i o n s q/2 put_pval 2, 1 Jo move T l i n t o r e g i s t e r 1 3, 2 Jo move T1X i n t o r e g i s t e r 2 put_pval deallocate Jo get r i d o f environment execute p/2 Jo l a s t c a l l  Even though the WAM  instructions are more complex than mine, more of them  are required in the inner loop (10 vs. 13); for deterministic append, the difference is even more dramatic: xpPAM has just 3 instructions in the inner loop against 8 for WAM  (these can be optimized to 2 and 7, respectively). Both have  about the same number of memory and register references (in addition to call frame allocation and deallocation, xpPAM has 2 references to the heap and 4 to the execution stack; WAM  has 9 memory references [Tick86]). It is difficult to  draw any general conclusions from this example; WAM appear to have similar efficiency. 140  and xpPAM design  7.2  Environments on the execution stack  Warren observed that passing arguments in registers rather than in stack frames has two advantages: tail recursion optimization can be easily performed and stack frames do not need to be created for unit clauses. However, his design keeps local variables in the stack. I have chosen to keep pointers to local variables in registers and to copy them to the execution stack only when they must be preserved across calls.  At first glance, my design appears less efficient. Although there are certainly cases where one or the other design is better, I believe that for "typical" programs [Matsumoto 1985], the two designs are similar in efficiency. In practice, only a few registers need to be saved around a call. In WAM  code  these registers would have to be loaded from the local or global stack anyway, so the number of executed instructions and stack references are about the same. The push/pop around a call in my design does not only save values over predicate calls; it also puts values into the correct registers. This simplifies compiler design because register allocation need only consider where the registers are needed between two adjacent calls - the compiler (exclusive of code for handling delays and for detecting deterministic predicates) is about 600 lines of Prolog (which took a week to write and debug). The compiler seldom has to move the contents of one register to another; in WAM,  instructions like  put_pval  appear quite often.  On a conventional machine, there is an interesting advantage to extending the heap or stack one cell at a time. If the next address can be made invalid, an 141  addressing exception will happen when there is no more room on the heap or stack. However, an environment requires extending the stack a number of cells at a time and the environment cells must be contiguous, so every time an environment is built, there must be a test for stack overflow. Thus, xpPAM's allocating from a free list can be more efficient than allocating from a stack.  7.3 Global and local stacks WAM has two execution stacks: "local" and "global." In the local stack, pointers are always from newer to older. Any pointers which cannot stick to this discipline are put onto the global stack (everything on the global stack is considered to be older than the local stack).  WAM's local stack is pushed and popped like a conventional execution stack (subject to frames being "protected" by backtrack information). The global stack is popped only on backtracking. However, the global stack can contain unused cells - for deterministic computations, the global stack must be garbage collected (some people use repeat, fail loops to get this effect - see section 8.7, "Cut" on page 173). That is, the WAM global stack is not really a stack - it is a global heap which can automatically remove some entries only on backtracking. WAM's use of the global stack allows for very good speedfiguresfor queries such as deterministic append (the "standard" LIPS figures use na'ive reverse, whose inner loop is deterministic append). WAM does not need to garbage collect the resulting list because the stacks will be popped after the query. Because XpPAM allocates everything in a heap, it suffers a speed penalty for such 142  queries. It is not clear how much better WAM  is than XpPAM for more "typical"  programs - if garbage collection is needed, both should be about the same. If there are delayed predicates, XpPAM is superior because there is no need to "globalize" variables.  XpPAM's design is intended for maximum speed for deterministic predicates. "Shallow backtracking" is a special case, which is handled by if-then-elses.  Full  backtracking is considered as an unusual situation - if some backtracking efficiency must be given up to gain deterministic efficiency, I consider that to be a good trade. I therefore see no advantage in WAM's ability to chop the global stack on backtracking.  Whenever two uninstantiated variables are unified, the newer points to the older. In WAM, age comparison is done by comparing stack addresses - the global stack is always older than the local stack. Therefore, there are never any pointers from the global stack to the local stack. Variables are recorded on the "trail" (backtrack stack) when they are older than the current choice point. The net result is similar to xpPAM's which keeps an "age" within each uninstantiated variable.  Compilers for WAM  must do some analysis to determine which variables are  local and which are global. There is no need for this with xpPAM. The local/global analysis also depends on the order of goals. In the presence of delays, this does not work, so a delay must "globalize" all the local variables, giving worse performance than xpPAM.  143  XpPAM's design allows treating predicates asfirstclass objects. This is somewhat trickier in WAM  because of the need to preserve and restore the state of the  current stack frame ("environment"). Also, xpPAM can delay and resume a predicate at any instruction whereas WAM  is more difficult to delay after an  environment (stack frame) has been allocated.  XpPAM does not distinguish between "local" and "global" variables as in  In WAM  WAM.  code, about half to three-quarters of all memory references are to the  global stack (extrapolated from [Tick 1986] and [Matsumoto 1985]). Because xpPAM keeps pointers to local variables entirely in registers, it can get similar  performance without the complication of stack shadow registers (a hardware implementation for xpPAM would probably cache the cells pointed by the registers, instead).  7.4  Saving environments (WAM)  The WAM  creates an environment on the execution stack whenever a predicate  has more than one goal. The environment is used to keep information across goal calls — it corresponds to the registers which are pushed and popped across a call for xpPAM. In addition, WAM  instructions must be able to access either  registers or offsets within the environment. This complicates WAM's unification instruction set, compared to xpPAM's instruction set which always works with registers.  There are two cases when a WAM  environment must be saved: when delaying a  predicate and when creating a "thunk." For a WAM-based design, all local  144  variables would have to be "globalized" first, to guarantee that there are no dangling pointers.  With the xpPAM design, there are no environments and hence nothing special needs to be done when a predicate delays or when a "thunk." is created.  7.5  Allocating from a list or from a stack  Because all the xpPAM object cells are the same size, allocating and deallocating on a linked list is as efficient as allocating and deallocating on a stack. Linked lists are easier to use when storage must be reclaimed (garbage collected). Here is allocation code (in C) for the two methods: Stack  Free list  i f (stack > top) error(); new = * s t a c k ; stack++;  i f ( l i s t - > t a g != f r e e ) errorQ; new = n e x t ; next = n e x t - > t l ;  As mentioned earlier, the ( l i s t - > t a g != free)test can be removed if an addressing exception will occur when there is nothing left on the free list (that is, the last element of the free list points at an invalid address; when it is used, an addressing exception occurs). The test for stack overflow is trickier because an addressing exception will only occur when the stack element is first used, which  145  may be somewhat later. However, hardware can do stack overflow checking in parallel, so the two methods may have the same speed with special hardware  18  Lists have the advantage that additional memory can be allocated as needed. However, a segmented stack architecture is possible - instead of initially allocating one large stack, a smaller stack can be allocated and new stack segments can be added as needed (all segments are chained together). For the WAM,  care must be taken because the "ages" of variables are compared by  comparing their addresses.  7.6  Cut  The original WAM  design did not handle cut. WAM  implementers have  handled this by adding new cut opcodes and doing some source level transformation. For example, p(X)  :-  q(X),  !,  r(X).  is transformed to something like p(X)  :-  p2(X,Z)  cutPoint(Z), :-  q(X),  p2(X,Z).  cutTo(Z),  r(X).  where the cutPoint and cutTo predicates translate directly to the new  cut  opcodes.  1 8  O n a conventional machine, if a page can be marked as invalid (that is, it will cause an addressing exception when read of written), stack overflow can be detected with a single instruction which simply attempts to read from the top end of the frame (this only works if the frame is less than half a page in size, unless multiple pages are marked as invalid).  146  If xpPAM allowed  cut,  such a technique would also work. However, it is not  necessary. "Shallow backtracking" situations allow handled by the  switch, skip  and  goto  if-then-elses  which can be  instructions. In some cases, choice points  must be created. The mkChAt instruction records the cut point information. A later  rmChAt  or  cutAt  does a "soft" or "hard" cut, respectively. Such  instructions could also implement "remote fail" or "deep bail out" situations these are definitely non-logical but are not prevented by xpPAM.  7.7  Code indexing  In deterministic predicates, considerable speed-ups are possible if code indexing can be done. These are similar to switch instructions in C. The WAM  has  special instructions for this which work only on the first parameter.  XpPAM  has generalized the WAM  indexing instructions so that it can be used  with any parameter. As shown by the example in section 5.5, "Optimized compiling of a predicate" on page 98, this allows taking advantage of all possible  if-then-else  situations, thereby minimizing the number of "create choice  point" instructions. The WAM  switch  instruction simply does a branch based on the object tag of  the first parameter. Usually, the next instruction is a list element unify which must check the tag again.  xpPAM  combines these two functions so that its  instruction is a generalization of the  unify list element  instruction. Similar  generalizations are possible for structures and atoms.  147  switch  7.8 RISCs, CISCs and in between Because xpPAM uses only registers for operands, it is a little more RISC-ish than WAM which allows either registers or environment slots for operands. On the other hand, some of xpPAM's instructions are more complex, such as which combines two of WAM's instructions  (switch_on_term  and  swXVNL  unifyJist)  into  one. There is little point in making claims about degree of RISC-ness; what counts is performance and both designs seem to promise similar performance (of course, I think that my design is better).  Both WAM and  xpPAM  may benefit from "reduced instruction set" (RISC)  design (see [Mills 1986] for an example).  XpPAM  has used this philosophy in a  few places. Sometimes simpler instructions are better; in other places, complex instructions are better. •  The separate  link,  push,  call  and  pop  instructions are used to do calls; a  more CISC-ish design (as in the VAX design for conventional computers) would combine these into one or two instructions (say, a call with a list of registers to be saved and restored ). The reasons for this decision are: 19  - It is faster to use the separate instructions, even for an interpreter (the interpreter loop is faster than a /or-loop for indexing over the registers to be saved). - It allows more flexibility in the compiler, so that a register can be pushed as soon as possible, to allow it to be re-used. In some cases, values can be saved on the backtrack stack instead.  19  T h e interpreter loop turned out to be faster than a / w - l o o p over the operands!  148  •  The mkCh;  pushB; pushB; ...  sequence is faster than a make  choice  instruction  which encodes the registers to be pushed onto the stack. An additional advantage is in compiling  if-then-else  code (see section 5.6, "Shallow  backtracking" on page 105) where the backtrack stack is used to save values across a call, saving one set of pushes. •  Earlier version of the implementation had separate add  reference, new free,  etc. instructions and no v, n or/annotation. This complicated the compiler and produced unification instructions that were less efficient. For example, the sequence new 1; eq 1,2 was used to copy register 2 into register 1, requiring a general unification. Tofixthis, copy instructions were created but they were just special cases of the •  xpPAM "eq"  instruction.  Earlier versions had a separate setFail instruction which put a failure address into a status register; whenever a unification instruction failed, the machine jumped to the failure address. The code at the failure address undid all the unifications and then executed a much simpler fail instruction. The advantage of a setFail instruction was that it simplified the  fail  instruction and it cut clown on the number of opcodes (for example, can be handled by a setFail plus  eq).  However,  setFail  eqSkip  was removed because  it resulted in longer instruction sequences and because it greatly complicated the compiler. A possible optimization to xpPAM is to add a variant of the v and/operand annotations to indicate that the register cannot point to a reference cell. This can speed up unification, although I do not know by how much. In WAM, such an optimization could be added by defining new instructions - the speed-up for WAM  is likely to be better than that for xpPAM because it would  149  speed up the  unify list element  instruction which immediately follows a  switch  instruction.  7.9  Instruction format  Both WAM  and xpPAM are deliberately vague on the exact format of  instructions. This is to be expected, because they are  abstract  machine designs.  The instruction layout will depend on the underlying architecture of the implementation. As with most computer designs, there is some trade-off among simplicity, speed and compactness.  Both WAM  and xpPAM have a fairly small number of instructions, supplemented  by "foreign" instructions ( " b u i l t i n " in xpPAM). This makes an interpreter fairly easy to write (a few thousand lines of code, plus memory management). XpPAM also has n, v and /annotation on its operands. Decoding these takes considerable time. Instead, the flags could be combined with the opcodes, resulting in perhaps 300 instructions (many flag combinations are impossible for most instructions). Of course, this makes the interpreter much larger but a doubling of execution speed is possible. Some other tricks are possible for interpreted abstract machine code. Instead of storing register numbers, register offsets (into the register vector) can be used to save shifts; pointers to the registers can produce greater savings, at the cost of larger code size. Similarly, opcodes can be replaced by pointers to the actual interpreter code. As more of these techniques are applied, the result approaches threaded code. Unfortunately, threaded code interpreters are not very machine  150  independent - if a portable implementation is desired, some significant performance degradation must be expected.  7.10  Functor and list storage  For simplicity, my implementation of xpPAM stores functors like lists. The only difference is a flag in the first element which distinguishes the two. The advantage of this technique is that constructs such as  "F(A)=g(x)"  are allowed  (unification is the same as in "[F,A]=[g,x]"). The disadvantage is that more storage is used. If pointers are 4 bytes (12 byte object table entries) and that the machine requires 4-byte alignment of objects in xArea (to handle the back-pointers), the string " f " requires 20 bytes (12 byte object table entry, 4 byte back-pointer, 2 byte string (null-terminated as in C), 2 byte padding). Space for nil is not shown as it is pointed at by many list elements. # args Functor bytes List bytes Functor 32 0 40 f() f(a) 64 64 1 f(a,b) 2 96 88 f(a,b,c) 112 132 3  List [] [f,a] [f,a,b] [f,a,b,c]  As can be seen, there seems to be little advantage of storing functors in  xArea  unless they have many arguments. Data in xArea must be compacted because they are of various sizes whereas data in oTab do not require compacting because they are all of the same size - the somewhat greater storage efficiency of storing functors as records is balanced by the greater increased time spent compacting  xArea  and by the greater complexity of unification.  151  Some additional thoughts on this are given in [Campbell and Hardy 1984]. I believe that x p P r o l o g provides the best of both: notational flexibility and compactness plus simple implementation. However, there is nothing stopping an implementation of xpPAM from using the more traditional record-oriented method for functors. The next section discusses some other aspects of this issue.  7.11  Object cell size  My design is somewhat wasteful of memory because it allows only one object cell size. Some objects (such as nil and reference) have wasted space and other objects (such as strings and code segments) require an extra pointer into the extension area  (xArea).  Furthermore, I do not take advantage of putting objects  within list cells and I do not provide compact list allocation such as encoding. The main reason for my scheme is to simplify implementation and to speed up unification. Suppose that a list cell could contain an uninstantiated variable object (instead of only a pointer to an uninstantiated variable object's cell). The action on unifying with a short integer would be to replace the object; unifying with a long integer or a string would require allocating a new cell and pointing at it. As most uninstantiated variables do eventually become instantiated, the only advantage of the more complex scheme is when dealing with lists which contain mainly short integers - but a lot of logic programming deals with strings which do not present any advantage in the more complex scheme. It should also be noted that the wasted space within an uninstantiated variable cell often gets fully used when the object becomes instantiated.  152  As the cost of increasing the complexity of unifying list elements more complex, cdr-coding could be used for storing list elements. Instead of allocating from a single free list, multiple free lists would be needed for the most common sizes; uncommon sizes would be kept in  xArea  (a similar technique could be used for  strings). Functors would then appear like cdr-coded lists.  7.12  Reference counts and garbage collection  Garbage collection and reference counting are still somewhat contentious issues. My implementation of xpPAM uses reference counting because reference counting frees unused memory as soon as possible. Reference counting also adds confidence to the correctness of the xpPAM implementation - errors in reference counts show up very quickly. As [Knuth 1973] notes: Some of the greatest mysteries in the history of computer program debugging have been caused by the fact that garbage collection suddenly took place at an unexpected time during the running of programs that had worked many times before. A common implementation strategy for Prolog uses local and global stacks. XpProlog  uses registers and the call save stack instead of the local stack and a  heap instead of the global stack. Stack-like operations are as efficient on this heap as they would be on a stack. Using the heap allows freeing value cells which would become "trapped" within a stack [Bruynooghe 1982] and simplifies the implementation of delay which would otherwise require a branching stack. 153  Free  annotation in xpPAM allows freeing objects as soon as they are no longer  needed. These instructions are easily generated automatically at compile time. For example: q s o r t ( U n s o r t e d , S o r t e d ) :- q s o r t x ( U n s o r t e d , S o r t e d , [ ] ) . q s o r t x ( [ ] , Sorted, Sorted). q s o r t x ( H . T , S o r t e d , S o f a r ) :s p l i t ( H , T, Lo, H i ) , qsortx(Lo, Sorted, H.SortHi), qsortx(Hi, SortHi, Sofar).  In the last clause, H and T and the first parameter [H.T) can be freed after split; Lo and H can be freed after the first  qsortx. Qsort  is deterministic if the  first parameter is fully instantiated and split is deterministic.  Reference counting also helps in implementing arrays. For example, the built-in predicate chElem(A,I,V,A2)  takes an input array A, an index /, a new value  V  and replaces that element to produce the new array A2. If A has a reference count of 1,  chElem  can produce  A2  by modifying  A  in place rather than copying  the entire array.  Reference counting and marking garbage collection seem to have about the same overhead [Krasner 1983] (about 15% if care is taken to avoid unneeded incrementing and decrementing the reference counts - this number has been verified for some simple examples with my implementation of xpPAM). Reference counting does not work well with circular lists but Prolog programs will not generate these, at least if they have unification with occurs check [Plaisted 1984]. A marking garbage collector could be used with xpPAM but care would  154  have to be taken to ensure that the marking phase does not cause noticeable pauses - the pauses in  xpPAM when  the  xArea  is compacted are unnoticeable.  Considerable research is still being done on garbage collection. Two interesting possibilities are scavenging garbage collectors and incremental garbage collection. Both of these remove the pauses which garbage collections normally impose. Furthermore, they allow garbage collecting in parallel, if special purpose hardware is used. However, on conventional machines, these techniques do not seem to offer better performance than other storage reclamation techniques - parallel garbage collection would be disastrous on a time-shared machine.  Another possibility is to use some kind of hybrid scheme. A cell's reference count might be replaced with a flag which indicates that the cell is possibly referenced by more than one pointer (this is all that the optimized array predicate chElem requires). A general marking garbage collector would be needed to remove circular lists and to handle cells referenced by more than one pointer. However, such cells would tend to be more permanent, so the garbage collector would not be needed nearly so often and it would have less garbage to scan than if the simplified reference counting were not used. Combined with Bruynooghe's and Kluzniak's schemes for static analysis, this hybrid scheme may reduce heap management overhead to a very low number.  155  Extended pure Prolog Conventional Prolog is inadequate for good logic programming. Some people have described it as the "Fortran of logic programming." When I think of Fortran, I am impressed by how advanced it was for its time and I am depressed because it has continued to be widely used, even though far superior languages have been invented later. Fortran was once a great advance in computer technology; it has become a hindrance.  It would be sad if Prolog were to become like Fortran. Until someone invents the Algol or Pascal of logic programming, we should consider some extensions to conventional Prolog which make it into a purely logical language: •  Adding flexibility to the execution order, including  if-then-else,  sound  negation and coroutining (the delay notation for these is described in section 1.5, "Delay notation" on page 8). •  Extensions to Horn logic, including all solution predicates (setof and 156  bagof).  Logical arrays and logical input/output. No "impure," non-logical predicates.  e description is divided into sections:  Section 8.0, "Execution order" on page 158 discusses how Prolog's top-down left-to-right execution strategy can be replaced by a more flexible strategy which xpPAM implements as efficiently as conventional Prolog's execution strategy. Section 9.0, "Coroutining, pseudo-parallelism and parallelism" on page 185 discusses how coroutining can be added to Prolog, including some examples which show the greater expressive power and efficiency that can be obtained. Extensions of coroutining to parallelism is also discussed. Section 10.0, "Extensions to Horn logic" on page 206 discusses negation, if-then-else,  all-solution predicates, higher order predicates and  meta-programming which contribute to making logic programming more expressive and practical. These extensions can be "purely" logical. Section 11.0, "Arrays and I/O done logically and efficiently" on page 222 discusses how arrays and I/O can be correctly incorporated into a logic programming language so that there is no need for predicates with side effects.  157  8.0 Execution order  Good order is the foundation  — Edmund Burke, Reflections  8.1  of all good things.  on the Revolution  in France  Language issues  Conventional Prolog uses a strict top-down left-to-right execution strategy. A detailed description is given in [van Emden 1982]. This execution order has a number of undesirable affects:  •  Some programs depend on the execution order to terminate successfully. These are inherently non-logical, typically using cut ("!") or var.  •  Some programs terminate correctly, not depending on the execution order, but use non-logical I/O predicates which do depend on execution order.  Because of such programs, some optimizations cannot be applied. The situation is similar to conventional programming languages where an optimizer is prevented from transforming f ( x ) + f (x) to 2*f (x) because of possible side effects inside f ( x ) (if you don't think this is a problem, just try this optimization on a program containing  rand()+rand()).  158  XpProlog  solves these problems by providing more expressive and safer control  constructs than  cut  ("!") and  var.  XpProlog  also provides logical I/O (see section  11.8, "Sequential Input/Output" on page 241).  Furthermore, more flexible execution order allows programming with coroutines. These are discussed in the next chapter.  8.2  Conventional Prolog's execution order  Conceptually, a Prolog program is a database of clauses which are processed in a rigid order. A clause is an expression of the form H: — G H  is the  clause head  and C, through  G„  make up the  u  G , ... , G„ where 2  clause body,  or fact has no goal in its body (the  composed of  is omitted); a rule  goals.  A unit clause  clause  has at least one goal in its body. Waterloo and subsequent IBM Prologs  use " <-" instead of  and "&"s instead of the commas separating goals.  20  The clause's head and goals are all  atomic formulas  where p is an n-place predicate symbol and the form f[t  u  t ,... , t„) 2  t  x  of the form  through  t„  are  p(t  t ,... , t )  u  2  Terms are of  terms.  where /is an n-place functor symbol and  n  t  ]  through  t„ are  variables or terms. 0-place functors are atomic, that is names (strings) or numbers.  2 0  Doth notations arose at about the same time, causing the original Marseilles notation (the head of a clause is marked with a "+" and the goals are marked with "—"s) to disappear. From a parsing point of view, the IBM notation is easier to handle, although it still overloads the dot ("."). 159  A  query  is of the same form as a clause body. Conventional Prolog computes  the solution to a query by trying the goals from left to right: Try a goal: The database is searched from the beginning for a matching clause  head: Found: The database position is remembered and variables are instantiated by unification (matching is often implemented by attempting a unification and undoing instantiations if the unification fails). If the clause body contains any goals, the goals are recursively tried from left to right. Once all the body's goals have been tried successfully, the goal has  succeeded.  Not found: The goal has failed.  Computation  backtracks  to the previous  goal which is re-tried. Re-try a goal: Any variables which had become instantiated by unification are uninstantiated. The database is searched from the remembered position and execution continues as for "try a goal" (either "found" or "not found."  Trying a goal is very similar to calling a subroutine in a conventional programming language. Prolog can be efficiently implemented with a call stack which is slightly modified to allow for backtracking. As variables become instantiated, they are recorded so that they can be uninstantiated on backtracking.  This execution strategy can be improved in several ways:  •  The clauses for a predicate can be grouped together so that only a small part of the database is searched when trying a goal. The clauses can be compiled 160  to an abstract machine code (only for predicates whose clauses are not dynamically added to the database). Abstract machine code may be implemented directly in hardware, interpreted, or translated to machine instructions on conventional computers. Conventional wisdom states that interpreting abstract machine code is 10-20 times faster than processing the raw clauses. If the abstract machine code is transformed to machine instructions, a further 2 — 6 times speed-up is common. Hardware implementations allow cost reductions of one or two orders of magnitude. •  When a clause head matches and no other clause heads for the predicate can possibly match, some book-keeping information does not need to be kept ("choice points" for backtracking do not need to be created).  •  When a variable which is newer than the last backtrack point becomes instantiated, it does not need to be recorded because it will automatically be deallocated when backtracking unwinds the stacks.  •  When a predicate is called with all its parameters fully instantiated (no variables anywhere within its parameters), there is no need to ever re-try it — if it were to succeed again, it would provide no new information.  •  When a predicate never instantiates anything within its parameters or does not instantiate a variable which is used by a later goal, there is no need to ever re-try it (this is a more general case of the previous case).  •  When the last goal of a clause is tried, some book-keeping information is not needed on the call stack  (tail recursion  optimization  (TRO)). If the goal is  also in the last clause of a predicate, additional savings are possible.  These optimizations do not affect the order of execution; they simply speed execution and reduce stack usage. Details of some of them are given in section 5.0, "Compiling Prolog" on page 85. 161  Although easy to implement (at least, without the above optimizations), conventional Prolog's left-to-right execution strategy is inflexible.  •  Unnecessary goals may be tried on backtracking. Sometimes this is merely very inefficient; in other cases, backtracking goes into an infinite loop.  •  Excessive stack may be used. Tail recursion optimization can often prevent this.  •  Negative goals can be executed correctly only in a restricted form which is not usually detected in conventional Prologs. Similarly, predicates such as bagof  •  and setof must be restricted.  Problems which can be naturally expressed as coroutines must be coded in an unnatural way. That is, x p P r o l o g allows writing some programs in a natural way but conventional Prolog requires changing predicates into an unnatural form.  A subtle defect in conventional Prolog's execution strategy is that many programmers write predicates which depend on the execution order. Not only is this poor programming style, contrary to the spirit of logic programming, but it also inhibits some automatic program optimizations. Most predicates which do not depend on the execution order are penalized by the minority which do. A similar situation exists for conventional programming languages which allow side-effects inside functions which prevents optimizing f ( y ) + f ( y ) to 2 * f ( y ) . .  162  8.3 Example of conventional execution order The conventional execution order, described above, has the flavour of a conventional machine with backtracking thrown in. Execution proceeds sequentially, goal by goal. When a goal cannot be satisfied (it fails because no matching clause can be found), the machine is restored to an earlier state and an attempt is made to try an alternative clause. For example, suppose the following clauses exist: a a bl b2  :- b l , b2, b3. :- b4. :- c l , c 2 . :- d l .  /* no c l a u s e f o r b3. */ b4. /* u n i t c l a u s e : no g o a l s */ c l . /* u n i t c l a u s e : no g o a l s */ c2. /* u n i t c l a u s e : no g o a l s */ /* no c l a u s e f o r d l . */  We can transform this into a table: Clause # Head  Goal 1  Goal 2  Goal 3  1  a  b2  b3  2  a  3 4 5  bl b2 b4  bl b4 cl  6 7  cl c2  c2  dl  We can treat the query " ? - a " as clause 0 with the single goal "a."  163  1. Goal a has two alternatives: cl ause 1 (a: -bi,b2,b3) or clause 2 (a: -b4).  The  first alternative is taken. 2. Clause 1 (a: -bl,b2,b3) tries its first goal (bl). 3. Clause 3 ( b l : - c l , c 2 ) matches and in turn tries goal c l .  4. Clause 6 ( c l ) matches. It is a unit clause and succeeds. The next goal in clause 3 is c2 5. Clause 7 (c2) matches. It is a unit clause and succeeds. This causes goal b l to succeed. The next goal is b2 in clause 1. At this point the call stack is as follows (in a clause, the "•" is after the goal being tried). Effects of tail recursion optimization (TRO) are not shown.  ?- a. • a :- b l , • b2,  b3.  Clause # Goal # Alternative clause # 0 I I I 2 3  2  -  6  -  -  a :- b l , b2, • b3.  1  2  -  b2  4  1  -  7  -  -  bl cl.  :- c l , c2. • • :- d l . •  c2. •  Nothing matches the goal d l , so it fails. There are no alternatives for c2 and b2, so they also fail. Clause 2 (a: -b4) is tried. This tries clause 4 (b4) and there is nothing else left to try, so the query succeeds:  164  ?- a. •  Clause # Goal # Alternative clause # 0 1  a :- b4. •  2  1  -  b4. •  4  -  -  As far as the correctness of the solution is concerned, the machine could have tried the clauses and goals in any order. The goals in clause 1 could have been tried in the order b3, b2, b l (which would have failed more quickly) or clause 2 could have been tried before clause 1 (which would have succeeded without any backtracking).  8.4  A more flexible execution strategy  Instead of viewing the computation as using a stack, we could think of it as a series of goal re-writings, much like a Markov algorithm. The above computation could be given as the series of states (in each line, the "•" is to the left of the goal which is about to be tried): Apply clause  State a •  a :- b l , b2, b3.  • b l , b2, b3  b l :- c l , c2.  • c l , c 2 , b2, b3  cl.  cl,  » c 2 , b2, b3  c2.  cl,  c 2 , • b2, b3  which gets to a failure point. Backtracking to clause 2 gives the alternative series:  165  Apply clause  State a • b4 •  a :- b4.  This succeeds because all the goals are unit clauses (here there is only one such goal: b4). In goal re-writing, any untried goal may be picked for re-writing. If it succeeds, it is replaced by all the goals of the matching clause - if there is none (a unit clause), the goal is removed from the set of goals which are candidates for re-writing. If it fails, the computation must backtrack to an earlier state and try different goals. The easiest goal re-writing strategy to implement is one which proceeds strictly from left to right. The machine keeps a position marker within the re-write goals — everything to the left of the pointer has already succeeded by matching a unit clause and everything to right of the pointer has not yet been tried. The entire computation succeeds when there is nothing left to try. This is conventional Prolog's left-to-right, top-down strategy.  XpProlog  uses an extended strategy which allows some goals to be delayed.  Delayed goals are removed from the re-write goals until something causes them lo be resumed.  A  resumed goal is put into the re-write goals immediately to the  right of the pointer. In this way, every goal is tried, but possibly in a different order from what conventional Prolog would do.  166  In the above example, assume that goal c l must delay until goal c2 instantiates something. Then the computation would be: State  Apply clause  - a|| a :- b l , b2, bl  b3.  :- c l , c2.  • b l , b2, b3 || • c l , c2, b2, b3 ||  delay c l • c2, b2, b3, || c l  c2.  {cl},  resume c l  c2, • c l , b2, b3 ||  cl.  c2, c l , • b2, b3 ||  The "||" is used to mark goals which have been delayed and {...} marks where they were originally.  The delayed execution has now reached a failure point. Backtracking will cause execution to continue as for the non-delaying execution (above) because clause cl has no other alternatives.  In this example, the delay computation adds steps because goal c l must first be delayed and then resumed. However, some problems can be sped up significantly by delays (see section 9.3, "Test and generate" on page 191). With goal re-writing, backtracking always restores the machine to an earlier state. The pointer is moved left one goal and all the goals to the right of the pointer are thrown away. The machine then tries a different clause match for the next goal. If this fails, backtracking will throw away that goal and move the pointer left one more goal. The entire query will fail if the pointer is moved all the way to the left. 167  When backtracking occurs, delayed goals must also be removed. This strategy - a left-to-right with delays - will get to every successful state that a sequential left-to-right computation would produce, although possibly in different order.  If execution tries to go past the "||," the answer is not "yes" or "no" but "maybe (insufficient information)."  8.5  Negation  Horn clauses contain only positive literals. Simple  not equals  can be considered  as a convenient abbreviation for the infinite list of clauses: a  ^  b  ± a.  b.  a + c.  1 i= 2. []  *  [x].  This definition is sound, even if one or both of the arguments are uninstantiated. The goal X^a will generate the solutions X=b, X=c, X=l, etc. on backtracking - a rather inefficient situation. Many other useful predicates such as less similar way to not  equals.  than (<), c a l l , name  and univ (=..) are  Such built-in predicates can also be thought of as  containing an infinite number of rules like:  168  call(p(Q))  :-  p(Q).  c a l l ( q ( A , B , C ) ) :- q(A,B,C). f ( A , B ) =.. [ f , A, B] g(A,B,C) =.. [g,A,B,C].  Conventional Prolog implements negation unsoundly; the not predicate ("\") is an attempt to implement negation as failure: n o t ( T e s t ) :- c a l l ( T e s t ) , not(Test).  !, f a i l .  /* f a i l s i f T e s t succeeds */ /* succeeds i f T e s t f a i l s */  Negation as failure [Clark 1978] requires that the terms be fully ground (do not contain any uninstantiated variables). If  Test  is completely ground, the above  definition of not is sound - it is equivalent to the definition given earlier with the "infinite" list of clauses. However, if  Test  contains any uninstantiated  variables, this definition of not may produce a wrong answer. Consider the list membership predicate: member(X, X . _ ) .  member(X, _ . R s t ) :- member(X, R s t ) .  The query ?- Z=4, n o t member(Z,  [1,2,3]).  correctly succeeds (all the terms are fully ground) but the logically equivalent ?- n o t member(Z,  [1,2,3]),  Z=4.  fails because the call to not contains an uninstantiated variable. Here is the conventional, incorrect, execution sequence: 1. The n o t predicate tries member(Z, [ 1 , 2 , 3 ] ) . Member's first clause succeeds, instantiating Z=l. 169  2. The  cut  ("!") in n o t removes all other choices for  member.  3. Z=4 is tried. Because Z is already instantiated to the value l, this fails. 4. Execution backtracks. The cut ("!") in n o t has already removed the other choices for  member,  so the entire query fails.  This can be remedied if not delays until its arguments are sufficiently instantiated. That is, not should act as if it had succeeded and be retried later when the variable which caused the delay becomes instantiated. Using the more flexible execution strategy given earlier, here is a correct execution sequence:  1. The  not  predicate delays trying  member(Z,  [1,2,3])  because of the  uninstantiated variable z. 2. Z=4 is tried, instantiating z to the value 4. 3. The n o t predicate is resumed by Z becoming instantiated. It now tries member(Z,  trying  [1,2,3]).  member(4,  Because Z already has the value  [1,2,3])  which fails, causing  not  4,  this is equivalent to  to succeed.  4. The entire query succeeds. The above method of handling negation by delaying is not the only one possible. For example, [Voda 1986b] describes a more active approach in which environment can be propagated in a negative context.  8.6 Single solutions For efficiency, the programmer may want member to succeed only once so that unnecessary backtracking is avoided - that is, the goal is "find the first X such that  Here is an attempted solution using cut. This technique is also often 170  used to minimize stack usage. This poor programming practice is unnecessary if the Prolog implementation has tail recursion optimization (TRO).  XpProlog  has TRO. memberl(X, X._) :- !. /* succeed o n l y w i t h f i r s t match */ memberl(X, Y.Rst) :- memberl(X, R s t ) .  The query ?- Z=2, memberl(Z, [ 1 , 2 , 3 ] ) .  correctly succeeds but the logically equivalent ?- memberKZ, [ 1 , 2 , 3 ] ) ,  Z=2.  incorrectly fails because the cut ("!") removes too many backtrack points. Memberl  executes wrongly when Z is a variable. This can be made explicit,  although rather clumsily, using var. memberla(X, memberla(X, memberlb(X, memberlb(X, memberlb(X,  L i s t ) :- n o n v a r ( X ) , !, memberlb(X, L i s t ) , L i s t ) :v a r ( X ) , !, member(X, L i s t ) . Y.Rst) - n o n v a r ( Y ) , X = Y, ! Y.Rst) v a r ( Y ) , X = Y. _ . R s t ) - memberlb(X, R s t ) .  But this is Still not sufficient. ?- member( [A,B], [ [ 1 , 2 ] , [ 3 , 4 ] ] ) , A=3, B=4. executes incorrectly. Instead, the  var_anywhere  predicate must replace the  var  check for uninstantiated variables anywhere within the term, with increased computational cost. A solution is to use an  if-then-else  which delays until the test is sufficiently  instantiated:  171  to  meraberlc(X, Y.Rst) :i f X=Y t h e n t r u e e l s e memberlc(X, R s t ) e n d i f . ?- memberlc(Z, [ 1 , 2 , 3 ] ) , Z=2  ("true"  is a predicate which always succeeds.)  The call to memberlc provisionally succeeds with the test Z=l delayed until Z becomes instantiated (Z=2). The resumed test (Z=l) fails and execution backtracks to the  else.  This tail recursively calls memberlc which again  provisionally succeeds with the test z=2 delayed until z becomes instantiated. The next goal z=2 instantiates z and the entire query succeeds. Backtracking will not re-do this computation because else  then  removes the backtrack point for the  (this can be done safely because the test is fully instantiated and will  therefore not produce any more information if it is retried).  If-then-else  is a convenient abbreviation for a combination of equals and  not  equals:  memberld(X, Y.Rst) :- X = Y. memberld(X, Y.Rst) :- X # Y, memberld(X, R s t ) .  An optimizer can transform this into the  if-then-else  form. The optimizer can  also notice that the equality/inequality test must not cause any variables to become instantiated: different unification instructions are used in this case. The computational cost of this member/ predicate is linear with the size of the list, even with full checking for delay situations.  172  8.7  Cut  [O'Keefe 1985] and [Debray 1986] describe the uses of cut. The common uses are:  •  Committing a clause.  •  Forcing computations to be functions (that is, deterministic).  •  Breaking out of failure-driven  (repeat/fait)  loops.  All of these can be eliminated from Prolog programs by if-then-else.  An  optimizing compiler can detect these situations in x p P r o l o g programs and put cuts into the generated xpPAM code when they can be justified by maintaining the meaning of the original, purely logical, predicate.  Because x p P r o l o g allows predicates to delay, the meaning of cut is very difficult to give. Consider ?- g o a l l , g o a l 2 , !, g o a l 3 .  where g o a l l has delayed and goa!2 has succeeded. The next goal is cut  ("!")  which removes all the alternatives within g o a l l and goal2. Goal3 causes g o a l l to wake up by instantiating a variable that g o a l l is waiting on. If g o a l l then fails, everything will fail because there are no alternatives. However, an alternative in g o a l 3 might produce a different value which would cause g o a l l to succeed. The problem is that g o a l l did not really succeed — it only provisionally succeeded, pending more information (from g o a l 3 in this case). The implementation of cut must be very complex to handle such cases properly - they can be handled much more cleanly by  173  if-then-else.  Tail recursion optimization removes the need for special  repeat  loops. In  is written  x p P r o l o g , repeat  r e p e a t ( P ) :- P, r e p e a t ( P ) .  which never consumes more stack than that required by P if P is deterministic. If it is not deterministic, the following is required: r e p e a t ( P ) :- i f e x i s t s P s u c h t h a t P t h e n r e p e a t ( P ) e n d i f .  where the  if-then-else  removes backtrack points. This is still logical but only the  first solution for each invocation of P will be found. Strictly speaking, repeat is not logical (if P has no side effects, it will always give the same result, so  repeat(P)  will either fail or loop forever. Typically, repeat is  used to handle a read-execute-write loop. As x p P r o l o g provides logical I/O with streams,  repeat  and  P  can be automatically transformed as described in section  11.8, "Sequential Input/Output" on page 241: r e p e a t ( P , S t d l n , StdOutRest, S t d l n 2 , StdOut) :P(StdIn, Std0ut2, Stdln2, StdOut), r e p e a t ( P , S t d l n , StdOutRest, S t d l n 2 , S t d 0 u t 2 ) . ?- s t r e a m ( ' < t e r m i n a l > ' , i n p u t , I n ) , /* i n p u t stream from t e r m i n a l */ s t r e a m ( < t e r m i n a l > , o u t p u t , O u t ) , /* o u t p u t stream t o t e r m i n a l */ r e p e a t ( P , I n , Out, I n 2 , [ ] ) • /* In2 i s r e m a i n i n g i n p u t */ 1  where  Stdln  1  and  StdOut  are the input and output streams (the other  variables are used to avoid calls to  Std  append).  Because many implementations lack tail recursion optimization and proper heap management, some programmers write the following for "infinite" loop programs (such as the top level of a read — eval interpreter):  174  repeat. r e p e a t :- r e p e a t . ?- r e p e a t , ..., fail.  Here, the f a i l chops the backtrack stack down and allows the execution stack to be freed. Because xpPAM fully supports tail recursion optimization and also garbage collects the global heap, there is no need to write in such a style.  The predicate p r e d ( [ ] , R) :- p l ( R ) . p r e d ( a . T l , R) :- p 2 ( T l , R ) . p r e d ( b . T l , R) :- p 3 ( T l , R ) .  can be written using  if-then-else  as  p r e d ( P l , R) :i f PI = [] then p l ( R ) e l s e i f P I = Hd.Tl t h e n if Hd = a t h e n p 2 ( T l , R) e l s e i f Hd = b then p 3 ( T l , R) end i f endif.  The programmer may use either form; the optimizer generates the same code for both forms.  If-then-else  delays the predicate if there is a variable anywhere in the test, not  just at the top level. This can be explicitly changed by adding " e x i s t s variable-list  suchthat."  For example:  175  lookup(Name, V a l u e , FoundV, NameList, NewNameList) :I f e x i s t s FoundV s u c h t h a t member(Name-FoundV, NameList) t h e n NewNameList = NameList e l s e NewNameList = (Name-Value).NameList, FoundV = V a l u e endif.  Here,  member  is used to try to find the Name-FoundV pair within NameList. If it  is found, the name list remains the same; if it is not found, the new Name-Value pair is added to the name list. In both cases, FoundV gets the value associated with the name.  If-then-else  first(P)  can be used to make a general first  solution  predicate:  :- i f e x i s t s P s u c h t h a t P then t r u e e l s e f a l s e e n d i f .  There is no delaying within P. Once it succeeds, no other alternative is tried. Member!  should be coded differently:  m e m b e r l ( X , L i s t ) :- i f e x i s t s X s u c h t h a t member(X,List) then true e l s e f a l s e endif.  which is less efficient than  memberIc  because List must be scanned to ensure it  has no uninstantiated variables.  8.8  Clause order  If a program uses only logical predicates, the order of clauses has no effect on the meaning of a predicate. However, the order can affect efficiency and, sometimes, whether or not it will terminate. Clause and goal order in x p P r o l o g can be considered as strong hints to the compiler. XpProlog's optimizer is allowed to change the order of clauses in a predicate (an example is given in section 5.5, "Optimized compiling of a predicate" on page 98). Usually, the 176  order will not be changed because the optimizer assumes that the most common cases have been put first. If a predicate is deterministic, then only one clause will be tried; the order of clauses does not matter. If the clause is not deterministic, the clauses will be tried in the order that they are given, although the optimizer may be able to skip over some.  To allow compatibility with conventional Prolog, a predicate can be marked as requiring conventional ordered evaluation. For example this non-logical Prolog program p(a,X)  :-  !,  q(X).  p(b,X)  :-  !,  r(X).  P(_,X)  :-  s(X).  /* A *  a,  A #  b */  is better written in x p P r o l o g using the special else predicate as: p(a,X)  :-  q(X).  p(b,X)  :-  r(X).  p(A,X)  :-  else(A),  s(X).  /* A *  which does not have any cuts. The  a,  A +  b */  xpProlog  order for correct execution of the query  version does not depend on clause  ?-p(Z,i)  —  on backtracking, it could  generate z=a, Z=b and Z = _ but the Prolog version would only generate z=a. 21  21  Strictly speaking, the solution Z=  should be the constraints Z#a, Z^b.  177  The programmer might prefer an  if-then-else  construct instead of the  else  predicate. The compiler would generate exactly the same code if the p predicate were written: p(A,X) :- i f A s u c h t h a t A=a t h e n q ( X ) elseif  A s u c h t h a t A=b t h e n r ( X ) else s(X)  endif.  If the  suchthats  were left out, the predicate would delay until  A  became  instantiated — a similar effect could be produced by adding ?- p r o c e e d p ( a ? , x ) .  8.9 AH solutions predicates The problem.of uninstantiated variables extends to second-order predicates such as  bagof  setof  and  setof.  The meaning of conventional implementations of bagof and  changes according to how the variables are instantiated when they are  invoked, giving different meanings to programs according to the computation order (see section 10.3, "Setof, bagof on page 211 for the definition). The usual (unsound) implementation is something like:  22  2 2  This  bagof (and  xpProlog's  bagof) returns nil ([  ]) if nothing is found. Some people prefer  predicates which fail if nothing is found. This can be easily defined: bagof_fail(Proto,  Pred,  Result)  : - bagof(Proto, Result  178  #  []  Pred,  Result),  b a g o f ( P r o t o , P r e d , L i s t ) :- ( b a g o f A s s e r t s ( P r o t o , Pred) ; t r u e ) ,  !,  (bagofRetracts(List) ; true), b a g o f A s s e r t s ( P r o t o , Pred) :- P r e d , assert(bagof(Proto)), f a i l . /* t r y another s o l u t i o n o f P r e d */ b a g o f R e t r a c t s ( X . R e s t ) :- c l a u s e ( b a g o f ( X ) ) , /* f a i l s when no more */ retract(bagof(X)), bagofRetracts(Rest). s e t o f ( P r o t o , P r e d , L i s t ) :- b a g o f ( P r o t o , P r e d , L i s t l ) , sortBag(Listl, List).  The bagof A s s e r t s first tries P r e d and then asserts the prototype Proto. A f a i l causes backtracking to try the next possibility of P r e d (backtracking over a s s e r t does not undo anything). The b a g o f R e t r a c t s simply picks up all the entries which have been added to the database, removing them as it goes. This definition will return [ ] if P r e d fails.  There are several problems with this approach: •  It does not always have a declarative reading.  •  The a s s e r t and r e t r a c t predicates are quite expensive. This can be solved by using special purpose internal database predicates.  •  The entire solution set must be generated, even if it is not needed. For example, if not is defined by bagof (_,Pred, []), all the solutions to Pred are  generated and then b a g o f R e t r a c t s fails as soon as the first element is found in the database. This also leaves un-retracted entries in the database, so not should actually be defined b a g o f ( _ , P r e d , L ) , L ? t [  ].  (Setof  must still generate  all the solutions because it must remove duplicates, usually by some kind of sorting.) 179  The sound implementation (using xpPAM) is very similar to that given above except that, when coded in abstract machine code, the solutions list can be generated directly. There is no need to enter and delete entries in a database. The xpPAM code is a little tricky and is not shown here. It is basically (see section 4.0, "Backtracking and delaying" on page 66 for a description of the instructions):  delay as necessary make choice point  (failOne)  make fresh copy of P r o t o call P r e d modify choice point  (failTwo)  append new list element with fresh copy of P r o t o fail /* doesn't remove copy of P r o t o on backtracking */ failOne:  remove choice point unify [] with (tail of) L i s t return failTwo:  fail /* the "add new list element" failed */ "Make fresh copy of P r o t o " means that a brand new copy is made, with no references to any free variables in P r o t o (see the meta and  unmeta  predicates in  section 10.4, "Meta-variables" on page 212). As these variables are newer than the top choice point, they will left alone on backtracking; however, the "append new list element" must not be recorded on the reset stack so that it will not be undone when the f a i l causes backtracking to the choice point within Pred. 180  When there is no more choice in Pred, the choice point is used to go to label fallOne  which terminates the solution list with  nil.  [Naish 1985c] advocates a notation which gives purely declarative readings to all solutions predicates. His design requires an implementation which can delay predicates until their arguments become sufficiently instantiated. XpProlog's meta-variables are a generalization of Naish's design. An all solutions predicate must delay until all non-local variables are ground. For example, in s e t o f ( [ X , Y ] , p r e d ( Y , Z ) , Ans)  the setof will delay until Z becomes ground. If this is written e x i s t s Z s u c h t h a t s e t o f ( [ X , Y ] p r e d ( Y , Z ) , Ans)  there will be no delay. X and Y are "local" to the The problems with problems with  not  bagof  and member  and  setof  1  setof  expression.  can be considered as special cases of the  because not and member  1  can be defined  n o t ( T e s t ) :- b a g o f ( X , T e s t , [ ] ) . /* n o t h i n g succeeded */ memberl(X, L i s t ) :b a g o f ( X , member(X, L i s t ) , X . _ ) . /* f i r s t s o l u t i o n o n l y */  8.10 Non-strict execution order: delays Most of the time, top-down left-to-right execution order is adequate for executing a Prolog program. But sometimes a more flexible execution order is needed.  Some conventional Prologs allow deviation from strict top-down left-to-right execution order by the  freeze  predicate. Using this, member 181  1  can be written:  memberle(X,  Y._)  :-  memberle(X,  Y.Rst)  freeze(X, :-  freeze(Y,  memberle(X,  This only works if the scope of the  (X=Y,  !))).  Rst).  cut  is the entire m e m b e r l d clause, not just the  freeze.  Actually, two kinds of freeze are needed: one which delays until the argument is sufficiently instantiated  (freeze)  and one which delays until the argument is  completely instantiated  (freeze_all).  Using this, we can produce one more  definition: memberlf(X,  List)  :-  freeze_all(X,  which is less efficient than  memberld  freeze_all(List,  memberl(X,  List)),  because it traverses the list an extra time, to  detect uninstantiated variables. Also, m e m b e r l f may delay unnecessarily on a variable, for example, for  memberlf (1, [ 1 , X ] ) .  Memberlc,  which uses  if-then-else,  is the most efficient and the easiest to understand.  Delays  are the general mechanism for sound negation, single solutions,  if-then-else  and declarative all solutions predicates. Delays also allow the  extensions to Prolog, including first-order quantifiers, described in [Lloyd and Topor 1984] and they allow coroutining with backtracking. The result is a very flexible design which permits a variety of programming styles. Delays are covered in section 9.2, "Coroutining example" on page 187. To take full advantage of the design, Prolog's syntax must be extended slightly - these extensions will be introduced as they are needed.  182  8.11  Input/Output  Input/output is a notoriously non-logical part of conventional Prolog. The most common design follows BCPL [Richards and Whitby-Stevens 1979]: input or output is always to the current stream and this stream can be changed at any time to be connected to a specified physical file. Predicates such as display and read  manipulate the input/output streams using side-effects. Most Prolog texts  give examples where the unwary programmer can be caught by I/O's illogical behaviour.  The main problems with conventional Prolog's I/O are •  Backtracking does not undo an I/O operation.  •  If predicates delay, the order of I/O may not be what the programmer intended.  •  The conventional I/O predicates work by side-effects, preventing some optimizations.  These problems can be solved by following a model similar to that in [Cheng 1986]. Input/output are still handled by streams but these streams are syntactically identical to lists. Meta-predicates associate these streams with physical files (section 11.8, "Sequential Input/Output" on page 241).  For debugging, it is often helpful to add output statements to predicates. XpProlog  has retained this feature (which is non-logical) but its use is  discouraged, except in debugging.  183  8.12  Efficiency  In my implementation of xpProlog (described by the xpPAM abstract machine), deterministic predicates are not slowed significantly by the delaying and backtracking features. Depending on the problem, the machine is well suited 23  to purely deterministic predicates and to complex generate-and-test predicates using coroutines and backtracking.  These implementation features are synergistic. For example, if determinism is recognized, then unnecessary choice points are not created. This speeds execution, allows tail recursion to take place, reduces stack and memory usage and decreases page faults.  2 3  T h e current implementation does have an overhead of a few percent because each unification which instantiates something must check if a predicate has been delayed, depending on that variable. A better implementation would use a different tag for such variables and there would be no run-time overhead -  they would just become another case in the jump-table  used by the unification routine.  184  9.0 Coroutining, pseudo-parallelism and parallelism  Time present and time past Are both perhaps present in time future, And  time future contained in time past.  — T. S. Eliot, Four  Quartets:  Burnt Norton,  Coroutining has been widely written about, but very few programming languages provide facilities for coroutining.  This chapter provides some examples of coroutining, to show how they lead to more natural programs than can be provided by purely procedural languages. In addition, coroutining follows naturally from logic programming's philosophy in that the order of execution does not affect the correctness of the solution. Thus, almost nothing must be added to the language to provide coroutining, although the implementation must change. Coroutining can be considered as a special case of parallelism. Section 9.6, "Parallelism" on page 198 has a short discussion of the two paradigms.  185  J  Coroutining is not a new notion for logic programming. It is discussed in [Kowalski 1979]. Some other coroutining proposals are discussed at the end of this chapter.  9.1  Notation  Coroutining can be explicitly indicated by the "?" notation introduced in section 1.5, "Delay notation" on page 8 The "?" notation does not change the logical meaning of a programming except that the same program without "?"s may either fail to terminate or may run much slower.  Coroutining can be deterministic or non-deterministic. The deterministic case is similar to what is provided in conventional languages such as Simula-67. The non-deterministic case exists only for logic programming languages; its main use is in speeding up generate-and-test programs. The "?" notation can be used within predicates or as separate  proceed  declarations (also described in section 1.5, "Delay notation" on page 8). Proceed  declarations have the advantage that they can be written separately  from a predicate's declarations. [Naish 1985a] shows how similar declarations (wait declarations for MU-Prolog) can be generated automatically. The algorithm is surprisingly simple and can be implemented in a few hundred lines of Prolog. Of course, such an algorithm cannot generate delay declarations for all predicates (which would solve the halting problem) but it works for most cases encountered in practice. 186  9.2  Coroutining example  Here is a short, artificial example program to demonstrate coroutining. It is typical of a large class of programs which have a natural representation as cooperating parallel processes or coroutines. Such programs often are much harder to understand if they are transformed into a strictly sequential style. [Kowalski 1979 pp. 114-118] has other examples. [Dahl, Dijkstra and Hoare 1972] has more general comments on why coroutines are desirable, and many more examples.  The problem is to generate a list containing 1 and all numbers divisible only by 2 and 3, in ascending order with no duplicates: [1, 2, 3, 4, 6, 8, 9, 12, 16, 18, 24, ...]. This is similar to an exercise attributed to R. W. Hamming for which a sequential, less understandable, solution is given in [Dijkstra 1976]. [Henderson 1980 chapter 8] gives a functional lazy execution solution to this problem. I have restated his solution with predicates, giving an elegant solution using coroutines (degenerate parallelism).  The problem can be re-stated:  •  1 is in List.  •  if TV is in List, so are IN  •  no other values are in  List.  •  if List contains [..., A,  B,  and  3/V.  ...] then  A <B  (no duplicates, in ascending order).  Pictorially, the flow of data through the program is: 187  timesList 2  timesList 3  Twos >-  Threes  ordMerge  Merge23  List  The predicate t i m e s L i s t generates a list of multiples in the third argument. For example: timesList(2,  [1,3,4], [2,6,8]).  The predicate ordMerge does an ordered merge of two ordered lists, removing duplicates. For example: ordMerge([1,3,4],  [2,4,5], [1,2,3,4,5]).  This program can be coded in x p P r o l o g : h a m m i n g ( L i s t ) :- L i s t = 1 . Merge23, ordMerge(Twos, T h r e e s , Merge23), t i m e s L i s t ( 2 , L i s t , Twos), t i m e s L i s t ( 3 , L i s t , Threes). o r d M e r g e ( [ ] , M, M). ordMerge(M, [ ] , M). ordMerge(A.X, A.Y, M) :- ordMerge(X, A.Y, M). /* remove dup's */ ordMerge(A.X, B.Y, A.M) :- A < B, ordMerge(X, B.Y, M). ordMerge(A.X, B.Y, B.M) :- A > B, ordMerge(A.X, Y, M). timesList(_, [ ] , [ ] ) . t i m e s L i s t ( N , A.X, An.Xn) :- An i s N*A, t i m e s L i s t ( N , X, X n ) .  This solves a set of simultaneous equations on List with Merge23  3  as temporaries. Each of the  timesList  Twos, Threes  and  processes produces a list of 2 or  times its input list. The ordMerge process merges them (in order) and the 188  resulting list is then fed back into the  timesList  processes. The duplicate  removal is not completely general (it assumes that each of the input lists is in strictly ascending order) but is adequate for this problem. [Henderson 1980] outlines a proof that the above computation is well-defined.  Delays are necessary for ordMerge and  timesList.  OrdMerge  must delay until  its arguments are sufficiently instantiated for the ordering test (or until an argument is nil); timesList must delay until the arithmetic "N*A" is possible: ?- p r o c e e d ordMerge(a?.x, b?.y, m). ?- p r o c e e d t i m e s L i s t ( n ? , a?.x, l i s t ) .  The program is very easily extended to deal with any number of t i m e s L i s t s by just adding more ordMerge stages - modifying the sequential solution is more complex. Programmers who are used to sequential programs often have difficulty with the notion of predicates delaying and then restarting as their arguments become instantiated. Tracing the sequential execution of these predicates is very tricky. If these are thought of as parallel processes, some of the difficulty goes away — coroutines are just special cases of parallel processing where each process runs until it cannot proceed or until is provides a value which is needed by a suspended process. With this in mind, here is how execution proceeds:  1. ordMerge  delays on Twos and Threes.  189  2. the two t i m e s L i s t predicates produce their first elements: [2, ...] and [3, ...]. They then delay. 3. ordMerge wakes up and produces itsfirstoutput: [2, ...]. It can proceed no further. 4. the two t i m e s L i s t predicates produce their next elements: [2, 4, ...] and [3, 6, ...]. They then delay. 5. ordMerge wakes up and produces its next output: [2, 3, ...]. It can proceed no further. 6. the two t i m e s L i s t predicates produce their next elements: [2, 4, 6, ...] and [3, 6, 9, ...]. They then delay. 7. ordMerge wakes up and produces its next output: [2, 3, 4, ...]. It can proceed no further.  And so on. In actual coroutining, the switching back and forth is done more frequently (as soon as a t i m e s L i s t predicate produces another element, ordMerge  is resumed) but the net effect is the same.  The notion of delaying on variables is very similar to the concept of  monitors  [Hoare 1985]. Each variable which has caused a delay (by being marked using "?" or proceed notation) is like a monitor which suspends execution until sufficient information is available. The concept of critical region, however, is slightly different because there is no destructive assignment in logic programming.  These proceed declarations were generated automatically by using the method in [Naish 1984] from the original program. The proceed declarations are purely  190  control information, separate from the logic of the program. However, they are necessary to ensure that the computation terminates.  9.3 Test and generate Not only does coroutining allow for easier to understand programs, it can produce significant speed-ups. Generate and test programs typically are written ?- g e n e r a t e ( [ a i , &2> 3 > ••• ]> S o l u t i o n ) , a  test(Solution).  where generate generates all potential solutions by permuting the problem's variables. The computational cost of this explodes factorially. Using coroutines, this can be written: ?- t e s t ( S o l u t i o n ) , g e n e r a t e ( [ a ^ , &2> 8 3 , ... ] , S o l u t i o n ) .  The  test  predicate typically contains a series of inequalities and arithmetic  predicates which have very large solution spaces. These are delayed.  Generate  then produces thefirstpotential solution. It instantiates Solution sequentially from thefirstelement. As the elements become available, the delayed tests are tried. If a test fails, it automatically removes all other potential solutions with that particular initial sequence - far fewer permutations are tried.  An example of this technique solves the arithmetic logic puzzle SEND + MORE  9567 + 1085 ====>  MONEY  10652  191  (s, E, N, D, M, 0, R and Y are all different; M and S are not 0): Here is the program (adapted from [Colmerauer 1982]). The solution has no delay annotation because all the delaying is done by the ' V predicates. money(S, E, N, D, M, 0, R, Y) :d i f f e r e n t ( [ S , E, N, D, M, 0, R, Y ] ) , M ± 0, S * 0, sum(Cl, 0, 0, 0, M), sum(C2, S, M, C l , 0 ) , sum(C3, E, 0, C2, N), sum(C4, N, R, C3, E ) , sum(0, D, E, C4, Y ) . /* sum w i t h c a r r y : C + X + Y = 10*C1 + Z */ sum(C, X, Y, C l , Z) :goodCarry(C), d i g i t ( X ) , digit(Y), T i s X + Y + C, s p l i t ( T , C l , Z). /* compute 2 - d i g i t number i n t o i t s two d i g i t s : N = 10*C + X */ s p l i t ( N , C, X) :- C i s N / 10, X i s N mod 10. /* enumerate t h e d i g i t s : */ digit(O). digit(l). digit(2). digit(5). digit(5). digit(7).  digit(3). digit(8).  /* F o r a d d i t i o n , t h e o n l y p o s s i b l e c a r r i e s goodCarry(O). goodCarry(1).  digit(4). digit(9). a r e 0 and 1: */  /* E n s u r e t h a t a l l elements i n a l i s t a r e d i f f e r e n t : different([]). d i f f e r e n t ( X . L ) :- notMember(X, L ) , different(L). notMember(X, [ ] ) . notMember(X, A.L) :- X # A, notMember(X, L ) .  192  */  The  different  The  sum  predicate sets up the delayed tests S#E, S^N, S#D,  E#N,  R#Y.  predicate generates possible sums and carries for D+E=Y, N+R=E, etc. by  backtracking through the possibilities produced by digit and values of S, E,  goodCarry.  As the  Y become instantiated, the inequalities are woken up. For  example, the first sum generates an initial answer ci=o, M=o. This is immediately rejected by the constraint M^o, so the other sum goals are never tried with those values for C l and M. This drastically reduces the search space over what would be tried using conventional Prolog (by about 100 times).  The conventional Prolog solution would first have to re-arrange the goal so that the test predicates are after the generate predicates. money(S, E, N, D, M, 0, R, Y) : sum(Cl, 0, 0, 0, M), sum(C2, S, M, C l , 0 ) , sum(C3, E, 0, C2, N ) , sura(C4, N, R, C3, E ) , sum(0, D, E, C4, Y ) , M * 0, S it 0, d i f f e r e n t ( [ S , E, N, D, M, 0, R, Y ] ) .  All the possibilities produced by the sum predicates will be tried. Greater efficiency can be gained by interleaving parts of the d i f f e r e n t predicate with the sum predicates in the goal - but this obscures the program and is still not as efficient as the delaying solution.  Using such techniques can speed up programs by two or three orders of magnitude. I have observed programs dropping from an hour to a minute or even a second.  193  Even better speed-ups are possible by using constraints because integer arithmetic is complete (at least, Pressburger arithmetic is complete). Using constraints, the problem could be solved by: money(S, E, N, D, M, 0, M * 0,  R, Y) :-  E, N, D, M, 0 ,  different([S,  R,  Y]),  S * 0,  0 is  ( S * 1 0 0 0 + E * 1 0 0 + N*10 + D +  M*1000 + 0*100 + R*10 + E)  - (M*10000 + 0*1000 + N*100 + E*10 + Y ) .  where different is as above. The xpPAM implementation cannot currently handle such constraints; it would need a generator for the values of S, E, the  9.4  digit  Y, such as  predicate given earlier.  Pseudo-parallelism  The structure of a typical parallel program is:  Requester  Server  initialize send request  loop wait for message  wait for reply  process message send reply  process reply  endloop  This is a producer-consumer coroutine, with message passing instead of coroutine calls. No matter how much parallelism is available, the speed of the 194  two processes is limited by the slower of the two. Using x p P r o l o g , the coroutining process is: c o r o u t i n e ( C , P) :- consume(L, C ) , p r o d u c e ( L , P ) . p r o d u c e ( H d . T l , P) :- makeOneMsg(Hd, P ) , produce(Tl, P ) . p r o d u c e ( [ ] , P) :- n o t e x i s t s Hd s u c h t h a t makeOneMsg(Hd, P ) . /* end o f produced l i s t */ consume(Hd?.Tl, C) :- /* d e l a y u n t i l l i s t element and Hd i n s t a n t i a t e d */ processOneMsg(Hd, C ) , consume(Tl, C ) . consume([]?, C ) . % s t o p a t end o f l i s t  The parameters C and P represent information which is used to control the consumption and production.  Consume  immediately delays on  makeOneMsg makeOneMsg.  L.  Produce  which instantiates Hd, wakens ProcessOneMsg  starts and continues with consume  then executes.  and suspends  Consume  repeats and delays on  the uninstantiated tail of the list - this returns to the suspended produce and the cycle continues. The coroutine halts when  makeOneMsg  fails to produce a  new element. Many kinds of control can be used; here, C and P are used to represent such control.  This program has one slight inefficiency. as it constructs a list element but  Produce  consume  will then immediately delay on the  head. More efficient, but logically equivalent: p r o d u c e ( L i s t , P) :- makeOneMsg(Hd, P ) , L i s t = Hd.Tl, produce(Tl, P ) .  195  will wake up consume as soon  The order of the  consume  and  produce  goals in coroutine is important for the  interleaved execution; it is not important for the actual meaning of the program (we could have produce generate the entire list  L  first, then have  consume  process it - but this would use up enormous amounts of memory if the list were very large). The list L needs to exist just one element at a time because it is used solely as a communication channel. If reference counting is used, the list cells are freed as soon as they are accessed and so the entire list never exists, just the current element.  The  produce  and  consume  predicates are both deterministic and tail recursive, so  the compiler can easily turn them into iterations. If makeOneMsg  are also deterministic, then the execution stack is bounded.  processOneMsg  9.5  and  Correctness and completeness  The order of goal selection does not affect the correctness of the computation [Lloyd  1984].  However, it may affect the completeness. The programmer will  want to be assured that any program which correctly terminates for conventional Prolog will also correctly terminate with  As long as there are no delays,  xpProlog  xpProlog.  chooses goals in exactly the same way  as conventional Prolog. For such programs, both languages have the same partial completeness. Consider the simple query: ?- x  < Y,  x  =  l,  x =  2.  196  Depending on the Prolog implementation, this will either generate a run-time error (because " <" can only work with ground arguments) or it will fail (because " <" requires numbers as its arguments and uninstantiated variables are not numbers). Putting this query to x p P r o l o g results in successful completion. Thus,  xpProlog  is "more complete" than conventional Prolog, if the  delays are used properly.  Unfortunately, the programmer can take a working conventional Prolog program and add delay annotations so that it will not work in  •  xpProlog:  The delays may be too strict so that execution halts with some predicates still delayed. A trivial example, the query ?- x  =  l.  will succeed but the query ?- x?  =  l.  will delay indefinitely. •  The delays may cause an infinite loop. A trivial example is the query ?-  X < Y,  inf(A),  X =  1,  Y =  2.  where inf is a predicate which does not terminate (for example, " i n f ( A ) : - i n f (A)").  This query will fail, as described earlier, for  conventional Prolog but will go into an infinite loop with  xpProlog  the "X<Y" will delay, allowing the inf predicate go start execution. 197  because  To summarize, any pure Prolog program will produce the same results in xpProlog.  Some pure Prolog programs will terminate correctly only in x p P r o l o g  — they will either produce wrong answers or will fail to terminate in conventional Prolog. However, the programmer must be careful in adding delays to programs - it is possible to cause an over-constrained program to deliver a "don't know" answer with a list of goals which cannot be computed:  •  the goal(s) might succeed.  •  the goal(s) might fail, allowing alternative goals to be tried (resulting in success, failure or a further "don't know").  9.6  Parallelism  Many problems which appear to be good candidates for parallelism and are written using parallel constructs are really disguised coroutines because critical sections must execute sequentially. A digital telephone switch has many thousands of processes (telephone calls) running concurrently. Most of these are suspended, waiting for an event such as someone picking up or putting down the telephone. The few active processes must all pass through one bottleneck: the circuit assignment process. Circuit assignment must be done sequentially and the capacity of the switch is determined by the speed of the circuit assignment algorithm.  This does not mean that some specialized problems cannot benefit from massive parallelism. Examples are equation solving by using relaxation techniques or searching large databases. These problems are characterized by needing only localized communication among processes: 198  Or-parallelism:  the problem results in a large set of simple solutions which can  all be tested independently. An example of this is a dictionary search for synonyms; another example is parsing with an ambiguous grammar. Generate and tests are not good candidates for or-parallelism because of their enormous search spaces; coroutining offers much greater speed-ups because it automatically prunes the search space. And-parallelism:  the problem can be broken down into two sub-problems which  are completely independent. The overhead of detecting independent and-parallelism is about 10% [Hermenegildo and Nasr 1986], so parallelism results in significant speed-ups.  In fact, the telephone switch does gain some speed by using parallelism: peripheral processors can do some autonomous event processing. When the peripherals communicate the central call-processing CPU, they actually become part of a large coroutining system. For the central CPU, the advantages of parallel processing are not speed, but clarity. Some parallel designs have retained standard Prolog with full backtracking [Hermenegildo and Nasr 1986]. Their intent is to retain the semantics of conventional Prolog, speeding execution when parallelism can be exploited. As xpProlog  provides full pure Prolog with coroutining, it can easily be used in  such a parallel machine — the coroutining predicates could transparently be executed on a fully parallel machine.  In contrast, guarded Horn clause languages (such as GHC [Ueda 1985], Concurrent Prolog [Shapiro 1983] and Parlog [Clark and Gregory 1984]) have abolished backtracking. Flat GHC has even discriminated against user 199  predicates by not allowing them in guards. Standard Prolog can be implemented in such languages - but that can be said of even Fortran. Rather (attribution uncertain): flat, safe, concurrent guarded Horn clauses = Occam + logical variable.  I believe that the indeterminacy of the guards - which has caused much semantic difficulty - is not a very valuable feature because the guards are usually mutually exclusive and can be transformed into an equally fast (or faster!) sequence of if-then-elses.  The valuable feature of these languages is their  ability to execute predicates in parallel, communicating via shared logical variables — a feature which can be exploited in any logic programming language which does not depend on execution order. XpProlog's design for coroutining does not rule out parallelism. Because x p P r o l o g allows only purely logical constructs, goals can be executed in any  order and a coroutining program can be executed on parallel hardware with no changes. It may run faster - very often it will run slower because the cost of creating and destroying processes is greater than the saving producing by running in parallel.  9.7  Comparison with other designs for delaying.  MU-Prolog &refNaish 1985d] has a slightly less general form of delaying - it can delay only on entire arguments, not on the individual parts of them. In xpProlog's notation, MU-Prolog can do (Hd.Tl)? but not Hd?.Tl (this is fixed in  the successor, NU-Prolog). [Naish 1984] describes how to automatically generate wait declarations which are similar to but weaker than 200  proceed  declarations. His algorithm detects situations where backtracking would produce infinite solution spaces and then generates wait declarations to prevent them. The hamming program given earlier was originally tested in MU-Prolog; it will not run deterministically in MU-Prolog, even if some rather strange tricks are tried (supplied to me by Naish), because MU-Prolog has only the equivalent of simple top-level "?" annotations (because the uninstantiated, the  less-than  called again - when  A  A  and  B  in  predicate simply delays, allowing  and B become instantiated, the  may be  ordMerge  ordMerge  less-than  to be  may fail,  causing backtracking).  Prolog-II [Colmerauer 1982] uses a different method of delaying: the  freeze  ( g e l e r ) predicate. This delays calling a goal until a particular variable has  become instantiated. The test for delaying is thereby associated with the calling predicate instead of with the called goal. XpProlog's method has the advantage of requiring the delay declarations in only one place; the extra flexibility of Prolog-II's method is not needed in practice. If desired,  freeze  can be defined in  xpProlog: f r e e z e ( X ? , Pred) :- Pred. /* wait until X becomes instantiated, then call Pred  */  Prolog-II can delay until one of several variables becomes instantiated ("or-delay") by a somewhat awkward - and not obvious - construct: o r F r e e z e ( X , Y, P r e d ) :- f r e e z e ( X , C o n t r o l = c ) , freeze(Y, Control = c ) , freeze(C, Pred).  201  The first freeze delays until X becomes instantiated; the second freeze delays until Y becomes instantiated. As soon as either of these variables becomes instantiated, the goal C o n t r o l = c is executed which wakes up the  freeze  for Pred.  Unfortunately, this leaves unevaluated predicates lying around, so that we cannot distinguish between a program which is "hung" or permanently delayed and a program where some computational alternatives are not needed. XpProlog  can do a cleaner  or-freeze:  ?- p r o c e e d o r F r e e z e ( x ? , y, p r e d ) . ?- p r o c e e d o r F r e e z e ( x , y ? , p r e d ) o r F r e e z e ( X , Y, P r e d ) :- Pred.  v  Freeze  is very unwieldy for the  every call to ordMerge(X,Y,M)  ordMerge  predicate given earlier. Using  freeze,  is replaced by  f r e e z e ( X , f r e e z e ( Y , ordM(X, Y, M ) ) ) .  where ordM is defined by: o r d M Q ] , Y, Y ) . ordM(X, [ ] , X ) . ordM(X.A, Y.B, M) :- f r e e z e ( X , f r e e z e ( Y , ordMerge(X, Y, M ) ) ) .  IC-Prolog [Clark, McCabe and Gregory 1982] has similar facilities as ours, adding unsynchronized parallel evaluation and parallelism with directed communication. Most practical parallel programs can be easily simulated with coroutines (see section 9.4, "Pseudo-parallelism" on page 194) so I do not consider my design to be deficient in this respect. There is no need for x p P r o l o g to be implemented by coroutining - full parallel processing could be used. 202  IC-Prolog requires duplicating the bodies of predicates which can delay in more than one way like  append  or  ancestor.  It uses a variety of control annotations in  the code (-, //,!,?, [, ]) whereas x p P r o l o g has a single separate control annotation (proceed). But these are not sufficient to handle the test-and-generate solution for eight queens (below). [Kluzniak 1981] has an interesting variant which seems to be more powerful than IC-Prolog's. He defines three control predicates:  spawn Call inserts a new process, in suspended state, immediate before the current process.  wait Variable delays until the variable becomes instantiated. The suspended process before the current process is activated. yield  suspends the current process and passes control to the next process.  Using this, the eight queens program is written (with "X?" being an abbreviation for walt(X)):  203  queens(X) :- spawn perm([1,2,3,4,5,6,7,8], X ) , s a f e ( X ? ) perm([], [ ] ) . perm(X.Y, U.V) :- d e l ( U , X.Y, Z ) , y i e l d , perm(Z, V ) . d e l ( A , A.L, L ) . d e l ( X , A.L, A.R) :- d e l ( X , L, R ) . safe([]?). s a f e ( X ? . Y ? ) :- spawn n o d i a g ( X , 1, Y ) , s a f e ( Y ) . n o d i a g ( _ , _, [ ] ? ) . n o d i a g ( B ? , D?, (N.L)?) :- B =\= N-B, D =\= B-N, D l i s D+l, y i e l d , nodiag(B, D l , L ) .  Kluzniak's wait is the same as xpProlog's ? The spawn is not needed in x p P r o l o g because any predicate is potentially suspendable. X p P r o l o g would  write the queens predicate's goals in the order safe, perm (that is, the safe test is implicitly spawned because it immediately delays, waiting for the first element of the permutation to be generated). X p P r o l o g does not require the yields. As soon as a predicate instantiates a variable which has caused another predicate to delay (by a wait), the delayed predicate is activated. Kluzniak's design permits greater control but at the cost of requiring the programmer to know that the code will be used for coroutining (for example, perm must have a yield added to it for coroutining). Kluzniak also notes that coroutining between non-adjacent goals is difficult and he is unsure of appropriate backtracking behaviour (the details of xpProlog's backtracking are in section 4.0, "Backtracking and delaying" on page 66). The x p P r o l o g solution to eight queens requires putting the safe goal before the perm goal, removing all the "?"s, spawns and yields and adding the following  (which were generated automatically, using Naish's program): 204  ?-  proceed perm(?,  -).  ?-  proceed d e l ( - ,  ?-  proceed  ?,  ?-  proceed nodiag(-,  -).  ?-  proceed perm(-,  ?-  proceed d e l ( - ,  ?). -,  ?).  safe(?). -,  ?).  These declarations make sure that the testing predicates {safe and nodiag) will not attempt to construct the solution list and the generating predicates  {perm  and del) will not go into an infinite loop on backtracking, generating longer and longer lists. The X=\ X2  = expr  predicate is an abbreviation for  X2 i s expr, X  — it will delay until its arguments are instantiated. Incidentally,  #  nodiag  should not be "optimized" to delay until /V becomes instantiated. With the proceed  declaration above,  nodiag  will spawn a series of inequalities - the  optimization would prevent these from being spawned. In predicates which produce something (such as ordMerge)  should have the delays propagated to the  head but those which test should not.  In summary,  xpProlog's  "?" and proceed can do everything and more that the  other delaying designs can do, but with simpler notation.  205  10.0 Extensions to Horn logic  ... but exhaust the realm of the  — Pinder, Pythian  possible.  Odes, III.  109  Although Horn clauses are sufficient to handle any first order logic, they are not very convenient for some programming situations. In particular, the following have proven useful: •  Negation and  •  All-solution predicates {setof and  •  "Higher order" predicates.  •  "Infinite" structures.  •  Dynamic creation of predicates (meta-programming).  if-then-else. bagof).  Negation and all-solution predicates are treated incorrectly by most Prolog implementations. .Sections 8.9, "All solutions predicates" on page 178 and 10.4, "Meta-variables" on page 212 describe language extensions for handling these properly.  Coroutining (discussed in section 9.0, "Coroutining, pseudo-parallelism and parallelism" on page 185) is not an extension to the logic but an extension to 206  the control strategy. Arrays and logical I/O (discussed in section 11.0, "Arrays and I/O done logically and efficiently" on page 222) are  conservative,  extensions  to the logic in that they can be implemented using the existing mechanisms of the language. Mode and type declarations and equality theories (discussed in sections 5.10, "Speeding up deterministic predicates - modes and types" on page 112 and 6.5, "Equality: "is" and " = "" on page 133) can also be treated as special predicates which are treated in a special way by the compiler.  10.1  Negation  Horn clause logic does not contain negation. But negation is very useful in practical programming. This section will discuss how negation can be handled properly in a logic programming language. Some of the material overlaps the section on execution order (see section 8.0, "Execution order" on page 158). Consider the  member  predicate:  member(X, X . _ ) . member(X, Y.Rest) :- member(X, R e s t ) .  There are three ways that not member(X,List) can be processed:  1. Delay until everything in X and member's  List  are ground; call member; invert  success or failure.  2. Execute member, in a special "negation" mode, with a list of variables which must not become instantiated (the predicate will delay rather than instantiate one of these variables). 3. Transform  member  to a new  notMember  207  and call that.  There is one disadvantage in the first approach: the query not memberd, [o, l , X ] ) will delay unnecessarily. Furthermore, the delay is detected  by recursively traversing all the arguments to member  -  this requires overhead  similar to that required by the occurs check. The second approach adds significant complexity to the abstract machine. Transforming  member  gets  around the problem.  Here is a transformation of member to its negative equivalent any  X  notMember.  For  and List, exactly one of member or notMember is true.  notMember(X, T a i l ) :- T a i l ¥= (_._). /* o r T a i l = [ ] */ notMember(X, Y.Rest) :- X * Y, notMember(X, R e s t ) .  Here, the  not equals ( ' V )  predicate does item by item delaying. When  notMember(l, [o, 1,X]) is tried, the first clause fails (because the second  argument is a list), so the second clause is tried. This succeeds, resulting in trying n o t M e m b e r d , [ l,x]). Again, the first clause fails. The second clause also fails, so the entire query fails, which is what we wanted. In contrast, using general not would have delayed until X became instantiated, resulting in unnecessary computation, or even never being tried if X never becomes instantiated.  This transformation can be done automatically. The method is:  1. Transform the predicate to an if-and-only-if  form (using the  [Clark 1978]). 2. Put in all universal and existential quantifiers. 3. Negate the predicate. 208  completion  4. Work all the negations through to the innermost level so that not appears only in front of single predicates. 5. Transform the predicate back to Horn clause form.  Here is the transformation for  1. member(X,List) List=  iff  Y.Rest & (X=  2. forA\l(X,List):  Y\  forA\\(A,B):  iff  Y.Rest & (X=Y\  notMember(X,List)  not exists^Y.Rest):  4. forA\l(X,List):  member(X,List))  member(X,List)  exists(Y,Rest): List-  3. torAVL(X,List):  member.  List-  member(X,List))  iff  Y.Rest & (X=  notMember(X,List)  Y|  member(X,List))  iff  X ^ A.B  | existsf Y,Rest):  5. notMember(X, L i s t )  List=  Y.Rest & X *  Y&  notMember(X.Rest)  :- L i s t # ( _ . _ ) .  notMember(X, Y.Rest) :- X ^ Y, notMember(X, R e s t ) .  In a typed system, a further transformation is possible because  "List^_._"  is  the same as " L i s t = [ ]" However, there is a subtle difference between the two because the inequality will delay if it encounters an uninstantiated variable but the equality will instantiate a variable to nil.  209  Such transformations can be done at run time. XpPAM allows a "compile on first use" so that a predicate is stored in some external form until it is first tried. At that point, it is compiled (the external form is kept). Similarly, if the predicate is tried in a negative context, it can be transformed and then compiled.  Another situation where negation arises is when there is an either-or situation. For example, let us suppose that a fungus is always either a mushroom or a toadstool and never both. In sequent calculus (that is, Horn logic allowing multiple "goals" on the left hand side), this can be stated: mushroom(X),  t o a d s t o o l ( X ) :- f u n g u s ( X ) .  From this we can derive two extended Horn predicates which state the situation, although without quite as much computational power because of the lack of general resolution: mushroom(X) :- f u n g u s ( X ) , not t o a d s t o o l ( X ) . t o a d s t o o l ( X ) :- f u n g u s ( X ) , not mushroom(X).  10.2  Closed predicates  The transformation of a negated predicate requires using the  closure  of the  predicate [Clark 1978]. The assumption is that the programmer has written the program using if-and-only-if  ifs  (":-" or " <-") but has really meant  if-and-only-ifs.  The  form is produced by simply or-ing the clauses for the predicate  (step 1 in the previous section).  210  The negation transformation cannot be done in the presence of assert or on the predicate. To indicate this, x p P r o l o g has a  closed  meta-predicate which  states that the predicate will not be subject to any more changes (by retract).  retract  assert  or  An attempt to negate a non-closed predicate results in a run-time  error.  Similar to closed predicates are  cannot fail  predicates. For such a predicate,  not  should always fail. When x p P r o l o g compiles a cannot fail predicate, it adds one more "catch-all" clause which generates an error message. Many predicates have this cannot fail property — failure for these predicates indicates a programming mistake.  10.3 Setof, bagof All solution predicates are often needed in practical logic programming. They are clearly second-order constructs outside of Horn clause logic and should not be simulated by using non-logical predicates such as assert and  retract.  All  solutions predicates should be built into the logic programming language. All solution predicates are general cases of negation and single solution predicates. For example, not could be defined as succeeding when the solution list is empty: n o t ( T e s t ) :- b a g o f ( X , T e s t , [ ] ) . /* n o t h i n g succeeded */  Similarly, predicate memberl which succeeds only once, for the first match can be defined: 211  memberl(X, L i s t ) :b a g o f ( X , member(X, L i s t ) , X . _ ) . /* f i r s t s o l u t i o n o n l y */  Because all solutions predicates are second order, they require special second order variables. [Naish 1985c] describes a notation which gives purely declarative readings to all solutions predicates. The notation is similar to xpProlog's exists notation for if-then-else.  He treats the problem as a control  issue, indicating that execution should delays until certain variables become instantiated (or execution may proceed in spite of certain variables not being instantiated). X p P r o l o g treats introduces second order variables to handle the situation. The next section discusses these  meta-variables.  Section 8.9, "All solutions predicates" on page 178 describes a sound implementation of all solution predicates.  10.4  Meta-variables  Conventional implementations of bagof and  setof  misuse Prolog's logical  variables. In the goal b a g o f ( P r o t o , P r e d , R e s u l t ) , the variables in " P r o t o " (prototype) and " P r e d " (predicate) are different from regular Prolog variables. They are really place holders: each time the predicate is evaluated, they are filled in with new logical variables. Such variables are common in higher-order predicates — I call them  meta-variables.  Meta-variables are limited in scope to  the formulas in which they occur. They are place holders in formulas.  In the expression s e t o f ( X , p ( X ) , L I ) , s e t o f ( X , q ( X ) , L 2 ) , ...  212  the two "X"s are distinct. The scope is obvious in all-solutions predicates, but there are other cases where the scope must be explicitly given.  Meta-variables also occur when predicates or formulas are manipulated. For example, a natural language query system takes sentence and transforms them into logical formulas (which look like predicates). At some later time, these formulas may be executed by filling in the variables. The input All men like  Mary.  might be transformed to the formula a l l ( X , man(X) -> l i k e s ( X ,  'Mary'))  which can be tested by the query ?- i f e x i s t s X s u c h t h a t man(X) & n o t l i k e s ( X ,  'Mary') t h e n f a i l  endif.  by using the transformations in [Lloyd and Topor 1984]. When the logical form was produced, it got a new uninstantiated variable for "X." However, this was not quite what was wanted: a new meta-variable should have been used. X p P r o l o g uses the meta predicate to create formulas containing meta-variables and unmeta to create predicates from formulas. For example, the following creates a logical formula in  Form,  turns it. into a predicate and  then executes it: m e t a ( [ X ] , a l l ( X , man(X) -> l i k e s ( X , 'Mary')), a l l ( N e t a V a r , Form)), unmeta([MetaVar], Form, [ V a r ] , P r e d ) , i f e x i s t s V a r s u c h t h a t c a l l ( P r e d ) then f a i l e n d i f .  Meta  looks like a predicate but it must be handled specially by the x p P r o l o g  compiler. The arguments are: 1. the list of meta-variables. 213  2. the formula (using meta-variables). 3. the resulting formula.  Unmeta's  arguments are:  1. the list of meta-variables. 2. the formula (containing meta-variables). 3. the resulting list of newly created variables. 4. the predicate with new logical variables substituted for the meta-variables in the formula. Meta-variables are a generalization of Naish's notation for all solution predicates. They clear up ambiguities with logical forms and get rid of all need for var or "=" predicates.  Meta-variables must be marked. For example, it is reasonable to ask for the set of all [/4,Aj" such that pred(X)  — there is no need to have A instantiated before  producing the set; A is not a meta-variable here but X is.  The notion of two kinds of variables (ordinary variables and meta-variables) opens up many issues in the design-of x p P r o l o g such as nice ways of marking them, scoping rules and automatically detecting them. These issues are beyond the scope of this report. I have taken the simple view that a variable is an ordinary variables unless:  •  it is defined by a  meta  or  unmeta  predicate; or  214  •  it appears in the prototype part of an all-solutions predicate and it has not been marked by an exists.  Thus, to get the set of all \_A,X] such that pred(X),  one must write:  e x i s t s A suchthat s e t o f ( [ A , X ] , Pred(X)).  which the compiler transforms to: meta([X], pred(X), MetaPred), e x i s t s A s u c h t h a t s e t o f ( [ A , X ] , MetaPred).  To ease the compiler's job, it is illegal to use a meta-variable twice within a clause: setof(X, P1(X), R l ) , setof(X, P2(X), R2).  would have to be written, substituting Y for the second X: setof(X, P1(X), R l ) , setof(Y, P2(Y), R2).  10.5  Constraints vs. delays  Delays are passive. For example, i n t e g e r ( X ) , 3 < X, X < 5.  will simply delay until X becomes instantiated. However, these goals could be used to deduce that X=4. Alternatively, some delayed goals can be inconsistent. For example, X > 3 & X < 5 can never be satisfied. Yet xpPAM will simply delay in these situation, only able to test whatever value that X might be given.  215  Delayed predicates can be considered as  constraints  in that they constrain the  range of values which an uninstantiated variable can be unified with. If we use set notation, the first example gives X being the set {Z | integer(Z)  &  3  <  {Z | X  X  <  5}.  >  3 &  Z  & Z XpPAM  < 5}; the second example gives keeps a list of all delayed predicates waiting  for a variable in a list pointed from the variable - this can be considered as being very similar to the description of the set, but in a purely passive manner. That is, x p P r o l o g merely tests for set membership instead of trying to built the set.  Constraints are easy to handle whenfinitesets are involved. When infinite sets are involved (such as the examples above), direct manipulation is impossible; the sets must be manipulated by their intensional meanings rather than by their extensional representations. This requires knowing additional theorems about the constraining predicates.  Multi-variable constraints are somewhat trickier to handle. For some predicates, these can still be processed. For example, X + 7  & X>0  Y = 25 & 2  integer(X)  & Y>0 has the two solutions X = 3, Y = 4 and X = 4, Y = 3.  Handling constraints is beyond the scope of this report. Constraints are very useful for solving certain kinds of puzzles which have arithmetic constraints (these puzzles may have practical uses, for example in helping compilers generate least cost code). It would be nice if a general mechanism could be found to allow specifying the theorems which pertain to constraining predicates. See [Jaffar and Lassez 1987] and [Voda 1986b] for further work on constraint programming. 216  Some problems are better solved by constraints and others are better solved by delays. Generally speaking, constraints are best for problems where a decision procedure can combine two or more constraints to produce a new constraint, such as solving inequalities (simplex programming, Diophantine equations, etc.). Delaying must be used for problems which cannot be handled this way. Also, delaying is useful for coroutining, which is a powerful program structuring tool but which can always be avoided by making more complex sequential programs.  10.6 The occurs check Most Prologs do not implement the "occurs" check. The reason is simple efficiency: the occurs check requires a complete traversal of both structures being unified, raising the cost of unification from  O(N)  to 0(N ). 2  [Plaisted  1984] gives some methods of detecting when the occurs check can be safely avoided, by doing global analysis of the predicates.  The occurs check is necessary for unification to work correctly. The simplest situation is "X=f ( X ) " which has no possible solution (if X is required to be finite). For example, consider append for difference lists: a p p e n d D i f f ( A - Z , Z-B, A-B).  The difference list can be turned into a regular list by a simple predicate, allowing appendD to be defined d i f f T o L i s t ( A - [ ] A ) . /* t u r n t a i l p o i n t e r i n t o [] */ appendD(A,B,C) :- a p p e n d D i f f ( A , B, Z ) , d i f f T o L i s t ( Z , C ) .  This produces the following results:  217  appendD(l.A-A, 2.3.B-B, C) = > C = [ 1 , 2 , 3 ] . appendD(A-A, X, X) ==> X = [ ] appendD(1.2.A-A, X, X) = > X = [1,2,1,2,1, ...]  Yet the last result is obviously wrong: it implies (by comparison with the second result) that [ ]=1.2. Often, the occurs check is quite cheap because most unifications can be transformed into simple assignments, equality tests, splitting up list elements or creating new list elements. The occurs check is needed when general structures are unified, for example the difference lists above. The cost of unification without the occurs check is O(N) whereas with the occurs check it is difference list for delays in  0(N ). 2  append  not  It should be noted that the occurs check makes the  work in the same time as the ordinary append.  or setof also multiplies the cost by O(N)  Checking  — it is desirable to  eliminate all such costs. The occurs check can be left out because in practice very few predicates require it; the cost of delaying with negation can be made acceptable by the technique discussed earlier (section 10.1, "Negation" on page 207).  Prolog-II takes a different approach to the occurs check by allowing "infinite" structures called "rational tree": x=f(X) X=1.X  => = >  x=f(f(...(...)...)) X=[ 1,1,1,...]  This neatly removes the need for the occurs check, at the expense of changing the logic (the Herbrand universe is defined over finite structures). But there 218  seems to be no reason to rule them out, especially as we are adding second order predicates (setof and  bagof)  and we can also handle potentially infinite list by  the delaying mechanism. Indeed, infinite lists are useful for some kinds of grammar notation. However, they do complicate the implementation: output predicates must check for infinite structures and reference counting may not be able to collect all garbage.  10.7  Assert and retract  There are two main uses for  assert  and  retract:  •  To create new predicates "on the fly."  •  To provide a database facility.  The first use is quite reasonable. For example, logic grammars (DCGs, etc.) are usually handled by being passed to a predicate which transforms the grammar notation to standard Prolog clauses, which are then added to the clause database. Such a use of  assert  is safe if it does not modify any predicate which  is executing.  Assert  can be used for meta-programming. As a simple example, logic  grammars can be processed by a translator (a logic program, of course) to produce a new logic program. This is completely safe if the translator simply outputs the new logic program which is subsequently read in and executed, independent of the translator. But this is inconvenient, so x p P r o l o g imposes no such restrictions; the user must be careful that an executing predicate is not  219  modified.  24  And  retract  being replaced using  XpProlog  is provided so that a predicate can be removed before  asserts.  implements assert by a "compile on first use" strategy. Whenever a  clause is added to (or removed from) a predicate, the compiled code for the predicate is thrown away and replaced by code which invokes the compiler. The next time the predicate is called, the compiler is invoked - the resulting code is then kept and executed.  The second use of assert and  retract  is poor. For one, thing, the clause database  is usually inefficient for general purpose database manipulations. Secondly, the predicates which do such operations rely on the side effects of the retract  assert  and  "predicates" — that is, such predicates are not declarative.  A correct logical treatment of databases is briefly discussed in section 11.7, "Databases" on page 240 which gives a completely logical use of databases in which backtracking will remove any changes to the database . This does not mean that a database cannot change with time - after a query is completed, any changes to the databases (and other files) are made permanent.  XpPAM  also provides a simple kind of associative table, defined by the predicates  tableAssign,  tableFind  and  tableDelete  (the details are given in section  Appendix D, "Built-in predicates" on page 300). These can have a declarative  24  In fact, the abstract machine is likely to go wildly wrong if an executing predicate is modified because it assumes that a predicate will not change while it is executing.  220  reading such that changes are undone on backtracking. Associative tables efficient internal databases.  221  11.0 Arrays and I/O done logically and efficiently  The simplest and most natural way to keep a linear list inside a computer is to put the list items in sequential locations, one node after the other.  ... This technique for representing  linear list is so obvious and well-known no need to dwell on it at any length. understand  the limitations  a  that there seems to be ... It is important  to  as well as the power of the use of sequential  allocation.  - [Knuth 1973] This chapter can be read as a separate essay. It describes how arrays can be added to a logic programming language using only logical constructs and how this can be done efficiently. The discussion of arrays is as an example of how, by applying a little thought, a language designer can both please the purists and the pragmatists. The purists want constructs which are purely logical; the pragmatists want constructs which are expressive and efficient.  222  11.1  Introduction  Some logic programmers believe in using "pure" logic programming regardless of its efficiency; some others believe in efficiency regardless of its affect on the declarative reading of programs. Efficiency should not be an over-riding concern when writing programs, especially in a "very high level language"; nor can efficiency be completely ignored. X p P r o l o g walks a middle path, allowing the programmer to pick appropriate data structures without having niggling worries about the efficiency of the resulting program. The idea is similar to that of the SETL project [Schwartz 1975] in which the programmer picks logical data structures which are natural for the problem and the compiler optimizes the data structures into particular implementations - the declarative reading of programs is retained while attaining acceptable efficiency.  Poorly thought out attempts to increase the efficiency of logic programs can actually backfire. Some programs can be sped up by a factor of 100 or more if a more flexible execution order is allowed (section 9.2, "Coroutining example" on page 187). Such optimizations can be used only if the programs are written in a purely logical (or declarative) style. When non-logical constructs are added to logic programs, these optimizations are not possible — the difficulties are similar to those encountered by optimizers for conventional programming languages when confronted with aliases caused by unrestrained use of pointers. In logic programming, the problem specification is the program. Clarity of specification is important but efficiency must not be forgotten. Conventional programs often attain efficiency by using arrays and destructive assignments. These facilities are not generally available in logic programming languages 223  destructive assignments are impossible within the declarative style of logic programming.  Some Prolog programs emulate arrays using lists or trees (see, for example, [Kluzniak and Szpakowiez 1985, pp. 113-120]). At best, these provide 0( log /V) access to arbitrary elements (/V is the number of elements in the simulated array) whereas true arrays provide constant time (0(1)) access.  25  However, arrays cannot always emulate lists. Arrays always have a known length but lists may be of indeterminate length if the tail of the last element is not instantiated.  I will present some array manipulation predicates which have obvious declarative meaning. These predicates can be implemented efficiently. Automatic transformations can take advantage of situations where destructive assignment can be used.  In most current logic programming languages, I/O is implemented by highly non-logical predicates. Many people have suggested that cut can be eliminated from logic programming languages. If var and "==" are also eliminated (by some form of delay mechanism or by meta-variables), the only remaining non-logical predicates are those which do I/O. The logical treatment of arrays can be extended to files and databases, removing the need for any non-logical predicates. The only difference between arrays and files is where the data are  2 5  This approach may not be entirely invalid: [Wise 1987]  provides an argument that for certain  numerical problems, trees are adequate. However, arrays have uses beyond matrix algebra and for these, constant time access appears to be crucial.  224  stored. Techniques which give efficient implementation of arrays also give efficient implementation of I/O.  With a rich choice of data structures and a syntax which supports them, programs can be written clearly, concisely and logically.  11.2 Array Operations I will show the operations only for 1-dimensional arrays (vectors); extension to multidimensional arrays are "left as an exercise for the reader."  The elements of an array do not all need to be the same and uninstantiated variables are allowed inside arrays. Space and time efficiency can be improved if the compiler knows that all the elements will be the same, without any uninstantiated variables. One particularly useful form of array is a string. A string is just an array of single characters. The operations which can be applied to a string are the same as those for an array. The typing predicate s t r i n g ( S ) has no effect on the logical reading of a program; it merely increases efficiency by guaranteeing that all elements of the array are single characters (sec section 5.10, "Speeding up deterministic predicates - modes and types" on page 112). Putting an uninstantiated variable into a string causes a delay and putting in anything else causes a run-time error.  First, the basic operations:  225  bounds(A, L , H ) The array typing predicate. L is the low bound of A and H is  the high bound. This predicate can either get the bounds of an array or create an array of the given size (with all uninstantiated elements). The high bound of an empty array is one less than the low bound this follows from the definition of length (predicate len below) which is high-bound—low-bounds  1.  eIem(A, I, X) Extract one element: X is the Ith element of A. That is, X is A[I]. chElem(A, I, X, A2) Change one element. A2 is the same as A except that the Ith element has been set toX, as if A2 := A, A2[l] := X.  These are sufficient to define all other array predicates.  For notational compactness, expressions may exist inside predicate calls. An expression is enclosed in [ ... } which is read "evaluate."  26  The expression is  evaluated by using the " i s " predicate (":=" in Waterloo or IBM Prolog). Thus factorial(N,  {N .* F } ) :- f a c t o r i a l ( { N - 1}, F ) .  has the same meaning as f a c t o r i a l ( N , NF) :- N2 i s N - 1, factorial(N2, F), NF i s N * F.  Some array operators can be used in " i s " expressions:  27  2 6  There is no ambiguity with grammar notation's use of curly brackets — expressions can occur only within goals whereas grammar rule curly brackets can occur only around goals.  2  ?  These are just extensions to the "built-in" arithmetic operations which could be defined  226  L  is  lob A : -  bounds(A,  H is  hib A : -  bounds(A, _ ,  0  is  len  L  is  len A : -  len(A,  X is  A[I]  :-  elem(A,  X is  first  A :-  X is  last  L,  _).  /*  low b o u n d * /  H).  /*  high  bound * /  [].  A  :-  L).  /*  I,  " l e n " defined later  X).  elem(A,  (lob  A},  X).  elem(A,  [hib A ) ,  X).  A2  is  A rangeFrom  A2  is  A rangeTo  I  A2  is  rangeFromSecond A : -  I  */  :-  rangeFrom(A,  :-  rangeTo(A,  I,  I,  A2).  A2).  rangeFrom(A,  [lob  /*  "rangeFrom":  /*  "rangeTo":  A +  1},  below  below  */  */  A2).  Following are some standard array predicates. These are built-in for efficiency, but they can all be implemented with just  bounds, elem  and  chElem  Many other  array predicates are possible. I have restricted myself to a list similar to those in [Dijkstra 1976].  swapElem(A, I, J, A2) Swap two elements. A2 is the same as A except that the Ith and Jth elements have been swapped: A2  :=  A, A2[I]  :=  A[J],  A2[J]  :=  A[I]  or, using c h E l e m : chElem(A,  [A[J]},  X is  X?  X is  A+B  :-  A2 i s  A , B2  X is  -A  :-  A2 i s  A,  etc. where  :-  I,  chElem(Al, J,  [A[I]},  A2)  atomic(X).  arilh  instantiated.  Al),  is  B,  arith(+,  A2?,  arith(-,  A2?,  B2?,  X2),  X=X2.  X2),  X=X2.  is a low-level predicate which expects all its arguments except the last to be  Note the usage of "?"s within  arilh  everything is sufficiently instantiated.  227  which delay the entire clause until  concat(Al, A2, A) Concatenate two arrays. A is the concatenation of. Aland A2. After this operation: bounds(A, { l o b ( A l ) } , { h i b ( A l ) l e n ( A 2 ) } ) . Concat  can be used with thefirsttwo parameters uninstantiated, in  which case it generates sub-arrays (or sub-strings) by backtracking. subrange(A, L , H , A2) Extract a subrange of an array. A2 contains the subrange of A delimited by L and H . If H is less than L , then A2 is unified with an empty array. IoEx(X, A, A2) Low extend an array. A2 is the same as A except that the low bound has been decreased by one and the new element is X. After this operation: {lob {hib  A2) A2]  = {lob A - 1}, = { h l b A},  subrange(A2,  { l o b A}, { h l b A j , A ) ,  {A2[lob A2]} = X  hiEx(A, X, A2) High extend an array. A2 is the same as A except that the high bound has been increased by one and the new element is X. After this operation: {lob {hib  A2} A2]  = { l o b A}, = {hib A + 1},  subrange(A2,  { l o b A}, { h i b A}, A ) ,  {A2[hib A2]} = X  reshape(A, L , H , A2) Reshape an array with new bounds. A2 contains the same elements as A but with different bounds defined by L and H . An error occurs (and the predicate fails) if A and A2 have different numbers of elements.  228  For convenience, the following are defined: low bound l o b ( A , L ) :- bounds(A, L, _ ) .  high bound h i b ( A , H) :- bounds(A, _, H). length  l e n ( A , {hib A - l o b A + 1}).  range from rangeFrom(A, L, A2) :- subrange(A, L, {hib A}, A 2 ) . range to  rangeTo(A, H, A2) :- subrange(A, { l o b A}, A 2 ) .  low remove one element loRem(A, {A[lob A ] } , {A rangefrom l o b A + 1 j ).  high remove one element hiRem(A, { A f h i b A]}, {A rangeTo h i b A - 1}). shift  s h i f t ( A , N, A2) :- reshape(A, { l o b A + N}, { h i b A + N}, A 2 ) .  An array predicate fails if an index falls outside the range of the bounds. A run-time error may be preferable to simply failing because an out of bounds condition usually suggests an error in program logic.  Array constants are defined by specifying either the lower or the upper bound, plus a list of all the elements (because they are declarative, these predicates also convert arrays to lists): arrayFrom(A, L, [ X I , X2, a r r a y T o (A, H, [ X I , X2,  XN]) /* a r r a y "A" w i t h low bound "L" */ XN]) /* a r r a y "A" w i t h h i g h bound "H" */  These predicates can be thought of as extensions of the standard "name" and "=.." (univ)  predicates.  The empty array is denoted [ ].  229  11.3  Transformation to allow destructive assignment  As a simple example, consider a predicate which increments each element in a list:  28  incrList([], []). i n c r L i s t ( A . R s t , A2.Rst2) :- A2 i s A + 1, i n c r L i s t ( R s t , Rst2).  For arrays, this is: i n c r A r r a y ( A , A2) :- bounds(A2, { l o b A}, { h i b A } ) , /* A2 has same shape as A */ i n c r A r r a y 2 ( A , { l o b A}, A 2 ) . i n c r A r r a y 2 ( A , I , A2) :- I > { h i b A}. i n c r A r r a y 2 ( A , I , A2) :- I =< { h i b A}, elem(A2, I , {A[I] + 1 } ) , /* A 2 [ I ] = A[I]+1 */ i n c r A r r a y 2 ( A , {1+1}, A 2 ) .  A general  incr  predicate is transformed to something like:  i n c r ( I n p u t , Output) :i f g r o u n d _ a r r a y ( I n p u t ) t h e n i n c r A r r a y ( P r e d , I n p u t , Output) e l s e i n c r L i s t ( P r e d , Input, Output).  2 8  This is a special case of the following Prolog idiom (which is much like the mapcar function in LISP):  map(Pred, [ ] , [ ] ) . map(Pred, X.Rest, Y.Rest2) :- Pred(X, Y ) , map(Pred, R e s t , R e s t 2 ) . If  Pred is  deterministic when its first argument is ground, the transformations in this section  can be applied.  230  The compiler can now generate the following Pascal-like code for the case when A is a ground array, by destructively modifying A: p r o c e d u r e i n c r A r r a y ( A : a r r a y , v a r A2 : a r r a y ) b e g i n i n t I; f o r I := l o b A t o h i b A do A [ I ] := A [ I ] + 1; A2 := p t r A; /* A2 p o i n t s t o ( m o d i f i e d ) A */ end p r o c e d u r e  Destructive assignment can only be used if A is not needed after  incrArray  is  called. Destructive assignment could not be used in the goal ?-  i n c r ( A , A2), p(A, A2).  because A is used by  p.  Instead, this must be transformed to (copy is defined  later): ?-  copy(A, A x ) , i n c r ( A x , A2), p(A, A2).  Destructive assignment is slightly more complex than in Pascal because of the possibility of backtracking. The Prolog abstract machine can record information for undoing the instantiation on the backtrack stack (called "trail" in the Warren Abstract Machine [Warren 1983]). If the value being instantiated (or, in this case, modified) is newer than the latest choice point, then nothing need be saved. If backtracking information must be recorded, the entire array A is not saved - only a pointer to A2 and the changed value. Backtracking can then restore the single changed value.  231  11.4  Efficient implementation of array operations  There is an extensive literature on methods for storing data in arrays (see, for example, [Knuth 1973]). Although any of these storage methods can be used, I will assume that an array is kept as a triple, containing the low bound, high bound and a pointer to the data. The data are kept in a contiguous vector. The elements are all pointers to other objects (including uninstantiated variables). Thus, any element can be reached by an ordinary array indexing operation in constant time.  The following predicates access elements of arrays and are very cheap: bounds(A, L, H)  (also l o b ( A , L ) , hib(A, H) and l e n ( A , L ) )  elem(A, I , X)  The following subrange predicates produce a new array which is some part of another array. They are also very cheap because they merely require creating a new object with different bounds, pointing to somewhere within the original array. subrange(A, L, H, A2)  (also rangeFrom(A, L, A2) and rangeTo(A, H,  A2)) loRem(A, X, A2) hiRem(A, X, A2) reshape(A, L, H, A2)  (also s h i f t ( A , - N, A2))  The following require changing array elements to create a new array: chElem(A, I , X, A2) swapElem(A, I , J , A2)  232  c o n c a t ( A l , A2, A) l o E x ( X , A, A2) h i E x ( A , X, A2)  These operations are cheap if they are done in-place (destructively) but can be expensive if they require array copying. LoEx usually requires an array copy but it is not a common operation because arrays are usually built using h i E x or concat.  HiEx  and concat can be efficiently implemented (as in the XPL  language [McKeeman, Horning and Wortman 1970]) by simply adding the new element(s) to the array if the heap space following the array has not been allocated. This is frequently the case because usually only one array is built up at a time. Thus, using h i E x to build an array is as efficient as pre-allocating the array (using bounds) and filling in the elements (using elem).  Array copying can be implemented: copy(A, A2) :- bounds(A2, ( l o b A}, { h i b A } ) , copy2(A, { l o b A}, A2). copy2(A, I , A2) :- I > [ h i b A}. copy2(A, I , A2) :- I =< { h i b A j , elem(A2, I , { A [ I ] ) copy2(A, {1+1}, A2). S  This is nothing more than a statement of what it means for two arrays to be equal; that is, A = A2 if the bounds are identical and the corresponding elements are identical.  Array copying should be avoided. But destructive assignment cannot be used everywhere because it can make programs non-declarative. For example, in Pascal: A [ l ] := 'a*;  A [ l ] := *b*; 233  is not the same as A[l]  V;  :=  A[l]  :=  'a';  as might be expected by a naive translation to Prolog (the "," ("and") operator in pure Prolog is semantically commutative). However, c h E l e m is declarative:. chElem(A,  1,  a,  A2),  chElem(A2,  1,  b,  A3)  chElem(A,  1,  a,  A2)  is clearly the same as chElem(A2,  1,  b,  A3),  if the Prolog implementation allows delaying goals  (chElem(A2, l,b,A3)  must be  initially delayed because A2 is as not yet instantiated).  The goals chElem(A,  1,  a,  A2),  p(A2)  can be replaced by destructive assignment A[l]  :=  a,  p(A)'  because the original array A is not used after it is modified.  29  If the original  array were needed later, a copy would be needed: copy(A,  A2),  A2[l]  :=  a,  p(A2),  q(A).  which is clearly equivalent to chElem(A,  1,  a,  A2),  p(A2),  It is also logically correct because  q(A). copy(A,  A2)  has the same meaning as A  =  A2.  Destructive assignment does not affect backtracking. In p(A),  29  chElem(A,  1,  a,  A2),  q(A2),  ...  More strictly, nothing else may point at A For example, if this were preceded by B=A and B were used later, then A could not be modified destructively.  234  if q fails and p has alternative clauses, p will be retried with the value of A before the destructive assignment. This is because destructive assignments to A (which create A2) are recorded on the backtrack stack and undone on backtracking.  In theory, it is possible to do a global analysis of all the predicates and determine which ones may use destructive assignments. Copy goals can then be inserted ahead of calls to these destructive predicates. Some of these copys may be unnecessary because a destructive predicate does not guarantee a destructive assignment, only the possibility. Moreover, the analysis is not simple because of subrange arrays. Some of the techniques of alias detection for conventional programming languages may be applicable. See [Bruynooghe 1986] and [Kluzniak 1987] for some preliminary work in this direction.  Another method of avoiding array copying is to associate a reference count with each array and make a copy only if the reference count is greater than one. The reference count must also include the counts of all subrange arrays. This technique ensures that array copying is done only if necessary. Some studies in [Krasner 1983] suggest that optimized reference counting has about the same overall cost as marking garbage collectors - by eliminating unnecessary array copying, reference counting may actually be more efficient than marking garbage collecting.  235  11.5  An example: Quick-sort  One of the favourite examples of logic programming is quick-sort for sorting a list: qsort(Unsorted, Sorted)  :- q s o r t x ( U n s o r t e d , S o r t e d , [ ] ) .  q s o r t x ( [ ] , Sorted, Sorted). q s o r t x ( P i v o t . R e s t , S o r t e d , SoFar) :- /* The "SoFar" parameter i s */ s p l i t ( P i v o t , R e s t , Lo, H i ) , /* used t o a v o i d c a l l i n g */ q s o r t x ( L o , S o r t e d , P i v o t . S o r t H i ) , /* "append". */ q s o r t x ( H i , S o r t H i , SoFar). split(Pivot, [], [],[]). s p l i t ( P i v o t , X.Rest, X.Lo, H i ) :- X =< P i v o t , s p l i t ( P i v o t , R e s t , Lo, H i ) . s p l i t ( P i v o t , X.Rest, Lo, X . H i ) :- X > P i v o t , s p l i t ( P i v o t , R e s t , Lo, H i ) .  These predicates are all deterministic if their parameters are sufficiently instantiated. With the tail recursion optimization (TRO [Warren 1986]), s p l i t can be transformed into iterative form. Yet, the resulting code is still far from what a "conventional" programmer would produce because he (or she) would sort not a list but an array and the sorting would be done in-place. Lists should be sorted using other techniques such as merge sort (in fact, if the keys are expensive to compare, merge sort may be better even for arrays).  Here is  quick-sort  as it would be written for arrays.  q s o r t ( U n s o r t e d , S o r t e d ) :q s o r t x ( U n s o r t e d , ( l o b U n s o r t e d j , {hib Unsorted}, q s o r t x ( S o r t e d , Lo, H i , S o r t e d ) :- Lo >= H i . q s o r t x ( U n s o r t e d , Lo, H i , S o r t e d ) :- Lo < H i , P i v o t = {Unsorted[Lo]}, 236  Sorted).  spl±t(Unsorted, Lo, H i , P i v o t , P I , S p l i t ) , q s o r t x 2 ( S p l i t , Lo, H i , Lo, P I , S o r t e d ) . q s o r t x 2 ( S p l i t , Lo, H i , PIO, P I , S o r t e d ) :- PI > H i , /* P i v o t v a l u e i s >= e v e r y t h i n g ; put a t end and reduce range s w a p E l e r a ( S p l i t , PIO, H i , S p l i t 2 ) , q s o r t x ( S p l i t 2 , Lo, {Hi-1}, S o r t e d ) . q s o r t x 2 ( S p l i t , Lo, H i , PIO, P I , S o r t e d ) :- PI =< H i , q s o r t x ( S p l i t , Lo, {PI-1}, S o r t 2 ) , qsortx(Sort2, PI, H i , Sorted).  */  s p l i t ( U n s o r t e d , Lo, H i , P i v o t , Lo, Unsorted) :- Lo > H i . s p l i t ( U n s o r t e d , Lo, H i , P i v o t , P I , S p l i t ) :Lo =< H i , [Unsorted[Lo]} =< P i v o t , s p l i t ( U n s o r t e d , {Lo+1}, H i , P i v o t , P I , S p l i t ) . s p l i t ( U n s o r t e d , Lo, H i , P i v o t , P I , S p l i t ) :Lo =< H i , {Unsorted[Lo] > P i v o t , s p l i t 2 ( U n s o r t e d , Lo, H i , P i v o t , P I , S p l i t ) . s p l i t 2 ( U n s o r t e d , Lo, H i , P i v o t , Lo, Unsorted) :- Lo > H i . s p l i t 2 ( U n s o r t e d , Lo, H i , P i v o t , P I , S p l i t ) :Lo =< H i , { U n s o r t e d f H i ] } > P i v o t , s p l i t 2 ( U n s o r t e d , Lo, {Hi-1}, P i v o t , P I , S p l i t ) . s p l i t 2 ( U n s o r t e d , Lo, H i , P i v o t , P I , S p l i t ) :Lo =< H i , {UnsortedfLo]} =< H i , swapElem(Unsorted, Lo, H i , U n s o r t e d 2 ) , s p l i t ( U n s o r t e d 2 , Lo, H i , P i v o t , P I , S p l i t ) .  This is a little longer than the algorithm for lists. It also implements a slightly different algorithm for s p l i t - the resulting split sub-arrays may have their elements in different orders than the sub-lists.  The array version can be easily modified to run more efficiently. A different method of choosing the pivot could pick a random member in constant time rather than the 0( log /V) time required for a list. Array lengths are always  237  known, so a different, more efficient sorting method can be used when the number of elements in a sub-array becomes small.  It is too much to expect an optimizing compiler to transform the list formulation of quick-sort into the array formulation, especially as the two algorithms are slightly different. However, the compiler can easily determine that s p l i t is deterministic if the first parameter is instantiated, resulting in: s p l i t ( U n s o r t e d , Lo, H i , P i v o t , Lo, S p l i t ) :i f Lo > H i t h e n S p l i t = U n s o r t e d e l s e i f (Unsorted[Lo]} =< P i v o t t h e n s p l i t ( U n s o r t e d , {Lo+1}, H i , P i v o t , P I , S p l i t ) . e l s e s p l i t 2 ( U n s o r t e d , Lo, H i , P i v o t , P I , S p l i t ) , endif. s p l i t 2 ( U n s o r t e d , Lo, H i , P i v o t , Lo, S p l i t ) :i f Lo > H i t h e n S p l i t = U n s o r t e d e l s e i f {Unsorted[Hi]} > P i v o t t h e n s p l i t 2 ( U n s o r t e d , Lo, {Hi-1}, P i v o t , P I , S p l i t ) else swapElem(Unsorted, Lo, H i , U n s o r t e d 2 ) , s p l i t ( U n s o r t e d 2 , Lo, H i , P i v o t , P I , S p l i t ) endif.  which can then be transformed by the tail recursion optimization into the more conventional iterative form.  238  11.6  Direct access I/O  Some conventional programming languages treat direct accessfileslike arrays. For example, XPL has the built-in pseudo-variable f i l e ( i , j) which maps to the jth record of the i t h file. X := f i l e ( i , j) results in a record being read into x; f i l e ( i , j ) := X results in a record being written from X.  The meaning of a program stays the same, regardless of whether the it uses direct access files or arrays; the sole difference is where the data are stored. A meta-predicate is used to associate afilewith an array name and no change is required to the program. Copying files is even more expensive than copying arrays, so ?-  destructive(A, A 2 ) ,  p ( A ) , ...  always results in a run-time error when A is accessed within p. This is not a desirable situation. However, all implementations have limitations (for example, available memory or stack depth). This restriction preserves the correctness of programs — if they succeed or fail, they will do so exactly the same as programs without the restriction but sometimes they will neither succeed nor fail but will halt with an implementation restriction error indication.  Undoing changes to a direct access file on backtracking is relatively cheap; the mechanism is similar to the method used to undo changes to arrays.  The implementation of direct accessfilesmay impose other restrictions on their use. For example, there may be a requirement that all elements in afilebe of the same type (this restriction can be relaxed with databases — see below). Also,filesusually do not handle the concept of uninstantiated variables. These 239  can be handled by keeping a separate list of output uninstantiated variables which get written when they become instantiated; a run-time error occurs if the program terminates with outstanding entries in this list.  11.7  Databases  A physical part of a database can be thought of as an array with non-integer indexes. The predicates which manipulate arrays can easily be extended to allow arbitrary indexes. Subrange predicates only make sense if the underlying index type can be ordered. Some additional predicates may be added, such as:  •  find all elements based on some search criteria  •  delete elements  •  update elements  • sort Using these predicates, a query language such as SQL can be easily implemented, if care is taken in avoiding the non-logical aspects of SQL. Databases are related to associative table, as direct I/O is related to arrays. Databases usually contain some notion of "commit" processing. The changes to the database become permanent only when a specific "commit" transaction makes them permanent — up to that point, the changes can be "backed out." This fits very closely with Prolog's execution strategy. When there is the possibility of backtracking, "commit" can not be done. As soon as there are no backtrack possibilities for the predicate which modified the database (conceptually, produced a "new" database), "commit" can be done. 240  Conventional Prolog confuses the notion of databases by using the retract  assert  and  predicates for both maintaining databases and for defining (or removing)  predicates. Databases can be implemented - inefficiently - by ordinary lists or functors; the definition of predicates lies outside the scope of logical inference (see also section 10.7, "Assert and retract" on page 219).  11.8  Sequential Input/Output  [Cheng 1986] has observed that sequential I/O streams act just like lists (related ideas are given in [Wilson 1985]) Both must be processed element by element and both are of indeterminate length. Programs which do sequential I/O are written just as if they are manipulating ordinary lists. To make a program process streamfilesinstead of lists, a meta-predicate is used to associate a file with a list. For input, whenever the unifier needs the head of this stream-list, it issues afileread. For output, whenever the unifier creates a new element in the list, it issues afilewrite.  Pascal uses a similar technique. A stream file is treated as a vector with a pointer to the current element. For logic programming, the stream pointer can be manipulated with input stream) and  loRem  hiEx  (to get individual elements from the beginning of an  (to add elements at the end of an output stream). The  high bound of a stream is unavailable - any predicate requesting the high bound or length of a stream delays until end-of-file is reached.  Again, there is no difference between programs which do I/O and those which manipulate list. A meta-predicate declares that a particular list is associated with a file. For example: 241  ?- i n s t r e a m ( i n f i l e , I n ) , o u t s t r e a m ( ' o u t f i l e , Out), q s o r t ( I n , Out). 1  1  1  will sort file "infile" into file "outfile" (this requires the list version of  quick-sort  rather than the array version). For convenience, the various standard files can be automatically added to predicates. p(A, B) :- p u t ( ' A = ' ) , d i s p l a y ( A ) , q(A, B ) , p u t ( ' => ' ) , d i s p l a y ( B ) , n l . q(A, B) :- put('<qqq>'), B i s A + 1. ?- P ( l , 2 ) .  would be transformed to p(A, B, S t d l n , S t d O u t R e s t , S t d l n 2 , StdOut) :StdOut = ('A = . A . Std0ut2), q(A, B, S t d l n , StdOut3, S t d l n 2 , S t d O u t 2 ) , stdOut3 = ( ' = > ' . B . '\n' . S t d O u t R e s t ) . q(A, B, S t d l n , StdOut, S t d l n , S t d 0 u t 2 ) :StdOut2 = ('<qqq>' . S t d O u t ) , B is A + 1 ?- i n s t r e a m ( ' < t e r m i n a l > , I n ) , o u t s t r e a m ( ' < t e r m i n a l > , O u t ) , p ( l , 2 , I n , Out, In2, [ ] ) . 1  1  1  This is similar to techniques used to transform grammar rules to Prolog. As with direct access files, I/O streams may impose some extra restrictions over lists. Because output may be to a non-erasable medium like a printer, an output stream has a buffer associated with it; backtracking can only undo values which 242  are still in the buffer. The f l u s h meta-predicate flushes everything in the buffer to the output device - a run-time error occurs if an attempt is made to flush an uninstantiated variable.  Flush  does not change the meaning of a program, nor  the final contents of a file, but it may cause a run-time error to occur in an otherwise correct program.  Input and output with streams is very similar to lazy evaluation. An input file is read only as far as needed and is automatically closed when the program terminates. An output file is closed by instantiating it to [ ]. When a program ends, all stream files are closed and the buffers flushed.  The following program will copy ?-  instream('inflie', instream('outfile , 1  In  inflle  to  outfile:  In), Out),  = Out.  Two input files are compared by: ?-  instream( infilel',  Inl),  instream('infile2',  In2),  1  Inl  =  In2.  In both these examples, the "=" can be done away with: ?-  ?-  instream( infile,'  CopyFile),  instream('outfile',  CopyFile).  instream('infilel',  CompareFile),  instream('infile2*,  CompareFile).  1  243  It should be noted that streams can be added to a logic programming language which implements delays. For example, the query ?- o u t s t r e a m ( ' f i l e 2 ' , O u t ) , q u e r y ( I n , O u t ) , i n s t r e a m ( ' f i l e l , I n ) , 1  reads in f i l e l and passes it through query to write out file2. The order of the goals are important to ensure lazy evaluation - the order does not affect the meaning of the query. The stream predicates can be written: i n s t r e a m ( F i l e , X.Rest) :- r e a d F i l e ( F i l e , X ) , i n S t r e a m ( R e s t ) . instream(File, []) :- r e a c h e d E o f ( F i l e ) . outstream(File, outstream(File,  X ? . R e s t ) :- w r i t e F i l e ( F i l e , X ) , o u t S t r e a m ( R e s t ) . []) :- c l o s e O u t p u t ( F i l e ) .  where r e a d F i l e , reachedEof, w r i t e F i l e and c l o s e O u t p u t are non-logical I/O predicates as in most Prologs. However, streams should be integrated into the unification mechanism or treated as lazy functions (see section 6.4.1, "Lazy thunks vs. delayed predicates" on page 132). Many varieties of the stream meta-predicate are possible. For example, we might want streams which pass single characters or strings (delimited by white space or delimiters) or even general purpose pattern matching, like that provided by the C library routine scanf.  11.9  Implementation of many object types  The above discussion has proposed adding some new objects to a logic programming implementation, including  •  arrays 244  •  arrays with reference count of 1  • strings •  strings with reference count of 1  •  sub-arrays (and sub-strings)  •  direct-access files  • databases •  I/O streams  All these objects have a similar appearance within programs but have markedly different implementations. Their various internal representations are invisible to the Prolog programmer.  Many logic programming implementations provide some form of clause indexing. For example, the WAM  provides a switch-on-term which causes a  jump to one of a number instructions depending on whether the object is a variable, nil, list element, atom or structure (WAM  restricts this to the first  parameter of a predicate although there is no need for that restriction (see 3.0, "The basic sequential inference engine" on page 28)). This instruction can be easily extended to include any number of other object types.  For arrays, the list expression A = Hd.Tl can be. interpreted as { l e n ( A ) } > loRem(A, Hd, T l ) .  For termination, [] unifies with a 0-lengh array. The  Prolog code P([],  ID-  p ( H d . T l , Hd2.T12) :- q(Hd, Hd2),  P(T1,  T12).  245  0,  can also be interpreted as p(A, A2) :- ( l e n ( A ) } = 0, bounds(A2, 0, - 1 ) . /* n i l a r r a y */ p(A, A2) :- ( l e n ( A ) } > 0, loRem(A, Hd, T l ) , q(Hd, Hd2), p ( T l , T12), hiEx(Hd2, T12, A2).  resulting in code something like this (in pseudo-C): pCparml, parm2) [ /* g e n e r a l case f o r any types o f parms */ again: switch (typeTag(parml)) { case t y p e N i l : unifyWithNil(parm2); break; case t y p e L i s t E l e m e n t : u n i f y L i s t E l e m e n t ( p a r m l , hd, t l ) ; u n i f y L i s t E l e m e n t ( p a r m 2 , hd2, t l 2 ) ; q(hd, h d 2 ) ; parml = t l ; parm2 = t l 2 ; goto a g a i n ; /* t a i l r e c u r s i o n => i t e r a t i o n */ case typeStream: readFromStream(parml, hd, t l ) ; /* ... c o n t i n u e as f o r t y p e L i s t E l e m e n t ... */ goto a g a i n ; /* t a i l r e c u r s i o n => i t e r a t i o n */ case t y p e A r r a y : i f ( r e f C o u n t ( p a r m l ) ) > 1) { a = copyArray(parml); p A r r a y ( a ) ; /* d e t e r m i n i s t i c update (code i s below) */ u n i f y ( a , parm2); } else [ pArray(parml); u n i f y ( p a r m l , parm2); }  break; ... cases f o r o t h e r types ...  246  default: error(...); } }  v o i d p A r r a y ( a ) { /* d e s t r u c t i v e u p d a t i n g v e r s i o n f o r a r r a y s */ f o r ( i = l o b ( a ) ; i < h i b ( a ) ; i++) { q(a[i], ax); a [ i ] = ax; } J  Any number of new internal types can be added to this skeleton. Of course, the amount of generated code will increase but execution speed will remain the same. If the logic programming language has a good optimizer, some of the cases in the above code can be determined at compile time.  11.10  Expressiveness  Many logic programs suffer from the "if you have a hammer, everything looks like a nail" syndrome. The main tool for structuring repetitive data is the list, so programs are written using lists, whether they are appropriate or not. A similar problem often arises in Fortran — structures are smashed into arrays because they are the only structuring tools in the language. For a partial discussion of one aspect of this problem, comparing records (functors) versus lists and selector functions, see [Wadler, 1987].  247  Consider a predicate which looks up a name N in a table T and returns its index Index  (starting from 1). If the name is not in the table, it is added to the end,  giving the new table NewT. lookup(N,  T , NewT,  lookup2(N,  [],  lookup2(N,  N.Rest,  lookup2(N,  N2.Rest,  Index)  Index,  lookup2(N,  [N],  Index, I,  Rest,  :-  lookup2(N,  1,  NewT,  Index).  Index). N.Rest,  N2.Rest2,  {I  T,  +  1},  Index)  :-  Index). Index)  Rest2,  :-  N *  N2,  Index).  Here is array based code: lookup(N, lookup2(N,  T, T,  NewT,  Index,  NewT,  lookup2(N,  Index)  :-  T,  Index >  hiEx(T, Iookup2(N,  T,  Index,  T,  Index)  :-  1,  T,  I,  NewT,  Index)  :-  I n d e x =< I n d e x =< {T[I]}  {hib  Index).  T}, / * NewT = T || N * /  T},  = N.  {hib  T},  * N,  looku 2(N, P  {hib  N , NewT).  {T[Index]} lookup2(N,  NewT,  T,  {I  +  1},  NewT,  Index).  Even though it is slightly longer, the array version is clearer because it says directly that the  Indexth  element of  T  is  N -  this must be deduced by an  inductive proof for the list version. Furthermore, the array version creates a new table only if the name N is not already in the old table (which can often be done by destructive assignment to the old table). If the list version were modified to create a new list only when the name is not in the old list, it would be longer and much less clear than the array version. If greater efficiency is 30  desired, the list version can be transformed to use a binary tree. But the  30  Some Prolog textbooks have truly horrible examples, using a list with an uninstantiated variable as the tail, then using var to determine whether or not the name is in the list. This  248  greatest efficiency is attained with a hash table which is efficient only because all elements are in an array and are accessible in constant time.  Arrays allow symmetric access to elements. An algorithm can as easily be written to process from the end of an array as from the beginning. Lists are best handled only from the beginning to the end. Some people would claim that logic programmers do not write algorithms, just specifications. But there is some concern for efficiency, otherwise we could just write:  31  sort(Unsorted,  Sorted)  :-  permutation(Unsorted,  Sorted),  ordered(Sorted).  Especially in "very high level" languages, programs often do not deal with lists or arrays but with sets or multi-sets, mappings and similar mathematical objects. These can be realized as lists or arrays but often they are better handled as abstract types. For example, the array handling predicates could be extended to work with lists (although, in this case, very inefficiently):  var technique does not work for more complex problems, such as are encountered by register allocation algorithms.  31  Although with all the clauses for o r d e r e d and p e r m u t a t i o n , this is actually longer than  quick-sort.  249  elem(L, /*  I,  treat  E)  :- islist(L),  a list  elemOfList(L,  as a 0 - o r i g i n  elemOfList(E.Rest,  I,  elemOfList(E.Rest,  L o , I,  array:  0,  I, E ) .  * /  I, E ) . E)  : - Lo < I, elemOfList(Rest,  chElem(L,  I,  E , LOut)  : - isList(L),  chElemOfList(X.Rest,  I,  chElemOfList(X.Rest,  L o , I,  chElemOfList(Rest, swapElem(A,  I,  J,  I,  E,  A O u t ) <-  elemOfList(A,  I,  Al) &  elemOfList(A,  J,  AJ)&  chElemOfList(A, chElem0fList(A2,  I,  chElemOfList(L,  0,  I,  I, E ) . E , LOut).  E.Rest).  E , X.RestOut)  {Lo+l},  {Lo+1},  I,  E,  : - Lo < I,  RestOut).  isList(A),  A J , A2) &  J,  A l ,AOut).  No one would want to use such predicates for quick-sort but they will do the job correctly. It is not unreasonable to build knowledge of such predicates into a Prolog compiler, so that the appropriate data conversions are done when needed. Thus, if the programmer entered ?-qsort(List,  OutputList),  where List and OutputList ?-listToArray(List, arrayToList(Array2,  q(OutputList).  are lists, the compiler would convert this to:  Array),  qsort(Array,  OutputList),  Array2),  q(OutputList).  by noticing that qsort's preferred data type is an array. Such data transformations have been investigated for a long time, for example in [Schwartz 1975]. Pure logic programming allows such a style of programming with optimizations being applied by a combination of automatic type inferencing and dialogue with the programmer.  250  FP [Backus 1978] and APL treat arrays as simple objects. Implementing these arrays requires optimizations similar to those given above.  FP depends heavily on functions which manipulate other functions. In logic programs, we can easily have predicates which manipulate other predicates (see 6.4, "Thunks, lazy evaluation and higher order functions" on page 127). Although predicates can be treated as "first class objects," they cannot be unified in the same way as other objects - there is no way to tell if arbitrary recursive predicates are the same.  Assert  and r e t r a c t  3 2  have two uses: to dynamically add new clauses and to  implement data bases. The former use can be problematic if the logic program is compiled; the latter use can be handled better by the array or database predicates described above.  Non-logical I/O is much less expressive than logical I/O. It requires that goals be in a particular order (contrary to the commutative nature of and Non-logical I/O does not easily support backtracking.  3 2  a d d a x and d e l a x in Watedoo and I B M Prologs.  251  (",")).  Conclusion Logic programming is still in its infancy. Just as Fortran was superseded by Algol, PL/I, Pascal, Ada and many others, Prolog will surely be superseded. Prolog is a first attempt at a new style of programming. As such, we should not criticize its defects, but applaud its boldness. We can only hope that its defects do not linger on, as do Fortran's. Innovations are needed.  Logic Programming is now well-established as a practical method of building computer systems. My own experience, and that of many others, is that logic programming can give programmer productivity increases of 10-20 times. The challenge is to make logic programs run as fast as those produced by conventional means, and to provide tools for building large logic programs. I have explored a variation on the popular WAM  implementation of a logic  engine (xpPAM) which retains WAM's efficiency, yet implements a more powerful  252  language than conventional Prolog, providing sound negation and coroutining. This design is also suitable for functional programming.  I have implemented the logic engine (except for virtual memory), including an prototype optimizing compiler, and have attained performance comparable to WAM  implementations. The compiler is short and simple because the abstract  logic engine's instructions lend themselves to easy compilation. Suitable language constructs - such as delays,  if-then-else  and meta variables - both  help the compiler and allow programmers to write clearer code.  Logic programming is a superior paradigm for many problems, compared to conventional programming techniques, because it improves programmer productivity, allows rapid prototyping and encourages writing provably correct programs. My work shows that logic programs can run as fast as conventional programs, without any need to introduce "impure" non-logical features.  Logic programming is a superior paradigm for many problems, compared to conventional programming techniques - it improves programmer productivity, allows rapid prototyping and encourages writing provably correct programs. My work shows that pure logic programs can be executed efficiently, with any need to introduce "impure" non-logical features. Continued work on techniques for compiling logic programs, such as those in this report, will give pure logic programs the efficiency of conventional programs.  " If anybody wants to clap," said Eeyore when he  253  had read this, "now is the time to do it." They all clapped. "Thank you," said Eeyore. gratifying  "Unexpected  if a little lacking in  — A. A. Milne,  254  and  Smack."  Winnie the Pooh  Glossary and Index  address  The index into memory of an object. See also "pointer," section3.1, "Objects" on page 29 .  algorithm  A method of computing a proof. Kowalski's definition is: algorithm  all-solutions predicate A  = logic +  control.  second-order predicate which computes all the solutions  for a predicate. These are usually called  bagof  and  setof.  is in computed order and may contain duplicates; setof  is  Bagoff  ordered  and contains no duplicates. In some implementations, no solution results in nil; in other implementations, the all-solutions predicate fails. See section 8.9, "All solutions predicates" on page 178.  argument  A value passed to a goal. See also "parameter."  255  atom  A simple value, typically a name (string) or a number. See also "compound term."  axiom  A predicate or clause.  backtrack  To reset the machine to an earlier state and pick an alternate clause to resume computation. See also "non-deterministic," section 4.0, "Backtracking and delaying" on page 66.  backtrack stack A stack which records information necessary to reset the machine to an earlier state. Also called "choice stack" or "choice point stack" See section 4.0, "Backtracking and delaying" on page 66. call  Attempt to compute a goal (also, "try" a goal).  choice point  A collection of information in the backtrack stack which contains information about an alternative clause to take on backtracking. See section 4.0, "Backtracking and delaying" on page 66.  clause  One alternative in a predicate. It consists of a "head" and zero or more "goals" which must all be satisfied if the clause is to be satisfied. See section 8.2, "Conventional Prolog's execution order" on page 159.  256  closed (completed) predicate A predicate which has all its alternatives given. Its negative can therefore be computed with the "closed world" technique. See section 10.1, "Negation" on page 207. code segment In xpPAM, the abstract machine instructions, plus constants for one predicate. A code segment may also contain instructions for predicates which are strictly internal to it. See section 3.6, "Code segments" on page 36. complete  Describing a proof procedure which will always terminate successfully if the goal is provable. See also "sound."  complex indeterminate A compound term which contains field (slot) names and associated values. Has possibilities for implementing frames (q.v., second meaning). See section 3.14, "Other object data types" on page 59. compound term A term of the former,, t ,... Q where each t may be either an 2  {  atom or a compound term. Also called a "functor." Some compound terms have special forms. For example, a list element is notated  "A.B"  "[A\B]  n  or ".(A,B)"  -  in all cases, the functor  name is "." and the arguments are A and B. computation  A series of steps that a machine takes in attempting to prove a query. See also "proof procedure."  257  coroutine  A predicate which may suspend before its computation is complete. Typically, one or more predicates cooperate by suspending and resuming. Coroutining may be considered as a form of parallelism. See section 9.0, "Coroutining, pseudo-parallelism and parallelism" on page 185.  cut ("!")  An "impure" predicate in conventional Prolog which "cuts" the solution space. Its declarative reading is "true" but it can change the generated solutions. Cuts can be avoided by use of suitable constructs (such as  if-then-else)  or by proper implementation (tail  recursion optimization, etc.).  declarative reading The logical reading of a Prolog predicate. If the predicate contains only purely logical predicates, the declarative reading can be produced by "or"-ing the clauses, replacing commas by "and"s and adding existential and universal quantifiers. If the predicate contains non-logical predicates (var, cut ("!"), I/O, etc.), a declarative reading is much more difficult.  delay  Postpone evaluation of a predicate until one or more arguments become sufficiently instantiated. See sections 1.5, "Delay notation" on page 8 and 8.10, "Non-strict execution order: delays" on page 181.  deterministic A computation which can only proceed in on sequence of steps, producing a single answer.  258  deterministic append  A predicate which appends two lists to make a third list.  This is the inner loop of "naive reverse." Append can also be non-deterministic, in which case it is used to split a list into two. append([],  append(X.A,  execution stack  X,  B,  X).  X.C)  :-  append(A,  B,  C).  A stack which keeps information across calls. Also called "call  stack." See section3.5, "Execution stack" on page 34 .  fact  A clause (or predicate) which contains no goals and completely grounds its arguments. See also "rule."  fail  In a computation, if a predicate does not succeed, it is said to fail. If a goal fails and the computation is sound, then there is no possible proof for the goal.  frame  Two meanings:  •  A collection of related information on the execution or backtrack stack.  •  A collection of information associated with an object (used for Al or object-oriented programming). Typically, the frame slots are values or procedures for computing the values. See also "complex indeterminates."  259  free list  A list of cells which are not yet allocated to objects. See also section 7.5, "Allocating from a list or from a stack" on page 145.  freeze (or geler) The predicate in Colmerauer's Prolog-II which implements  "delay." See also section9.7, "Comparison with other designs for delaying." on page 200 . function  A mapping from one set of objects to another. An ^-argument function can be transformed to an (n + /J-argument deterministic predicate by "returning" the answer in the (n + l)th argument.  functor  A compound term.  fully instantiated (or ground) Has a value with no uninstantiated variables. See "instantiated" and "sufficiently instantiated." functor element cell If a functor is represented internally as a list of cells, this is like a list element cell but with a special flag so that it will print out in functor notation and will only unify with another functor. See also "list element cell," sections 3.1, "Objects" on page 29, 7.10, "Functor and list storage" on page 151. goal  One of the terms on the right hand side of a clause which must be satisfied if the clause is to be satisfied. See section 8.2, "Conventional Prolog's execution order" on page 159.  260  ground  Has a value. See "instantiated."  head  The name and parameters of a clause.  heap  An area of memory within which objects may be allocated and freed in any order. See also "stack," section 7.5, "Allocating from a list or from a stack" on page 145.  Horn clause  A clause containing a head and zero or more goals. The more general form (from Gentzen's sequent calculus) allows more than one term in the head.  KLIPS  Thousand LIPS.  if-then-else  A control construct which can be defined: p :- i f t t h e n q e l s e r .  is the same as: P t , q. p :- n o t t , q.  See section 8.6, "Single solutions" on page 170.  impure predicate  A predicate which cannot easily be given a logical or  declarative meaning. For example, cut ("!") and var require knowledge of the computational mechanism; I/O predicates  261  require transformations (with extra arguments) to be given declarative meanings. instantiate  To cause an uninstantiated variable object to receive a value. This normally happens during unification. See section 3.10, "Unification" on page 40.  instantiated  Has a value. Also called "ground." When an object is created during computation, it is initially uninstantiated; it may become instantiated during unification. Once instantiated, an object cannot change its value; backtracking may cause it to become uninstantiated again.  LIPS  Logical inferences per second. Usually computed with the "naive reverse" benchmark which typically gives speeds about three times faster than "typical" predicates. There are rumours that some implementations have a built-in "naive reverse" instruction (or deterministic append instruction) to give high LIPS figures. See section 5.9, "Optimized compiling to conventional machine code" on page 110.  list element cell An object containing a "head" and a "tail" (corresponding to LISP's car and cdr). For example, the list  contains two  list element cells: thefirst'shead is A, the first's tail points at the second cell; the second's head is B and the second's tail contains nil. See also "functor," sections 3.1, "Objects" on page 29, 7.10, "Functor and list storage" on page 151. 262  logical predicate A "pure" predicate which can easily be given a logical or declarative meaning in first (or second) order predicate calculus.  meta-predicate A predicate which has no effect on correctness of a computation but which may control some aspect of the representation or of the order of computation. For example, a meta-predicate may supply delaying information or it may associate an input list with a file. Sometimes, "second order" predicates (such as call are called meta-predicates. See section 11.8, "Sequential Input/Output" on page 241 for examples.  meta-variable A variable which is used as a "place holder" within a predicate skeleton. This is typically used for all-solutions predicates or for knowledge representation which uses predicate forms. See section 10.4, "Meta-variables" on page 212. MIPS  Million instructions per second (also called "meaningless indicator of processor speed").  MLIPS  Million LIPS.  naive reverse A predicate which reverses a list using a rather simple-minded algorithm which mainly consists of calls to deterministic append. Naive reverse is often used to give LIPS figures. naive_reverse( [ ],  []).  naive_reverse(X.A,  R)  :-  naive_reverse(A, append(A2,  263  [X],  A2), R).  negation  The contra-positive of a predicate. That is, if a predicate evaluates to "true," its negation evaluates to "false." Some logic systems allow making negative statements; others, such as Horn clause logic allow only making positive statements. There are a number of kinds of negation, including: classical negation which requires full resolution in the general case (for example, given "all politicians exaggerate" and "John does not exaggerate," this could derive "John is not a politician"). closed world negation which will only correctly negate full ground predicates by checking if they are not in the database of clauses. It is simple to compute but less complete that classical negation See sections 10.1, "Negation" on page 207, 8.5, "Negation" on page 168.  non-deterministic A computation in which the order is not known ahead of time. There are two main kinds of non-determinism: order of clause selection and multiple answers. non-logical  Not capable of being given an simple first or second order predicate calculus meaning. See "impure predicate."  264  nonvar  An "impure" predicate in conventional Prolog which succeeds if its argument is instantiated. It has no possible declarative reading.  Nonvar  can be avoided by use of suitable control  constructs (such as "delay"). object  A piece of allocated memory with a tag indicating its type. See section 3.1, "Objects" on page 29.  parallel computation A computation in which more than one predicate are computed in parallel. There are two main forms: and-parallelism requires all the parallel predicates to succeed; it waits until the last predicate succeeds and is typically used for divide-and-conquer algorithms. or-parallelism  requires one of the parallel predicates to succeed;  it is typically used for database searches. parameter  A value passed into a clause. See also "argument."  pointer  An object which contains the address of an object. See also "address," section3.1, "Objects" on page 29 .  predicate  A term of the form  p{f\,f , ••• ,f„) 2  which is supposed to have a  value of "true" or "false" for each instantiated substitution for its variables. A predicate may be (recursively) computed, so some substitutions may not be computable.  265  Prolog  "Programming in Logic": a logic programming language, originating from the work of Colmerauer and Kowalski. It has a number of dialects, of which the most popular is Edinburgh (DEC-10) Prolog and its derivatives (C-Prolog, etc.). IBM Prolog uses a somewhat different syntax which can (usually) be mechanically transformed to Edinburgh Prolog. Most of the differences among Prolog dialects are minor, usually in details of built-in predicates, especially "impure" predicates.  proof procedure  A mechanical method of proving a goal. Sometimes called a  computational method. A proof procedure typically selects some path through the tree of all possible proofs. See section 8.3, "Example of conventional execution order" on page 163.  pure  Containing only logical predicates. "Pure Prolog" is sometimes, mistakenly, referred to as a subset of Prolog.  query  A goal or conjunction of goals which the machine will attempt to prove (in the original theory, a query is a single negated goal and the computation attempts to find a substitution which results in a contradiction; this corresponds tofindinga substitution which satisfies the axioms and original goal).  register  In xpPAM, a special piece of memory containing the address of an object. A register may be empty, in which case its contents are meaningless. See section 3.2, "General registers" on page 32.  266  The status of the machine is kept in special status registers, see section 3.4, "Status registers" on page 33. register annotation One of the letters v, n,f, x or s which gives information about the contents of a register and what is to be done with it in an instruction. The annotation c indicates that the operand is not a register but an index into the constants' vector. See section 3.7, "Machine instruction format" on page 36. reset stack  A stack which contains information about variables which have become instantiated and which require being "un-instantiated" on backtracking. Also called "trail." See section 4.0, "Backtracking and delaying" on page 66.  resolution  A method discovered by Robinson based on Gentzen's cut rule which provides a single rule for doing a proof. Full general resolution can have exponential cost.  resume  To continue execution of a predicate where is was earlier suspended. See also "suspend," "coroutine."  rule  Sometimes used to mean "clause" or "predicate." A rule normally has at least one goal. See' also "fact."  satisfy  To try a goal and succeed or to find a substitution during a unification.  267  sequent calculus  A method of proving first order predicate calculus goals, using  an alternate notation which was invented by Gentzen. Horn logic is a subset of sequent calculus.  sound  Describing a proof procedure which always gives a correct answer (success or failure corresponding to provable or unprovable). See also "complete."  stack  An area of memory from which objects may be allocated and freed in strict LIFO order. See also "heap," section 7.5, "Allocating from a list or from a stack" on page 145.  string  Zero or more characters. If the string starts with a lower case letter and contains only printable non-space characters, it can be written without enclosing quotes, like an atom. Otherwise the quotes are not needed:  substitution  "abc='abc\"  A set of values which satisfy some equalities (usually, computed by unification).  succeed  Of a predicate, to be proved. See also "fail."  suspend  To cease execution of a predicate and start or resume execution of another before the current predicate has terminated with either success or failure. See also "resume," "coroutine."  268  tail recursion optimization An optimization which turns the call to the last goal in a clause into a kind of go to. Also called "TRO." TRO is used to transform recursion into iteration. See section 3.9, "Tail recursion optimization (TRO)" on page 39 and the "last call" instructions. term-rewriting A general proof procedure (used, for example, by sequent calculus) in which an arbitrary goal is chosen and its right hand side is substituted for it. See section 8.4, "A more flexible execution strategy" on page 165. Similar to Markov algorithm. If the left-most goal is always chosen, term-rewriting is the same as conventional Prolog's depth-first left-to-right computation rule. TRO  Tail recursion optimization.  try  Attempt to compute a goal (also, "call" a goal).  thunk  An object which contains a predicate name and some or all of the arguments for it. If all the arguments are present, it is very similar to a suspended predicate. See section 6.4, "Thunks, lazy evaluation and higher order functions" on page 127.  trail  WAM  terminology for the reset stack.  269  sufficiently instantiated (or ground) Has enough of a value for computation to proceed. For example, the computation may require that an object be instantiated to a list element, but the head or tail may still be uninstantiated. See also "instantiated" and "delay." unification  The process whereby two terms (possibly containing uninstantiated variables) are made "equal." Unification may fail.* Unification may cause variables to become instantiated, always to the "most general unifier." For example, "X=Y,Y=a" results in X=a and Y=a; "X=Y,Y=Z" results in X=Y, X=z and Y=z (the most general unifier) even though x=Y=Z=a will satisfy the equation. See section 3.10, "Unification" on page 40.  var  An "impure" predicate in conventional Prolog which succeeds if its argument is uninstantiated. It has no possible declarative reading. Var can be avoided by use of suitable control constructs (such as "delay").  variable  WAM  Two meanings: •  a name in a predicate, usually starting with a capital letter.  •  an uninstantiated variable during the course of a computation.  Warren Abstract Machine.  270  Warren Abstract Machine A design by D. H. D. Warren for an abstract machine which has proven to be very efficient. It is the basis of most recent hardware and software implementations of Prolog. See section 7.1, "Comparison with the Warren Abstract Machine instructions" on page 138.  weak delay  A delay which has an associated "cost." Unlike a regular delay, a weak delay may resume without its argument(s) being instantiated, but the computation is likely to be expensive (the cost gives an indication of the expense). See section 4.6, "Weak delays: dynamic reordering of clauses" on page 81.  xpPAM  The x p P r o l o g abstract machine. An abstract machine design suitable for implementing conventional Prolog, x p P r o l o g or functional languages. It has some similarities with WAM but is both more simple and more flexible, while retaining WAM's efficiency. XpPAM can be interpreted on conventional machines, used an intermediate language when compiling to conventional machines, or xpPAM could be implemented in hardware. See sections 3.0, "The basic sequential inference engine" on page 28. and 4.0, "Backtracking and delaying" on page 66.  xpProlog  Extended pure Prolog.  The pure "subset" of Prolog, extended  with constructs (such as if-then-else, meta-variables, delays, etc.) which make "impure" predicates unnecessary.  271  References Abramson, H . [1984] A Prological Language  with Unification  Definition  of HASL  Based Conditional  a Purely  Functional  Binding Expressions.  New  Generation Computing, 2. Springer-Verlag. Auslander, M . and Hopkins, M [1982] An Overview of the PL.8 Compiler.  Proc.  SIGPLAN 1982 Symposium on Compiler Construction. Backus, J. [1978] Can programming (A.CM.  Turing Award  be liberated from  Lecture).  style?  Communications of the A.C.M., 21(8).  Bosco, P. and Giovanetti, E. [1986] IDEAL: Language.  the von Neumann  An Ideal DEductive  Applicative  Proc. IEEE 1986 Symposium on Logic Programming.  Bratko, I. [1986] Prolog Programming  for Artificial  Intelligence.  Addison-Wesley 1986 Bruynooghe, M . [1982] The Memory Management  of Prolog  Implementations.  In "Logic Programming": Clark, K.L., Tarnlund, S-A. (ed.). Academic Press.  272  Bruynooghe, M . [1986] Compile time garbage collection Report CW43,  Department Computerwetenschappen, Katholieke Universiteit Leuven. Burge, W. [1975] Recursive Programming  Techniques.  Addison-Wesley.  Campbell, J. and Hardy, S. [1984] Should Prolog be list or record oriented?. In  "Implementations of Prolog": Campbell, J. (ed.). Ellis Horwood (1984). Cheng, M . [1986] Logical I/O for Prolog.  Dept. of Computer Science,  University of Waterloo. Clark, K. [1978] Negation as failure.  In "Logic and Databases": Gallaire, H.  and Minker, J. (ed.). Plenum Press. Clark, K. and Gregory, S. [1984] PARLOG:  Parallel  Programming  in Logic.  Research report DCO 84/4, Dept. of Computing, Imperial College, London. Also in A.CM. Transactions on Programming Languages and Systems 8(1) (January 1986). Clark K., McCabe, F. and Gregory, S. [1982] IC-Prolog  Language  Features.  In  "Logic Programming": Clark, K.L., Tarnlund, S-A. (ed.). Academic Press. Clark, K. and McCabe, F. [1984] micro-PROLOG:  Programming  in Logic.  Prentice-Hall. Clocksin, W. and Mellish, C. [1981] Programming  in Prolog.  Cohen, P and Feigenbaum, E. (ed.) [1982] The Handbook of Intelligence,  Vol. 3.  Springer-Verlag. Artificial  HeurisTech Press.  Colmerauer, A. [1982] PROLOG-II  Manuel  de Reference et Modele  Theorique,  Groupe Intelligence Artificielle, Univ. d'Aix-Marseille II. Dahl, O-J., Dijkstra, E. W. and Hoare, C. A. R. [1972] Structured Programming.  Academic Press.  Debray, S. [1985] Register allocation in a Prolog machine.  85/10, State University of New York, Stony Brook.  273  Technical Report  Debray, S. [1986] Towards Banishing the Cut from Prolog.  Proc. IEEE 1986  International Conference on Computer Languages. Debray, S. and Warren, D. S. [1986] Detection and Optimization Computations  in Prolog.  of  Functional  Proc. Third International Conference on Logic  Programming. Springer-Verlag 1986. DeGroot, D. and Lindstrom, G. [1986] Logic Programming and Equations.  Functions,  Relations  Prentice-Hall.  Dijkstra, E. W. [1968] Go to statement considered harmful.  Communications of  the A.C.M. 11 (March 1986). Dijkstra, E. W. [1976] A Discipline of Programming. Dobry, T. P. [1987] A high performance  architecture  Prentice-Hall. for Prolog.  Ph. D. thesis,  report UCB/CSD 87/352, University of California at Berkeley. Dobry, T. P., Patt, Y. N. and Despain, A. M . [1984] Design decisions the microarchitecture  for a Prolog machine.  influencing  Micro 17 Proceedings, October  1984. Gabriel, Linkholm, Lusk and Overbeek [1985] A Tutorial Abstract  Machine for Computational  Logic.  on the Warren  Argonne National Laboratory  Report ANL-84-84. Gabriel, R. P. [1985] Performance  and Evaluation  Golderg, A. [1983] Smalltalk-80:  The Language  of Lisp Systems. and its  MIT Press.  Implementation.  Addison-Wesley. Henderson, P. [1980] Functional Implementation.  Programming:  and  Prentice-Hall.  Hermenegildo, M . and Nasr, R. [1986] Efficient AND-Parallelism.  Application  Management  of Backtracking  Proc. Third International Conference on Logic  Programming. Springer-Verlag 1986. Hoare, C. A. R. [1986] Communicating  Sequential  274  Processes.  Prentice-Hall.  in  Hodges, W. [1971] Logic.  Penguin Books.  Ingerman, P. [1961] Thunks - A way of compiling procedure some comments on procedure  statements  with  Communications of the A.CM.  declarations.  4, 1. Kleene, S. [1967] Mathematical  Logic.  John Wiley & Sons.  Kluzniak, F. [1981] Remarks on Coroutines in Prolog In "Papers in Logic  Programming I.":, Report 104, University of Warsaw, Institute of Informatics (also for the closed Workshop on Logic Programming for Intelligent Systems, 18-21 August 1981, Long Beach Harbor, California). Kluzniak, F. and Szpakowicz, S. [1985] Prolog for Programmers.  Academic  Press. Kluzniak, F. [1987] Compile time garbage collection for ground  Prolog  (unpublished) Warsaw University Institute of Informatics. Knuth, D. E. [1973] The Art of Computer Algorithms,  Programming,  Vol. I:  Fundamental  pp. 417-420. Addison-Wesley.  Kowalski, R. [1979] Logic for Problem Solving. Kowalski, R. [1979b] Algorithm  = Logic  Elsevier North Holland  + Control.  Communications of the  A.C.M. August 1979 Krasner, G. (ed.) [1983] Smalltalk-80:  Bits of History,  Words of Advice.  Addison-Wesley. Jaffar, J. and Lassez, J.-L. [1987] Constraint  Proc.  Logic Programming.  Conference on Principles of Programming Languages. Landin, P. [1966] An abstract machine for designers of computing  languages.  Proc. IFIP Congress 65, Vol. 2, Washington. Spartan Books. Lloyd, J. [1984] Foundations of Logic Programming. Lloyd, J. and Topor, R. [1984] Making  Springer-Verlag.  Prolog more Expressive.  Logic Programming, 4. 275  The Journal of  Matsumoto, H. [1985] A Static Analysis of Prolog Programs.  SIGPLAN  Notices, V20 #10, October 1985. Mellish, C. [1982] An Alternative a Prolog Interpreter.  to Structure  Sharing  in the Implementation  of  In "Logic Programming": Clark, K.L., Tarnlund, S-A.  (ed.). Academic Press. McKeeman, W., Homing, J. and Wortman, D. [1970] A Compiler implemented for  Prentice-Hall.  the IBM System/360.  Mills, J. W. [1986] A high performance  Generator  LOW RISC  machine for logic  Proc. IEEE 1986 3rd International Symposium on Logic  programming.  Programming. Moss, C. [1986] CUT & PASTE  - defining the impure Primitives  of Prolog.  Proc. Third International Conference on Logic Programming. Springer-Verlag 1986. Mukai, K. and Yasukawa, H . [1985] Complex Indeterminates Application  to Discourse Models.  Naish, L. [1985a] Automating  in Prolog and its  New Generation Computing, 3.  Control for  Logic Programs.  The Journal of  Logic Programming, Vol. 2, Num. 3, October 1985. Naish, L. [1985b] Negation and Control in Prolog.  Melbourne. Also available as Science #238  Ph. D. Thesis, University of  Springer- Verlag Lecture Notes in Computer  (Goos, G. and Hartmanis, J., eds.) (1986).  Naish, L. [1985c] All solutions Predicates in Prolog.  Proc. IEEE Symposium on  Logic Programming (July 1985). Naish, L. &Ibrkl985d] The MU-Prolog  3.2 Reference Manual.  Department of  Computer Science, University of Melbourne Naish, L. [1986] Negation and quantifiers  in NU-Prolog.  Proc. Third  International Conference on Logic Programming. Springer-Verlag 1986.  276  O'Keefe, R. [1985] On the Treatment  of Cuts in Prolog Source-Level  Tools.  Proc. 1985 Symposium on Logic Programming. Periera, F., Warren, D. H . D., Byrd, L. and Pereira, L . M . [1984] C-Prolog User's Manual  Version 1.5.  SRI International, Menlo Park, California.  Pirsig, R. [1974] Zen and the Art of Motorcycle  Maintenance.  William Morrow  (also Bantam Books). Plaisted, D. [1984] The Occurs-check Problem in Prolog.  Proc. IEEE 1984  International Symposium on Logic Programming. Quine, W. [1941, revised 1965] Elementary  Logic.  Richards, M and Whitby-Stevens, C. [1979] BCPL compiler.  Harper and Row. - the language and its  Cambridge University Press.  Sahlin, D [1986] Making  tests deterministic  SICS,  using the reset information.  Sweden. Schwartz, J. [1975] On Programming:  an interim report on the SETL  project.  Courant Institute of Mathematical Sciences, New York University. Sergot, H . [1983] A Query-the-U ser facility for logic programming.  Proc.  European Conference on Integrated Interactive Computer Systems: Degano, P. and Sandewall, E., eds. North-Holland. Shapiro, E. [1983] A Subset of Concurrent Prolog and its Interpreter.  ICOT  Technical Report TR-003. Sterling, L. and Shapiro, E. [1986] The Art of Prolog.  The MIT Press.  Tick, E. [1985] Prolog Memory-Referencing  Stanford University  Behavior.  Technical Report No. 85-281 (September 1985) Tick, E. [1986] Memory performance  of Lisp and Prolog programs.  Proc. Third  International Conference on Logic Programming. Springer-Verlag 1986.  277  Tick, E. and Warren, D. H. D. [1984]  Towards a Pipelined  Prolog  Processor.  Proc. IEEE 1984 International Symposium on Logic Programming. Also in New Generation Computing, 2, Springer-Verlag. Turner, D. A. [1979] A Languages.  New Implementation  Technique for  Applicative  Software Practice and Experience, 9.  Ueda, K. [1983] Guarded  Horn  ICOT Technical Report TR-103 (June  Clauses.  1983). van Emden, M. [1982] An  interpreting  algorithm for  Proc.  Prolog programs.  First International Logic Programming Conference, University of Marseilles. Reprinted in "Implementations of Prolog": Campbell, J. (ed.). Ellis Horwood (1984). Van Roy, P. [1984] A  Prolog compiler for  the PLM.  Master's Report Plan II,  Computer Science Division, University of California, Berkeley. Voda, P. [1986] Choices  in, and Limitations  of, Logic Program.  Proc. Third  International Conference on Logic Programming. Springer-Verlag 1986. Voda, P. [1986b]  Pre-complete  Negation  and Universal  Quantification.  Technical Report 86-9, Dept. of Computer Science, University of British Columbia. Wadler, P. [1987] A  Critique  Better than Scheming.  of Ahelson and Sussman or Why Calculating  SIGPLAN Notices Vol. 22 #3 March 1987.  Walker, A. (ed.), McCord, M., Sowa, .1. and Wilson, W. [1987] Systems and Prolog: Language  Processing.  A Logical  Approach  to Expert  Knowledge  Systems and  Natural  Addison-Wessley.  Warren, D. H. D. [1977] Implementing Programs.  is  Prolog - Compiling  Predicate  Technical Reports 39 and 40, Department of Artificial  Intelligence, University of Edinburgh.  278  Logic  Warren, D. H. D. [1983] An  Abstract  Prolog Instruction Set: SRI  Technical  Note 309, Menlo Park, California. Warren, D. H. D. [1986] Optimizing  Tail Recursion.  In "Logic Programming  and its Applications": van Canemghen, M. and Warren, D.H.D (eds.). Ablex Publishing. Wilson, W. [1985] PureLog Meta-declarations.  Wise, D. [1987] Matrix  I: Pragmatic  Logic Programming  with  Ph. D. Thesis, Syracuse University. Algebra and Applicative  Programming.  Conference on  Functional Programming Languages and Computer Architecture, Sept. 1987.  279  Appendix A . Sample machine code The code in section 2.3, "C code for deterministic append predicate" on page 21 can be transformed to machine code, for ultimate speed. These transformations cannot easily be done by a C compiler, because certain values must be kept globally in registers for ultimate performance.  A.l  Machine code for deterministic append predicate  The C code for  append  can be translated directly to machine code. Here is  sample IBM/370 code. The parameters are in general purpose registers 1, 2 33  and 3. The next free object cell's address is in register RFREE. Register RBACK points to the top of the backtrack stack (used for determining a variable's age). RZERO  always has zeroes in the high three bytes, to avoid clearing it before  loading a character (the IBM/370 does not clear the high bytes when loading a character). RTEMP is a temporary register. For efficiency, the object tags are 0, 4, 8, etc. instead of 0, 1, 2, etc., to avoid a shift instruction.  33  This code is offered more as an example than as a definite method (Warning: the code has not been tested).  280  I w i l l use a number of macros, of  "SWITCH reg,Ai,A2,...,An" j u m p s to one o f a list  addresses depending on a tag; it expands to something like (it is assumed that  RZERO  always has the high 3 bytes zero):  SWITCH1  IC L BR DC DC  RZERO,TAG(,reg) RTEMP,SWITCH1(,RZERO) RTEMP A(A1) A(A2)  DC  A(An)  "ALLOC newtag" gets a (the first two  load t a g s w i t c h on t y p e branch table  new cell f r o m the free list; it expands to something like  instructions c a n be eliminated if the last unallocated cell can point  to an i n v a l i d address, thereby raising an addressing exception when there is no more  space left): CLI BNE  TAG(RFREE),4*TGUNALL0C next c e l l OVERFLOW  MVI  TAG(RFREE),4*newtag  free?  s e t t a g f o r new  cell  The code for the second (slower) version is: START SWITCH R1,NIL1,LISTEL1,ERROR,REF1,VAR1 s w i t c h on t y p e o f p i NIL1 ... code o m i t t e d f o r b r e v i t y REF1 L Rl,ASREF(,R1) p i = pl->u.asReference B START goto s t a r t LISTEL1 SWITCH R3,NIL2,LISTEL2,ERROR,REF2,VAR2 s w i t c h on t y p e o f p3 REF2 L R3,ASREF(,R3) p3 = p3->u.asReference B LISTEL1 back t o s w i t c h VAR2 C RBACK,VARAGE(,R3) i s p3 newer than l a s t BNH NEWER choice point? yes: s k i p ... code o m i t t e d : push p3's i n f o onto r e s e t s t a c k NEWER MVI TAG(R3),4*TGREF p3->tag = t g R e f e r e n c e ALLOC TGLISTEL RFREE = <new> ( l i s t elem) LR RTEMP,RFREE <new> = RFREE  281  NEWTAIL  L RFREE,NEXT(,RFREE) ST RTEMP,ASREF(,R3) ST R1,HEAD(,RTEMP) ST RFREE,TAIL(,RTEMP) L R1,TAIL(,R1) ALLOC TGVAR RBACK,VARAGE(,RFREE) ST RFREE,NEXT(,RFREE) L  p o i n t t o next f r e e c e l l p3->u.asReference = <new> <new>.asListElem.head = p i < n e w > . a s L i s t E l e m . t a i l = <new2> pi = pl->asListElem.tail RFREE = <new2> ( v a r i a b l e ) <new2>.varAge = c h o i c e p o i n t age p o i n t t o next f r e e c e l l  SWITCH R1,NIL1,LISTEL1,INT1,REF1,VAR1 s w i t c h on type o f p i  The inner loop is 19 or 23 instructions, depending on whether heap overflow is detected in-line or by an exception.  A.2  Machine code for append function  The above machine code is not optimal. As soon as a variable is found for the third argument, a variant of the loop in thefirstexample C program can be used. The code from NEWTAIL on is replaced by: inner: s w i t c h (p->tag) { case t g L i s t E l e m : *p3 = a l l o c L i s t E l e m ( p l - > u . a s L i s t E l e m . h e a d , pi = pl->u.asListElem.tail; p3 = & ( ( * p 3 ) - > u . a s L i s t E l e m . t a i l ) ; goto i n n e r ; case t g N i l : *p3 = 2 ; break; case t g R e f e r e n c e : p i = pl->u.asReference; goto i n n e r ; default: error(); P  }  282  &dummyCell);  which in assembler is: NEWTAIL INNER  * INNER2  DONE  LA R3,TAIL(,RTEMP) p3 = & <new>.tail b r a n c h t o t e s t a t end o f l o o p B INNER2 RFREE = <new> ( l i s t elem) ALLOC TGLISTEL ST RFREE,0(,R3) *p3 = <new> MVC HEAD(RFREE),HEAD(Rl)1 <new>.u.asListElem.head = pl->u.asListElem.head p3 = & ( ( * p 3 ) - > u . a s L i s t E l e m . R3,TAIL(,RFREE) LA RFREE,NEXT(,RFREE) p o i n t t o next f r e e c e l l L pi = pl->u.asListElem.tail L R1,TAIL(,R1) CLI TAG(Rl),4*TGLISTEL w h i l e p l - > t a g = t g L i s t E l e m BE INNER o r p l - > t a g == t g R e f e r e n c e CLI TAG(R1),4*RGREF BNE DONE deref p R1,ASREF(,R1) L INNER2 and t r y a g a i n B i f ( p l - > t a g != t g N i l ) CLI TAG(Rl),4*TGNIL error() BNE ERROR R2,0(,R3) *p3 = p2 ST  This inner loop from the INNER label to the BE INNER instruction is just 8 or 1 0 instructions, depending on whether heap overflow is detected in-line or by an exception. On a "1 MIPS" machine, the above loop will run at over 100 KLIPS (thousands of Logical Inferences Per Second). The inner loop has the same instruction count, for either allocating from a heap or for allocating from a stack. For reference counting, three extra instructions are needed (load, add one, store). For this particular example, it appears as if reference counting is significantly slower than a marking garbage collector. However, a marking garbage collector must eventually scan the list and it probably will take more than three  283  instructions to mark a cell. See section 7.12, "Reference counts and garbage collection" on page 153.  However, the Warren Abstract Machine has a very fast way of releasing storage once a query is finished - it simply pops the global heap to its initial position (some people force this using a repeat, fail loop). However, in the general case, where intermediate structures are created, the Warren Abstract Machine must also garbage collect its global heap.  284  Appendix B. Details of xpProlog syntax X p P r o l o g is syntactically very similar to conventional Edinburgh syntax [Periera,  Warren, Byrd and Pereira 1984]. The exact syntax is not very important; to follow micro-Prolog's or Waterloo (IBM) Prolog's syntax would require very little work. I will describe xpProlog's extensions to conventional Prolog's syntax with examples.  34  B.l Functors and lists Functor and lists are treated in similar ways. That is, the name of a predicate 35  may also be a variable where conventional Prolog requires that it be a name (string). X p P r o l o g allows " p r e d ( F ( A r g s ) ) :- ..." where conventional Prolog  3 4  A confession: not everything described here has been implemented — I ran out of time. The main thrust of my work was to show that pure Prolog can be implemented efficiently; I considered the exact syntax of the resulting language to be less important.  3 5  I prefer the infix notation " H d . T l " rather than " [ Hd|Tl ]." accept this by the command : -  op(600,  xfy,  285  ('.')).  C-Prolog can be persuaded to  would require " p r e d ( F A r g s ) :- FArgs =.. ( F . A r g s ) , ,..." In other words, we 3 6  can define univ by: F f j A r g s ) =.. ( F . A r g s ) .  /* F(|Args) =. . [F|Args]. */  which also works if there are no arguments (in which case, Args is unified with [ ]). Also, s() = s but [S] * s. In a clause, if a goal is a functor or a list, it is called. There is no need to use univ  before calling, nor to explicitly use the c a l l meta-predicate. That is:  p r e d ( F , A r g s ) :- ( F . A r g s ) , ...  is the same as p r e d ( F , A r g s ) :- F ( A r g s ) , ...  which would have to be written like this in conventional Prolog: p r e d ( F , A r g s ) :- C a l l =. . F.Args, c a l l ( C a l l ) , ...  (Some Prologs allow leaving out the " c a l l " ) .  B.2 Lexical and syntactic details This is a brief attempt to explain some of the intricacies of the rules for parsing Prolog. The explanation in [Clocksin and Mellish 1984] leaves a few things somewhat unexplained. Also, my method is slightly different from that used in C-Prolog [Periera, Warren, Byrd and Pereira 1983] (it is "cleaner," I think).  37  3 6  The parentheses around "F.Args" are necessary because the period could otherwise be interpreted as the end of the clause.  37  Prolog syntax is currently being standardized by the British Standards Institute (BSI). The process is not yet finished. It will almost certainly differ somewhat from x p P r o l o g .  286  The details of Prolog syntax are not very Important; many of these details can be changed by simply changing some lexical tables.  First, the basic lexical items:  string  Also called "name," "atom" or sometimes "id." This is a lower-case letter followed by any number of letters and digits. Any other characters can be included if the entire item is surrounded by single quotes ('""). Unprintable characters and single quotes can be included if they are preceded by a backslash ("\"), following C syntax (for example, '\\\'a\'\n' which, when output, produces a single backslash, a beep, a quote and a new-line).  variable name An upper-case letter or underscore followed by any number of letters, digits or underscores. A single underscore is an anonymous variable, unique from all other variable names in the clause. quoted string A string surrounded by double quotes ("""). Again, C syntax conventions are followed for unprintable characters and double quotes. "Abc" is equivalent to [ 'A', 'b', 'c' ].  number  A sequence of digits, optionally preceded by a negative sign ("-") and optionally containing a decimal point.  However, there will be little difficulty in modifying x p P r o l o g to conform to the new standard.  287  delimiters Most other characters are considered as delimiters ("!@#$%-'&*()-" etc.)- Except for the bracketing symbols ("()[]{}"), delimiters are treated like strings. When an operator is defined (by the op built-in predicate) and the operator's name is made up entirely of delimiters, then the entire string is treated as a single item (for example, "-->" or ":-"). white space Ignored between lexical items - includes blanks, tabs, new lines and comments. Comments may either be bracketed by "/* ... */" or marked by "%" which causes everything up to the end of line to be ignored. statement terminator A period ("."). Unfortunately, it can also be used as a decimal point in numbers and as the "cons" operator for lists. Therefore, "p(X):-x=a.b." must be written " (X):-x=(a.b)." to P  avoid ambiguity (the former phrase would be interpreted as two clauses: "p(X):-x=a." and "b.").  An operator is defined by the op built-in predicate: :- o p ( P r i o r i t y , A s s o c i a t i v i t y , OpName).  The operator name ("OpName") must be enclosed in single quotes if it contains any delimiter characters. The operator name is then added to a lexical table so that it will be subsequently treated as a single item.  288  Unlike "normal" compilers, higher priorities binds looser. Here are the usual 38  arithmetic operators: :::-  op( 500, op( 500, op( 500, op( 500, op( 400, op( 400, op( 400,  yfx, yfx, fx, fx, yfx, yfx, yfx,  +) -) +) -) *). /). /).  % unary ( p r e f i x ) % unary ( p r e f i x )  The associativity is of the forms:  39  x f x x f y y f x y f y (for infix operators) fx xf  fy  (for prefix operators) yf  (for postfix operators)  If there are no parentheses, a y means that the argument can contain operators of the same or lower precedence (same or tighter binding) while x means that the argument can contain operators of strictly lower precedence (tighter binding). Thus, y f x is left-to-right and x f y is right to left. Some operators (such as comparisons) are defined xfx. This prohibits, for example, a < b < c although ( a < b) < c would be legitimate (but meaningless).  3 8  T h e BSI committee is proposing to change this.  3 9  In I B M (or Waterloo) Prolog, the associativity is one of p r e f i x , s u f f i x , l r or r l which doesn't allow quite the same degree of control as does C-Prolog's values. O n the other hand, I B M Prolog is more readable.  289  Some ambiguities are still possible. An x f y operator next to a y f x operator can be parsed in two ways if they have the same priority. Such cases are flagged as errors by the parser ("ambiguous operator juxtaposition").  Note that monadic (prefix and postfix) operators also have priorities. For example, ?-  X i s 1+2.  is parsed ?-  (X i s ( 1 + 2 ) ) .  whereas  -a+b  is parsed  (-a)+b. An operator is treated as such only if it is in a position which allows it to be an operator. For example, in "a opx + b," the opx can potentially be either a 40  postfix operator or an infix operator. If opx is a prefix operator, then the "+" must be an infix operator ("(a opx) + b"); if opx is an infix operator, then the "+" must be a prefix operator  ("a  opx  (+b)").  The decision is done left-to-right  with no backtracking. C-Prolog distinguishes between operators and functor names. It considers "not x"  and  "not(x)"  to be distinct.  XpProlog  does not make this distinction.  Furthermore, there is an ambiguity with prefix operators and comma operators (recall that ?-op(iooo,xfy, ,') - that is, comma is a right-to-left infix 1  operator). If "+" is a prefix operator, then  +(l,2)  can be parsed as either the^  binary operator + applied to the two operands l and 2 or it could be the unary operator applied to the single operand which is itself a binary operand applied to the two operands  1  and  2. X p P r o l o g  takes the former meaning; if the latter is  desired, it should be written " + ( ( 1 , 2 ) ) . " (C-Prolog distinguishes between the  40  Some Prologs, including the draft BSI standard insists that an operator cannot be enclosed within quotes. I have not implemented this, although it could easily be done. 290  cases by looking for a blank after the "+"; I (and the BSI standard) consider this to be a kludge).  Note that operators may be enclosed in parentheses to have them considered as ordinary strings. For example, ( + )*(:-) = (*)( + , :-).  41  Lists may be entered either in bracket form or dotted form. Here are some examples (IBM and Waterloo Prologs use "!" instead of "|" and also allow curly brackets ("{}") instead of square brackets ("[ ]")): [a  | b]  [a] [a,  = a  .  = [a b]  =  [a,  b | []] b | []]  = a  .  []  = a  . b  .  []  = a  .  (  b  .  [])  Curly brackets are also allowed, for grammar rule notation. Note that here comma is used as an operator rather than a separator. {a} [a,  b}  =  *U'(a)  =  'n'((a,b))  =  ' H ' C V C a . b ) )  B.3 Critique of Prolog's syntax Prolog's syntax badly overloads several characters, particularly  "." (period):  41  •  end of clause  •  "cons" for lists  With the BSI convention, ' * ' ( + 1  1  ,  ':-')  achieves the same affect.  291  •  decimal point in a number  "," (comma): •  argument separator inside predicates, functors and lists  •  and-operator, separating the goals of a predicate  •  ordinary operator, for example inside "[}"  Period is especially troublesome. For example, "1.2.3." can be interpreted as "[ 1,2,(3]" or "[l|2.3]" or "[1.2|3]."  Some other nasty surprises are possible because most "syntactic sugar" is done using monadic and diadic operators. Many programmers have been unpleasantly surprised by what happens with "p->q,r;s,t" (the "... - > ... ; ..." notation is used for  if-then-else).  I propose the following solution: a y>  •  Use only as a decimal point  •  "cons" is represented by the colon ":" (I consider the "[ A|B]" to be hard to read).  •  end of clause can normally be inferred (as in BCPL [Richards and Whitby-Stevens 1979]); in the few places where it cannot be inferred, an explicit "end" is used (alternatively, ";" could be used; "or" probably should use the more obvious "|" symbol instead).  292  it  »  •  Used only as an argument separator.  •  "And" ("&") is used as a goal separator. This "&" can also be an operator.  •  "{a,b}"  is interpreted as  processor for grammar rules because arguments into  - this can easily be handled in a  "'{}'(a,b)"  " {}' (|Args)" 1  will put all the  "Args."  Prolog's method of entering predicates by listing out the clauses makes things a little difficult for a compiler, because the compiler cannot know when a predicate has been completely defined. Normally, all the clauses for a predicate are kept together; in fact, separated clauses usually indicate an error in typing the predicate. Therefore, x p P r o l o g enforces that all the clauses for a predicate appear together.  B.4  Debugging extensions  Misspelling a variable name or predicate name is a very common error. XpProlog  gives an warning message if a variable name is used only once within a  clause (this message can be suppressed by prefixing the name by an underscore). At run-time, if an undefined predicate is tried, an error message is output and the user is prompted for a definition of the predicate ("query the user"). This 42  42  This has not yet been implemented; currently the user is prompted only for a simple success or failure. For a full discussion of "query the user," see [Scrgot 1983]. 293  can be suppressed by defining the predicate as always failing (using the built-in predicate "fail").  One other troublesome point in Prolog is what happens if a predicate is tried with the wrong type of argument. Most such cases can be caught if the programmer indicates that a predicate should always succeed. If the programmer writes "?-neverfall(pred(_,_))," then when "pred/2" fails, an error message is produced. This is done by automatically adding an extra catch-all clause to "pred/2" which produces the error message.  Type inferencing can help in detecting undefined predicates and wrong types to predicates. It is discussed in section 5.10, "Speeding up deterministic predicates  o  - modes and types" on page 112.  294  Appendix C.  Implementation status  An interpreter for the xpPAM has been completed. It includes a simple assembler and loader.  The implementation is incomplete in the following ways:  43  •  Some opcodes are implemented inefficiently. The opcodes should be combined with the operand types to allow faster decoding (as mentioned in section 3.7, "Machine instruction format" on page 36). This would increase the number of cases in the interpreter by a factor of three or four, while approximately doubling its speed.  •  The implementation of delays is inefficient. Currently, a variable which caused a delay has a flag set; when such a variable becomes instantiated, it causes a search for a predicate which can be resumed. A much more efficient way would be to have "uninstantiated variable which caused a delay" to get a separate tag and to have the cell point at a list of predicates which depend on it (as described in section 4.4, "Delays" on page 75).  43  I simply ran out of time and concentrated on writing this report; the remaining work will probably take a few months to finish.  295  Some opcodes have not been implemented:  —  testskip  —  gete  — caseGoto —  cutAt  —  chopBack  —  delayCost  —  mkTHunk  —  delayRec  Only some built-ins have been implemented. A complete list of the existing built-ins is given in Appendix D, "Built-in predicates" on page 300 together with their implementation, to indicate how more built-ins can be easily added. The "virtual choice depth" is not recorded on the backtrack stack (it can easily be added). The implementation uses three stacks. This could be changed to two by combining either the execution and backtrack stacks (the WAM  choice) or  the reset and backtrack stacks (more preference). Note that combining the reset and backtrack stacks requires saving the "virtual choice depth." User defined unification and complex indeterminates have not been implemented.  296  The unification algorithm is recursive, not using the Deutsch-Schorr-Waite algorithm (recursive unification is detected as described in section 3.10, "Unification" on page 40). However, freeing memory does use the D-S-W algorithm.  An optimizing compiler exists only for an earlier version of the abstract machine. Some re-writing is necessary for the present abstract machine.  An un-optimizing compiler has been written. It generates  xpPAM  assembler.  It has not yet been integrated (bootstrapped) with the interpreter. Consequently, predicates must be first run through the compiler, then processed by the interpreter. A fully integrated compiler will implement the compile on first use strategy described in the thesis: a predicate's text is compiled when thefirstattempt is made to execute it and the compiled code replaces the text.  The compilers only produce abstract machine code; they do not compile directly to machine code (as in the chapter 2.0, "Fast append on a conventional machine" on page 15). It is not clear how translation to actual machine code should be done: by using abstract machine instructions as templates and then doing "peephole" optimizations to recognize special cases; or by compiling directly to machine code, using the abstract machine instructions as a guide to the algorithm.  The compilers do not handle "?" notation nor  proceed  must be added by hand to the generated assembler.  297  declarations; they  •  The "closed predicate" notion has not been implemented, nor has negation transformation (section 10.1, "Negation" on page 207).  •  The parser contains code to properly handle Prolog style operators (albeit with slightly different rules than conventional Prolog - see Appendix B, "Details of xpProlog syntax" on page 285.). However, the compilers do not yet handle the if-then-else constructs.  •  The array and I/O predicates described in section 11.0, "Arrays and I/O done logically and efficiently" on page 222 have not been implemented.  •  The compilers do not handle explicit if-then-else. can detect if-then-else  The optimizing compiler  situations and generate appropriate code. The  compilers also do not handle the special else predicate nor the suchthat notation. •  Meta-variables have not been implemented.  •  Object paged virtual memory has not been implemented; the compacting scheme in section 3.15, "Virtual memory" on page 61 has been implemented.  To test the speed of the final interpreter, a number of test programs have been written to test the speed of deterministic append (the inner loop of the naive reverse benchmark). These try the optimized abstract machine (with opcodes and operand types combined) and also directly generated machine code (effectively, removing the interpreter loop and directly executing the code for the 298  various abstract instructions). These have given results in the speed range which has been advertised for some of the better commercial implementations (about 20KLIPS on a 1 MIPS machine), without any attempts to optimize for the particular machine. In addition, the reference counting mechanism has been experimentally turned off, giving approximately 15% speed-up, as suggested by the Smalltalk papers (see section 7.12, "Reference counts and garbage collection" on page 153).  299  Appendix D . Built-in predicates Here is a list of the built-in predicates which have been implemented. The final list will be much larger. These built-ins can be considered as extensions to the basic xpPAM. The following gives definitions of the built-ins, plus their xpPAM code.  Some of these built-ins are non-logical. They exist for compatibility with conventional Prolog, and also as building blocks for logical predicates. call(X)  X must be instantiated to a term which can be interpreted as a goal. The c a l l ( X ) goal succeeds by attempting to satisfy X. X may be an atom, a functor or a list. c o d e ( c a l l , 1, [ ] , [ lCall(O.f)]).  X,Y  % c a l l ( X ) :- X.  The same as c a l l ( X ) , c a l l ( Y ) . code(V,  2, [ ] , [ link, push(1.f) , call(O.f), pop(O), unlink, lCall(O.f)]).  % (X, Y) : - c a l l ( X ) ,  300  call(Y).  The same as c a l l ( X ) , c a l l ( Y ) . The definition is identical to ","  X & Y  (above).  X ;Y  Can be defined by (X;Y) :- c a l l ( X ) . (X;Y) :- c a l l ( Y ) c o d e ( ' ; , 2, [ ] , [ % (X ; Y) :- X. pushB(l.f), mkCh(cl2), lCall(O.f), label(cl2), popB(O), lCall(O.f)]).  (X ; Y) :- Y.  f  /  <goal>  /  This is an internal predicate which is used to handle top-level queries. It is used as template which is modified. The top level of the interpreter produces two structures: one with all the variable names replaced by uninstantiated variables and the other showing the bindings. For example: foo(A,B,A), b a r ( A ) .  generates something like this (note that "," is the "and" predicate given above); the  more  and  nornore  (foo(_5,_7,_5), bar(_5)),  built-ins are described later):  '<more>*(['A ._5, ' B ' . _ 7 ] ) .  This is compiled to:  301  1  /* c o n s t a n t 0: t o p l e v e l g o a l t o c a l l : f o o ( A , B , A ) , b a r ( A ) . */ /* c o n s t a n t 1: l i s t o f v a r i a b l e names ( f o r <more>): ['A'i mkChAt eq link push call pop unlink eq builtin free free stop fin: builtin stop  X = Y  l,fin nO, cO fl fO nl  Jo  remember c h o i c e p o i n t  Jo  c a l l the goal get c h o i c e p o i n t i n t o r e g 1  Jo  a r g f o r <more>  Jo  nO, c l II  5, * B \ _ 7 ] */ Jo must have a c h o i c e p o i n t on % get g o a l i n t o r e g 0  it  more fO fl  "noMore" % always succeeds  unifies X and Y. c o d e ( ( = ) , 2, [ ], [ eq(0.f, l . f ) , return]).  true  Jo X = X.  always succeeds. c o d e ( t r u e , 0, [ ] , [ return]).  fail  % true,  always fails c o d e ( f a l l , 0, [ ] , [ fail]).  % fail  302  :- < f a i l > .  X is Y  Y must be instantiated to a structure that can be interpreted as an arithmetic expression. X is then unified with the result. All arithmetic is done double precision floating point (note that this gives perfectly accurate results for integers up to a few billion). I am investigating further enhancements to is. For example, we may add more constants (by adding to the associative memory) or allow calling user-defined predicates. Is delays if anything in the structure is uninstantiated. c o d e ( i s , 2, [ ] , [ builtin(4), free(l), eq(0.f,2.f), return]).  % Left i s Right % e v a l : r 2 := e v a l ( r l )  The "eval" built-in opcodes expects register 1 to contain a structure to be evaluated; the result is put into register 2 which must be empty (on failure, register 2 is left alone). If there is an uninstantiated variable to be evaluated, the built-in delays (this is slightly inefficient if the expression is complex; the method of defining is as given in section 6.5, "Equality: "is" and "="" on page 133 would be preferable). Here are some further examples of using the "eval" built-in opcode. They implement plus which works if any argument is uninstantiated (delaying if two or more are uninstantiated and doing addition or subtraction otherwise); and  less-than:  303  c o d e ( p l u s , 3, % r O + r l = r2 /*0*/ [const([]), /*!*/ const('+ ), /*2*/ const('-')], [eq(3.n,2.f), % move r2 i n t o r 3 , label(restart), % f r e e i n g r2 varGoto(0, varO), varGoto(l, v a r l ) , % from h e r e , o n l y r3 i s p o s s i b l y u n i n s t a n t i a t e d e q l s t ( 4 . n , l . f , O.c), % r4 = r l . [] e q l s t ( 5 . n , O.f, 4 . f ) , % r5 = rO . r4 % = rO . r l . [] e q l s t ( l . n s , l . c , 5 . f ) , % r l = '+' . r5 % = +'(r0, r l ) builtin(4), % e v a l : r2 := e v a l ( r l ) free(l), eq(2.f, 3 . f ) , return, l  ,  label(varO), varGoto(l, delay), varGoto(3, delay), % from h e r e , o n l y rO i s p o s s i b l y u n i n s t a n t i a t e d e q l s t ( 4 . n , l . f , O.c), % r4 = r l . [] label(vl), e q l s t ( 5 . n , 3.f, 4 . f ) , % r5 = r3 . r4 % = r3 . r l . [] e q l s t ( l . n s , 2.c, 5 . f ) , % r l = '-' . r5 % = '-'(r3,rl) builtin(4), % e v a l : r2 := e v a l ( r l ) free(l), eq(2.f, O.f), return, label(varl), varGoto(3, delay), % from h e r e , o n l y r l i s p o s s i b l y u n i n s t a n t i a t e d e q l s t ( 4 . n , O.f, O.c), % r4 = rO . [] eq(0.n, l . f ) , goto(vl),  304  label(delay), nonvarGoto(0, d e l a y O ) , delay(0, r e s t a r t ) , label(delayO), nonvarGoto(1,delayl) , delay(1,restart), label(delayl), nonvarGoto(3, d e l a y 3 ) , delay(3, r e s t a r t ) , label(delay3), builtin(3)]).  % error/3 % ( s h o u l d n ' t happen)  c o d e ( ( < ) , 2, % A < B :- 1 i s (A < B) [const([]), c o n s t ( l ) , const(<)], [ e q l s t ( 3 . n , l . f , O.c), % r 3 = B . [] e q l s t ( 2 . n , O.f, 3 . f ) , % r2 = A . r3 % = A . B . [ ] e q l s t ( l . n s , 2.c, 2 . f ) , % r l = '<' . r 2 % = '<'(A,B) builtin(4), % e v a l : r 2 := e v a l ( r l ) free(l), eq(2.f, l . c ) , % r2 = 1 ? return]).  X \= Y  Succeeds if X and Y cannot be unified. It delays if either argument is insufficiently instantiated. c o d e ( ( \ = ) , 2, [ ] , [ builtin(5), free(O), f r e e ( l ) , return]).  % noteq: rO \= r l % noteq  This is an old form which should be replaced by:  305  code((\=), 2, [ ] ,  [  t e s t s k i p ( 0 . f , l . f ) , % delays i f necessary fail % success becomes f a i l u r e return]). % f a i l u r e becomes success  Note that not builtin(5))  equal  must be inside a predicate. The testskip (or  will cause its enclosing predicate to delay, which gives the  appearance of replacing the opcode by a return. Thus, if the not equal  were in-line, it would cause execution to not consider any of  the following goals until the  not equal  had been resumed.  name(A,L) The characters for the atom A are the list L. For example, name(ab, [ a , b ] )  or name(ab,"ab"). Note that this is different from  C-Prolog which puts the ASCII codes into the list but x p P r o l o g puts the individual characters in. code(name, 2, [ ] , [ varGoto(0, varO), builtin(6), free(O), eq(l.f,2.f), return, label(varO), varGoto(l, varOl), builtin(7), free(l), eq(0.f,2.f), return, label(varOl), delayOr(0,0), delay(l.O)]).  306  % name(String,  List)  % r 2 := s t r T o L i s t ( r l )  % r 2 := l i s t T o S t r ( r O ) ^  The built-in opcode "strToList" turns a string into a list. It expects a string in register 0 and assigns the list expansion of the string to register 2.  The built-in opcode "HstToStr" turns a list into a string. It expects a list in register 1 and assigns the string value to register 2.  include F  switches input to be the file F. This will probably change to the C-Prolog "consult."  audit  audits the memory to look for wrong reference counts, etc. (This is not a true predicate right now.)  code(N,NP,C,M) defines the name N with NP parameters by the code defined by the constant list C and the machine code list M. The format is given in an appendix. The old definition, if any, is replaced. Note that the code segment constants always have one level of indirection so that replacing a predicate automatically changes it everywhere it is used. If a predicate is used which hasn't yet been defined, the predicate is automatically defined to produce a warning message ("Undefined predicate") and then fail. A "query the user" facility can be added by simply replacing the default "undefined" predicate. code(code, 4, [ ] , [ builtin(8), free(P), free(l), return]).  % code(Name, A r i t y , C o n s t L i s t , % MachCodeList) % code(rO,rl,r2,r3) free(2), free(3),  307  delcode(N,NP) removes the definition of the name N with NP parameters. Again, because code segments are always handled with one level of indirection, this removes the definition everywhere (it is actually replaced by the "Undefined predicate" warning). code(delcode, 2 , [ ] , [ bulltin(9), free(O), f r e e ( l ) , return]).  % delcode(Name, A r i t y ) % delcode(rO,rl)  dbgFlag(X) turns off debugging flag " X " ("-dx" in the command line can set this on initially). "dbgFlag(' ')" turns on the general debug flag. About the only other useful thing is d b g F l a g ( t ) and sometimes d b g F l a g ( i ) to do tracing. c o d e ( d b g F l a g , 1, [ ] , [ builtin(lO), free(O), return]).  % dbgFlag(Char) % dbgFlag(rO)  nodbgFlag(X) turns on debugging flag " X . " % nodbgFlag(Char) 1 nodbgFlag(rO)  code(nodbgFlag, 1, [ ] , [ builtln(ll), free(O), return]).  op(P,A,0) defines operator O with priority P and associativity A (see section B.2, "Lexical and syntactic details" on page 286 for parsing and the standard operators).  308  code(op, 3, [ ] , [ % op(Prio,Assoc,Name) builtin(12), % op(r0,rl,r2) free(O), f r e e ( l ) , free(2), return]).  tableAssign(T,I,V) assigns value V to the associative table T at index I. V can be anything at all (it should be instantiated or else strange things will happen); T and I must be strings (atoms). c o d e ( t a b l e A s s i g n , 3, [ ] , [ % assign(Table,Item,Val) builtin(13), % assign(rO,rl,r2) free(O), f r e e ( l ) , free(2), return]).  tableFind(T,I,V) finds value V in the associative table T at index I. T and I must be strings (atoms). c o d e ( t a b l e F i n d , 3, [ ] , [ builtin(14), free(O), f r e e ( l ) , eq(2.f,3.f), return]).  % tableFind(Table,Item,Val) % e v a l : r 3 := f i n d ( r O , r l )  tableDelete(T,I) deletes the associative table T at index I. T and 1 must be strings (atoms). c o d e ( t a b l e D e l e t e , 3, [ ] , [ builtin(15), free(O), f r e e ( l ) , return]).  % tableDelete(Table,Item) % delete(rO,rl)  Incidentally, a backtracking version of tableAssign would be:  309  t a b l e A s s i g n B ( T , I , V ) :i f e x i s t s VOld s u c h t h a t t a b l e F i n d ( T , I , V 0 1 d ) then (tableAssign(T,I,V); tableAssign(T,I,VOld)) else (tableAssign(T,I,V); tableDelete(T,I)) endif.  The semi-colon (";") is an or. The first time in, the t a b l e A s s i g n ( T , l , V ) is executed. On backtracking, the tableAssign(T,I,Void)  or t a b l e D e l e t e ( T , I ) is executed to undo the  affect of the original t a b l e A s s i g n .  put(Item) output one item on the output "stream." No quote marks are added ( d i s p l a y adds  quote marks).  code(put, 1, [ ] , [ builtin(16), free(O), return]).  nl  % put(Item) % put(rO)  output a new line. c o d e ( n l , 0, [ c o n s t ( ' \ n ' ) , c a l l ( p u t , 1 ) ] , [ eq(0.n,0.c), lCall(l.c)]). % put('\n')  % n l ( )  display(Item) display an item, in such a way that it can be read back in (that is, with quote marks added as necessary). c o d e ( d i s p l a y , 1, [ ] , [ builtin(17), free(O), return]).  % display(Item) % display(rO)  310  D.l  Built-in opcodes.  A built-in opcode is not normally used directly; rather it is embedded inside a "shell" predicate which sets up the registers for it. After each built-in is given the shell predicate (often with the same name) which set it up.  The built-in opcode is used to extend the instruction set of the basic inference machine. A built-in may succeed, fail or delay, not necessarily in the same way as its "shell" predicate.  Most of the built-in opcodes have been given above.  Undef:  an undefined predicate gets set to the following code: b u i l t - i n "under" fail % always does t h i s  The  undef builtm  uses information from the current code segment to  display information about the predicate. The built-in always succeeds and frees all the registers that were set up for the call (again, by using information from the current code segment). The instruction following the built-in is a fail, so the predicate fails (the built-in was made to succeed so that it could also be used for a query-the-user facility). More  is used with '<goal>' to prompt the user for more solutions.  More  expects register 0 to contain a list of variables (in the form [namel.valuel,name2.value2 , ...]) and register 1 to contain the original choice point.  More  displays the values. 311  •  If the list of variables is empty, it succeeds.  •  If the backtrack stack is empty (no other solutions possible), it succeeds.  •  If input is from the terminal, it waits for the user to type in either a semi-colon (";") or just < return '<more>'  >  . In the former case, -  fails to cause backtracking; in the latter case, it succeeds  so that no more solutions are produced. The initial choice point must have been created by '<goal>' so that a final backtrack has somewhere to go to. •  If input is from an included file, it either succeeds of fails depending on the value of a user-settable flag — if the flag is on, backtracking will generate all possible answers.  NoMore  takes no registers. It displays "No (more) solutions." and succeeds. It should always be followed by a stop instruction.  Error  uses information from the current code segment to display information about the predicate. This built-in is used directly in user code; typically for a case from a swXVNL argument. This built-in always fails.  312  instruction for an invalid  The end  This page intentionally left blank, to indicate the end of the thesis.  313  


Citation Scheme:


Citations by CSL (citeproc-js)

Usage Statistics



Customize your widget with the following options, then copy and paste the code below into the HTML of your page to embed this item in your website.
                            <div id="ubcOpenCollectionsWidgetDisplay">
                            <script id="ubcOpenCollectionsWidget"
                            async >
IIIF logo Our image viewer uses the IIIF 2.0 standard. To load this item in other compatible viewers, use this url:


Related Items