Hope (programming language)
This page was last modified on 1 June 2016, at 17:17.
Paradigm | procedural |
---|---|
Designed by | Computer Technology Department of the University of Edinburgh, Scotland |
First appeared | 1970-1980 |
Typing discipline | Strong, dynamic |
Website | None |
Influenced by | |
LISP, NPL, ISWIM, PROLOG |
Hope – is a functional programming language developed in the 1970s at Edinburgh University.
Contents
Review
An applicative language called HOPE is described and discussed. The underlying goal of the design and implementation effort was to produce a very simple programming language which encourages the construction of clear and manipulable programs. HOPE does not include an assignment statement; this is felt to be an important simplification. The user may freely define his own data types, without the need to devise a complicated encoding in terms of low-level types. The language is very strongly typed, and as implemented it incorporates a typechecker which handles polymorphic types and overloaded operators. Functions are defined by a set of recursion equations; the left-hand side of each equation includes a pattern used to determine which equation to use for a given argument. The availability of arbitrary higher-order types allows functions to be defined which 'package' recursion. Lazily-evaluated lists are provided, allowing the use of infinite lists which could be used to provide interactive input/output and concurrency. HOPE also includes a simple modularisation facility which may be used to protect the implementation of an abstract data type.
Introduction
As hardware gets cheaper, people become increasingly aware of the high cost of software and particularly of those components of the cost which are due to bugs in programs and to inflexible programs. A simple yet powerful programming language in which one has a good chance of avoiding mistakes and in which programs are transparent enough to be easily altered would be an effective means of countering this trend. We believe this goal is worth working towards even if it requires radical changes in our programming habits.
In this paper we will first list some aspects of programming languages which contribute to accuracy and flexibility. We illustrate them by describing an applicative language, called HOPE, which we have developed and implemented. This is followed by an example of a complete HOPE program. Finally, we discuss the strengths and weaknesses of HOPE.
Desirable programming language features
The following points seem important in a language for writing correct and flexible programs. They are listed rather briefly but they are illustrated in the description of HOPE which follows.
No assignment: referential transparency
Applicative languages which work in terms of expressions and their values, using recursion instead of loops, seem much clearer and less error-prone. The expressions are transparent in the sense that their value depends only on their textual context and not on some notion of computational history. Each variable is given a value just once where it is declared. Assignments which alter data structures are particularly prone to cause bugs; disallowing them greatly simplifies the language although it does slow down execution.
Maximum use of user-defined types
The user should define his own types whenever possible; thus type 'age' rather than type 'integer'. The machine must check these types. The language should allow polymorphic types so that code can be as general as possible; for example, 'list of alphas' rather than specifically 'list of numbers', where alpha is a type variable which can be instantiated to any type. We suspect that a great many errors are caused by the complications introduced when encoding data in terms of the commonly-supplied low-level types; the provision of a simple and powerful facility for defining types should greatly simplify the programmer's task.
Overloaded operators
It is convenient to use common operation symbols such as "+" or "eval" with a variety of meanings, according to the type of the arguments. The implementation should be able to disambiguate such 'overloaded' operators (this requires some subtlety for polymorphic types).
Exhaustive case analysis
Each data type will have a set of disjoint subtypes, each with a different constructor function; e.g., lists are made with cons or nil It should be easy to do case analysis with respect to these constructors and easy for a compiler to check whether the analysis is exhaustive. This avoids mistakes like forgetting to include a test for the empty list.
Avoiding unnecessary names
It should not be necessary to introduce a large number of different names for functions to construct data structures (e.g. cons ), to test which subtype it is (e.g. the predicate null ), or to select components (e.g. car and edr ). We can use just the constructor name and by using pattern matching syntax let it denote testing and selecting as the context dictates.
Abolishing explicit recursion as far as possible
Given that we want an assignment-free language, we will use recursion instead of loops. However, undisciplined use of recursion is not at all desirable. It is much better to use standard constructions related to the data structures; for example, { x2 I x ∈ S } as in mathematics (and SETL [12]) instead of recursion over the elements of S. In a language which allows higher-order functions (that is, procedures taking procedures as parameters), we can easily define useful higher- order functions which 'package' recursion to save writing it explicitly. An example is mapcar in LISP.
Making secuences into data object
In traditional imperative-style programming, sequences are sometimes represented as successive values of a variable; thus a sequence of integers is represented as values of an integer variable. It is also possible to represent them by an array or list, but this has the disadvantage that all the elements of the sequence must be in memory at the same time. An idea aptly called 'lazy evaluation' by Henderson and Morris [18] (due originally to Wadsworth [29]) enables us to deal with such sequences as lists but have them evaluated one element at a time as needed. Amongst other things, this enables one to achieve the same effect as coroutines in an imperative language but in a more perspicacious way, by passing a lazy list as the result of a function. It also enables one to handle interactive input/output and concurrency in an applicative language.
Modularity and data abstraction
The main trick in writing large bug-free programs is to construct them in smallish pieces, each as self-contained as possible, with the pieces communicating with each other in a disciplined, explicit manner. An important technique for achieving this is data abstraction (see for example Hoare in chapter 2 of [10]). This has been explored as an aid to good programming style in such recent languages as ALPHARD, CLU and ADA, which may be thought of as developing one aspect of the SIMULA class concept.
An abstract data type, such as 'set' or 'queue', is defined by a collection of procedures which create elements of the type and operate on them. The implementations of these procedures and the representation of the type itself are hidden from the user; his code relies solely on the specified properties of the abstract type.
Conventional block structure scope rules make it difficult to maintain the desired separation of use and implementation, so these languages adopt different rules for identifier scopes.
Description hope language
A programming language called HOPE which illustrates the features mentioned above has been designed and implemented at Edinburgh University. A brief informal description of HOPE follows; full details can be found in. A precursor, NPL, was described in.
The aim throughout was to design a programming language which was very simple and encouraged clarity and manipulability of programs. Major influences in the design of HOPE have been LISP and Landin's ISWIM. We were not trying to be original; we sought a judicious selection of well- understood ideas. HOPE seeks some blend of LISP power with the discipline of strong typing and modularity. It bears some resemblance to a number of other languages, including PROLOG, ML, SASL, OBJ, SCRATCHPAD, and languages by Burge and Backus.
HOPE (in its present form) is an experiment in language design and a means of testing certain ideas in programming methodology rather than the ultimate in programming languages.
Part of the HOPE implementation was carried out by Michael Levy and we express our appreciation of his efforts.
Data declarations
Conceptually, all data in HOPE is represented as terms consisting of a data constructor applied to a number of subterms, each of which in turn represents another data item. The tips of this tree are nullary data constructors or functional objects. An example is succ(suce(0)) in which suec is a unary constructor and 0 is a nullary one (i.e., a constant). Constructor functions are uninterpreted; they just construct.
A data declaration is used to introduce a new data type along with the data constructors which create elements of that type. For example, the data declaration for natural numbers would be:
data num == 0 ++ succ (num)
defining a data type called num with data constructors 0 and suco . So the elements of num are 0 , suco(O) , succ(succ(O)) .... ; that is, O, 1, 2, ....
To define a type 'tree-of-numbers' we could say
data numtree == empty ++ tip (num)
++ node (numtree # numtree)
the sign # gives the cartesian product of types.
One of the elements of numtree is:
node (tip(succ(0)),
node (tip(succ(succ(0))), tip (0)))
But we would like to have trees of lists and trees of trees as well, without having to define them all separately. So we declare a type variable typevar alpha
which when used in a type expression denotes any type (including second- and higher-order types). A general definition of tree as a parametric type is now possible:
data tree (alpha) == empty ++ tip (alpha)
++ node (tree (alpha) # tree (alpha))
Now tree is not a type but a unary ~vne constructor -- the type numtree can be dispensed with in favour of tree(hum).
Another example of a data declaration is
Data graph == mkg(set vertex
# (vertex#vertex->truval))
This says that a graph is (the data constructor mkg applied to) a set of vertices together with a binary relation which tells if there is an edge between any two vertices.
HOPE currently comes equipped with the data types hum , truval , char , list , and set .
Expressions
The simplest expressions of HOPE are constants (i.e., data constructors and functions -- the 'usual' concept of a constant is just the class of nullary functions and data constructors) and variables.
An application may be formed by simply Juxtaposing two expressions:
factorial 6
For functions of several arguments we use tuoles, formed with commas; thus 3,4 is a 2-tuple. Parentheses are used for grouping, for example:
g (3,4)
In the expression:
(f x) y
the subexpression f x would have to produce a function; thus the types would be f : T1 -> T2 -> T3 with x : T1 and y : T3.
It is possible to use function symbols as infix or postfix operators if they are declared and given a precedence; for example:
infix +, - : 8
A similar form is used to assign a precedence to a prefix symbol.
Some convenient notations have been implemented for built-in types; thus e1:: (e2:: …. :: nil)
is abbreviated [el,e2, ...] , ['a','b', .] is "ab..." and sets are written {el,e2, ...} . Note that we write cons as infix :: .
There are two equivalent forms of conditional expression:
; option 1
e1 if c else e2
; option 2
c then e1 else e2
in many languages written if c then e1 else e2 .
Local variables may be introduced and associated with values using either of the equivalent forms:
e1 where p == e2
; or
let p == e2 in e1
where p is an expression formed by application of data constructors to a number of distinct variables (this is called a pattern). For example:
a+b where a:: (b:: l) == f (t)
Upon evaluation, f(t) is expected to yield a value which 'matches' the pattern a::(b::l) . The corresponding subterms in the value of f(t) are then bound to a , b , and 1 while evaluating a+b .
Defining function
Before a function is defined, its type must be declared. For example:
dec reverse : list (alpha) -> list (alpha)
HOPE is a very strongly-typed language, and the HOPE system includes a polymorphic typechecker (a modification of the algorithm in [27]) which is able to detect all type errors at compile time. Function symbols may be overloaded. When this is done, the typechecker is able to determine which function definition belongs to each instance of the function symbol.
Functions are defined by a sequence of one or more eeuations, where each equation specifies the function over some subset of the possible argument values. For example:
--- reverse (nil ) <= nil // (1)
--- reverse (a : : l) <= reverse (l ) <> [a] // (2)
(the symbol <> is infix append ). This defines the (top-level) reverse of a list; for example:
reverse (1 : : (2 : : nil ) ) = reverse (2 : : nil ) <> [1]
= (reverse (nil ) <> [2] ) <> [1]
= (nil <> [2] ) <> [1]
So reverse [I,2] = [2,1] (by two applications of equation 2 followed by a single application of equation I). The left-hand-side patterns will normally be disjoint and be related to the structure of the type definition:
data list alpha == nil ++ alpha :: list alpha
The set of equations defining a function should exhaust the possibilities given in the data- statement introducing the argument types.
For example, a definition of the Fibonacci numbers:
dec fib : hum -> num
--- fib (0 ) <= 1
--- fib (succ (0 ) ) <= 1
--- fib (succ (succ (n ) ) ) <= fib (succ (n ) ) + fib (n )
In this case the three patterns 0 , succ(O) , and succ(succ(n)) exhaust the set of values belonging to num. The pattern ; may be used as shorthand for succ(O) .
Nullary 'functions' may also be defined; for example:
dec pi : rational
--- pi <= mkrational (22,7)
which assumes that the type rational has been defined. Lambda-expressions are defined similarly. For example, a function to compute the conjunction of two truth values (already available as the function and ):
lambda true, p => p
I false, p => false
Another example of a lambda-expression occurs in the definition of functional composition:
typevar alpha,beta,tau
dec compose : (alpha->beta)#(beta -> tau )
-> (alpha -> tau )
--- compose (f, g) <= lambda x => f (g (x ) )
Modules
Any sequence of statements may be made into a module by surrounding it with the statements:
module name
and
end
Data types defined in a module may be referred to outside only if a statement pubtype tname
is included in the module.
Similarly, constants (including data constructors) may be referenced only if a statement pubconst cname
is included.
Nothing defined outside a module may be referenced within it, unless the module includes the statement uses mname
.
In this case, all of the types and constants declared as public to the indicated module are available. In addition, certain global types and constants (num, truval , char , list , and set together with some primitive operations) may be referenced within any module.
This is an effective tool for the encapsulation of data abstractions; if the primitive constructors and low-level operations on the data representation are not declared public, then the implementation of the abstraction is hidden from the rest of the program.
Lazy evalution and input/output
Lazy evaluation was mentioned briefly in section 2; there we described it as a desirable language feature. In HOPE, lazy evaluation has thus far only been implemented for lists, which makes available most of the power of the mechanism. Lazy lists are created using the special constructor icons in place of the usual :: . This constructs a list which is identical to a normal list (in particular, the pattern a::l will match icons(c,d) ) except that the arguments to icons remain unevaluated until their values are required during subsequent use of the list.
Consider the function which produces the (potentially) infinite list consisting of a number and all its successors:
dec allsuccs :num -> list num
--- allsuccs (n ) <= lcons (n, allsuccs (n+1 ) )
Now if we define a 'lazy mapcar' function:
Typevar alpha, beta
dec <*> : ( alpha-> beta)#list alpha -> list beta
infix <*> : 6
--- f <*> nil <= nil
--- f <*> (a::al) <= lcons (f (a ), (f <*> al ) )
then the infinite list of the squares of all numbers is just:
square <*> allsuccs (0)
We can use lazy lists to provide interactive input/output in HOPE. The first step would be for the system to provide two new functions, input and output (not yet available). Input takes an argument of type device specifying an input device, where and produces a lazy list where new items are read from the device when needed.
data device == tty ++ file(list(char)) ++ ...
A special character (control/Z) signals end of input, and consequently the end of the lazy list. Output is a function of type device->( list (alpha)->void)
. Given a list, output(dev) evaluates each element and produces it as output on the indicated device.
As a simple example, here is a program to read a number from the terminal, output its square, and then repeat (until control/Z is typed):
output ("tty") (square <*> input("tty"))
Since input("tty") in this expression is of type list(num) , the teletype will accept as input only expressions of type num.
This is a special example of the general problem of communication between concurre, processes. Lazy lists provide a means of implementing communication channels, but at present the notation of HOPE is not adequate to specify an arbitrary network of communicating processes. We have not yet given much thought to this problem; however, notations exists which are sufficient and which could be adapted for our purposes (see for example).
Implementation notes
The HOPE system consists of a compiler (from HOPE programs to code for an abstract stack machine) and an implementation of the target machine. The system is written in POP-2, and currently runs in approximately 55K words (plus a 15K shareable segment) on a DEC KI-IO. The compiled code should be quite portable, as the abstract machine simulator could easilybe written for a new machine. The compiler itself is less portable since it requires availability of POP-2.
Our preliminary tests indicate that a program written in HOPE runs approximately 9 times slower than the same algorithm coded in LISP running under the Rutgers/UCI interpreter (and 23 times slower than compiled LISP). A large program run recently ran more slowly, presumably because of page thrashing. A machine code implementation of the interpreter should run a lot faster.
Example
An example of a complete HOPE program is given in figure I. This illustrates how we can use HOPE to implement a data type (ordered trees), and then how that type can be used in a program for treesort.
Ordered trees
The first module contains an implementation of the abstract type ordered-tree-of-numbers (data type otree in the program). An otree is defined to be either empty , a tip (containing a number), or a node containing two otrees and a number. The special property of otree is that for any term node(tl,n,t2) , all numbers contained in tl are less than n , which is in turn less than or equal to all numbers contained in t2 .
We define three public constants:
- empty – the empty otree
- insert – adds a number to an otree, preserving the 'orderedness' of the otree
- flatten – inorder traversal of an otree
Ordinarily an abstract data type would have a few more operations; we have only included those which are used in the remainder of the program.
Note that the data constructor node is not public. Consequently, the only functions available to the 'outside world' for constructing and modifying otrees are empty and insert . Both of these preserve the properties of otrees, so the integrity of the implementation is assured. However, insert is not a data constructor, and hence may not be used in patterns.
Example complete HOPE program:
Module ordered_trees
pubtype otree
pubconst empty, insert, flatten
date otree == empty ++ tip(num)
++ node(otree#num#otree)
dec insert : num#otree -> otree
dec flatten : otree -> list num
--- insert(n,empty) <= tip(n)
--- insert(n,tip(m))
<= n<m then node(tip(n),m,empty)
else node(empty,m,tip(n))
--- insert(n,node(tl,m,t2))
<= n<m then node(insert(n,tl),m,t2)
else node(tl,m,insert(n,t2))
--- flatten(empty) <= nil
--- flatten(tip(n)) <= [n]
--- flatten(node(tl,n,t2))
<= flatten(tl) <> (n::flatten(t2))
end
module list_iterators
pubconst*, **
typevar alpha, beta
dec * : (alpha->beta)#list alpha -> list beta
dec ** : (alpha#beta->beta)#(list alpha#beta)
-> beta
infix *, ** : 6
--- f * nil <= nil
--- f * (a::al) <= (f a)::(f * al)
--- g ** (nil,b) <= b
--- g ** (a::al,b) <= g ** (al,g(a,b))
end
module tree_sort
pubconst sort
uses ordered_trees, list_iterators
dec sort : list num -> list num
--- sort(l) <= flatten(insert ** (l,empty))
end
List iterators
This module defines two second-order functions which apply a given function to every element of a list and collect the results. These two functions are representatives of a group of functions which are widely used in HOPE programs in an attempt to eliminate explicit recursion as far as possible.
The first function, * , is identical to mapear in LISP. It produces a list containing the results of applying the function supplied to each element of the given list. This operation is not actually used in the example.
The function ** is slightly more complicated. When supplied with a function g of alpha#beta -> beta
, a list of alphas, and an 'initial' beta-object, it applies g to each element of the list, beginning with the given beta-object as a second argument and subsequently recycling the result of the previous application. This operation is analogous to the 'reduction' operator of APL [21]; an example of its use would be to compute the union of a list of sets:
union ** (setlist,emptyset)
In this case, the module facility is used as a means of packaging a number of related functions rather than as a device for protecting a delicate abstraction. However, if one of the operations requires an auxiliary function which has no utility of its own, then it might be desirable to keep this function local to the module.
Tree sort
A function for sorting a list of numbers is now defined using the primitives developed in the preceding modules. The ** operation from llst_iterators is used to successively insert the llst elements into an initially empty otree . The result is then flattened to produce the final answer.
Conclusion
Some desirable features for programming languages were discussed in section 2. Since HOPE was designed with these in mind, it not surprisingly displays them all, although some are exhibited better than others (an exception is set expressions, still to be implemented). But more important is the question of whether the resulting language provides a better interface with the computer than the one presented by existing languages.
Probably the most pleasing result of our efforts has been the observation that it really is significantly easier to construct programs in HOPE than in any other programming language we know. In particular, we have found that it is rather easy to write programs which are absolutely correct the first time they are run. It seems quite difficult to commit an error which remains undiscovered for long -- the simple errors are caught during compilation by the typechecker, while the more fundamental errors (stemming usually from an insufficient understanding of the problem) display themselves glaringly during even a casual test. Since testing and debugging account for more than half the cost of many software projects, this could yield enormous savings.
An important aim of language design is to make it easier to verify that a program meets a given specification. In this respect applicative languages such as HOPE seem to offer considerable advantages; the absence of assignment statements and the consequent replacement of iteration by recursion gives programs a simple and easy to analyse form. Powerful verification systems for applicative languages have been written by Boyer and Moore and by Aubin.
Another advantage of an applicative language is the fact that programs lend themselves very well to the technique of program transformation [7], whereby a simple but inefficient program is transformed into an acceptably efficient one by steps which maintain its correctness. A very simple example of program transformation would be the production of the following linear-time program for generating Fibonacci numbers from the equivalent program in section 3.3 which requires exponential time.
dec g :num -> num#nnm
--- g(O) <= 1,1
--- g(succ(n)) <= (a + b),a where a, b = = g (n)
dec fib' : hum -> num
--- fib'(O) <= 1
--- fib'(1) <= 1
--- fib'(succ(succ(n)))
<=a+b where a, b = = g (n)
In addition, there is another (as yet unmentioned) advantage of applicative languages which may come to our rescue: applicative languages are not so tightly bound to the notion of a sequential machine as are imperative languages. The value of the function applicatione0 (e1...., en)
is independent of the order of evaluation of the expressions eo, ... ,en (if parameters are passed 'by value'); ~his is guaranteed by the absence of an assignment statement. If a parallel machine is available, c0, ... ,en may be evaluated simultaneously. Not onIy that, but if c0, ... ,en are themselves function applications, then their arguments may all be evaluated simultaneously. With the rapidly falling cost of hardware, this is feasible, although we would have to build the machine ourselves.
HOPE has faults, too; one is illustrated in figure 1. The sorting program will only sort a list of numbers, because otree is 'ordered-tree-of- numbers. We want a more general sorting program, and this depends on a more general definition of ordered trees; we would like to define 'ordered- tree-of-alphas'. The data declaration is easy to generalise. But to generalise insert to type:
alpha#otree(alpha)->otree(alpha)
we must have a more general order relation than < , which is defined only for numbers.
One solution to this problem would be to require otree users to supply the appropriate order relation explicitly when dealing with an otree. It could be added as an extra argument to insert, or alternatively it could be built into the trees themselves (supplied as an argument to empty , and propagated to new tips and nodes by insert ). Unfortunately, a 'bad' order relation (for example, one which is not transitive) would violate the integrity of the data type, causing unpredictable results.
Another possibility would be to define a general order relation as a new HOPE primitive. This was our solution to the analogous problem involving the equality relation = . Two data values are equal if they have the same representation as terms of constructors, except that for a newly-defined data type the user may provide a nonstandard equality which is automatically incorporated into the standard system equality. For example, we might want to define equality for otrees so that tip(n) = node(empty,n,empty) . Allowing an order to be associated with each type would be a satisfactory solution to our present difficulties. But how many more such operations should be associated with a type?
The best solution is of course to associate a collection of operations with each data type (so types become al~ebras instead of simply sets). Rather than generalising to otree(alpha) we could generalise to otree(alpha[<]) , requiring an order relation to exist on the parameter type. This is the approach taken in CLU [25] and in the specification language CLEAR [8]. We really want HOPE modules to have ~r.~_~r_~., a collection of types and operators, just as CLU clusters have parameters.
As a further example, refer again to figure I and note that the module tree_sort does not depend on the fact that otrees are trees, but just on certain properties of insert and flatten . We may substitute a module ordered_lists for ordered_trees , where empty becomes nil , insert becomes the obvious order-preserving insertion in an ordered list, and flatten is the identity function. Essentially, tree_sort is a parametrised module which may be 'applied' to any module satisfying certain (nontrivial) properties.
The CLU cluster parameter, however, is just an explicit list of types and operators:
T with < : T#T -> truval
In CLEAR such an entity (called a theory and provided with axioms) can be named, thus 'Ordered-Set'; what is more various operations for building such theories are provided. We would llke to extend HOPE to have parameterised modules where the parameters and the interfaces between modules are nameable and manipulable as in CLEAR.
Links
References
- Aubin, R. Strategies for Mechanizing Structural Induction. Proc. 5th Int. Joint Conf. on Artificial Intelligence, Cambridge, Massachusetts, August, 1977, pp. 363-369.
- John Backus, Can programming be liberated from the von Neumann style?: a functional style and its algebra of programs, Communications of the ACM, v.21 n.8, p.613-641, Aug. 1978
- Boyer, R.S. and Moore, J.S. A Computational Logic. Academic Press, 1980
- Burge, W.H. Recursive Programming Techniques. Addison-Wesley, 1975.
- Burstall, R.M. Design Considerations for a Functional Programming Language. Infotech State of the Art Conference: The Software Revolution, Copenhagen, October, 1977.
- Rod M. Burstall, Electronic Category Theory, Proceedings of the 9th Symposium on Mathematical Foundations of Computer Science, p.22-39, September 01-05, 1980
- R. M. Burstall , John Darlington, A Transformation System for Developing Recursive Programs, Journal of the ACM (JACM), v.24 n.1, p.44-67, Jan. 1977
- Burstall, R.M. and Goguen, J.A. Putting Theories Together to Make Specifications. Proc. 5th Int. Joint Conf. on Artificial Intelligence, Cambridge, Massachusetts, August, 1977, pp. 1045-1058.
- Burstall, R.M. and Sannella, D.T. HOPE User's Manual. In preparation.
- O. J. Dahl , E. W. Dijkstra , C. A. R. Hoare, Structured programming, Academic Press Ltd., London, UK, 1972
- Dahl, O-J., Myhrhaug, B. and Nygaard, K. The SIMULA 67 Common Base Language. Publication S22, Norwegian Computing Centre, Oslo, 1970.
- Robert B. K. Dewar , Arthur Grand , Ssu-Cheng Liu , Jacob T. Schwartz , Edmond Schonberg, Programming by Refinement, as Exemplified by the SETL Representation Sublanguage, ACM Transactions on Programming Languages and Systems (TOPLAS), v.1 n.1, p.27-49, July 1979
- Feather, M.S. A System for Developing Programs by Transformation. Ph.D. Th., University of Edinburgh, 1979.
- Goguen, J.A. and Burstall, R.M. CAT, a System for the Structured Elaboration of Correct Programs from Structured Specifications. In preparation.
- Goguen, J.A. and Tardo, J.J. An Introduction to OBJ: A Language for Writing and Testing Formal Algebraic Program Specifications. Specifications of Reliable Software Conf. Proc., Cambridge, Massachusetts, April, 1979.
- Michael J. C. Gordon, The Denotational Description of Programming Languages: An Introduction, Springer-Verlag New York, Inc., Secaucus, NJ, 1979
- M. Gordon , R. Milner , L. Morris , M. Newey , C. Wadsworth, A Metalanguage for interactive proof in LCF, Proceedings of the 5th ACM SIGACT-SIGPLAN symposium on Principles of programming languages, p.119-130, January 23-25, 1978, Tucson, Arizona
- Peter Henderson , James H. Morris, Jr., A lazy evaluator, Proceedings of the 3rd ACM SIGACT-SIGPLAN symposium on Principles on programming languages, p.95-103, January 19-21, 1976, Atlanta, Georgia
- Henderson, P. and Snowdon, R. An Experiment in Structured Programming. BIT 12, 1 (1972), 38-53
- J. D. Ichbiah, Preliminary Ada reference manual, ACM SIGPLAN Notices, v.14 n.6a, p.1-145, June 1979
- Kenneth E. Iverson, A programming language, John Wiley & Sons, Inc., New York, NY, 1962
- R. D. Jenks, The SCRATCHPAD language, Proceedings of the ACM SIGPLAN symposium on Very high level languages, p.101-111, March 28-29, 1974, Santa Monica, California, USA
- Brian W. Kernighan , P. L. Plauger, Software Tools, Addison-Wesley Longman Publishing Co., Inc., Boston, MA, 1976
- P. J. Landin, The next 700 programming languages, Communications of the ACM, v.9 n.3, p.157-166, March 1966
- Barbara Liskov , Alan Snyder , Russell Atkinson , Craig Schaffert, Abstraction mechanisms in CLU, Communications of the ACM, v.20 n.8, p.564-576, Aug. 1977
- George Milne , Robin Milner, Concurrent Processes and Their Syntax, Journal of the ACM (JACM), v.26 n.2, p.302-321, April 1979
- Milner, R. A Theory of Type Polymorphism in Programming. Journal of Computer and System Sciences 17, 3 (December 1978), 348-375.
- Turner, D.A. SASL Language Manual. University of St. Andrews, 1979.
- Wadsworth, C.P. Semantics and Pragmatics of the Lambda Calculus. Ph.D. Th., Programming Research Unit, Oxford University, 1971.
- David H D Warren , Luis M. Pereira , Fernando Pereira, Prolog - the language and its implementation compared with Lisp, Proceedings of the 1977 symposium on Artificial intelligence and programming languages, p.109-115, August 15-17, 1977
- Wulf, W.A., London, R.L. and Shaw, M. An Introduction to the Construction and Verification of Alphard Programs. IEEE Trans. on Software Eng. SE-2, 4 (December 1976), 253-265.
Присоединяйся к команде
ISSN:
Следуй за Полисом
Оставайся в курсе последних событий
License
Except as otherwise noted, the content of this page is licensed under the Creative Commons Creative Commons «Attribution-NonCommercial-NoDerivatives» 4.0 License, and code samples are licensed under the Apache 2.0 License. See Terms of Use for details.