github.com/PDP-10.its
1
0
Fork 0
mirror of https://github.com/PDP-10/its.git synced 2026-04-26 12:17:41 +00:00
Code Issues Releases Wiki Activity

CSG memo 144: Abstraction Mechanisms in CLU.

Browse Source
This commit is contained in:
Lars Brinkhoff
2021-08-26 13:22:30 +02:00
parent 76e5b7cb8b
commit 74c3facf09
12 changed files with 2294 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

11
build/timestamps.txt
View File

@@ -156,6 +156,17 @@ clib/-read-.-this- 198002261810.43
clib/tv.128 197908312338.58 clib/tv.128 197908312338.58
clu/clu.order 197711161922.32 clu/clu.order 197711161922.32
clu/action.refman 197806022022.04 clu/action.refman 197806022022.04
clu/clukey.r 197909041308.12
clu/clup0.r 197711261712.18
clu/clup1.r 197711261712.27
clu/clup2.r 197711261714.24
clu/clup3.r 197711261712.44
clu/clup4.r 197711261716.29
clu/clup5.r 197711261713.02
clu/clup6.r 197711261713.08
clu/clup7.r 197711261713.18
clu/clupap.header 197711261651.18
clu/clupap.r 197711261651.32
clu/clusym.r 197806271243.01 clu/clusym.r 197806271243.01
clu/exampl.refman 197805301747.35 clu/exampl.refman 197805301747.35
clu/except.refman 197806061946.59 clu/except.refman 197806061946.59

27
doc/clu/clukey.r Normal file
View File

@@ -0,0 +1,27 @@
.
string registers for CLU keywords
.
.de bold
.fr i 0 nargs-1
.sr \\i 1\\i*
.en
.em
.
.bold any array
.bold begin bool break
.bold cand char cluster continue cor cvt
.bold do down
.bold else elseif end except exit
.bold false for force
.bold has
.bold if in int is iter itertype
.bold nil null
.bold oneof others own
.bold proc proctype
.bold real record rep resignal return returns
.bold sequence signal signals string struct
.bold tag tagcase then true type
.bold up
.bold variant
.bold when where while
.bold yield yields

293
doc/clu/clup0.r Normal file
View File

@@ -0,0 +1,293 @@
.nd started 0
.nr do_refs 0
.if ~started
.nr do_refs 1
.en
.so clu/clupap.header
.so r/ref3.rmac
.if csg_memo==0
.ls 1
.if narrow
.new_font 1
.ef
.new_font 3
.en
.nf c
.vp 2i
Abstraction Mechanisms in CLU
.if narrow
.new_font 0
.ef
.new_font 1
.en
.sp .5i
Barbara Liskov
Alan Snyder
Russell Atkinson
Craig Schaffert
.sp .3i
Laboratory for Computer Science
Massachusetts Institute of Technology
545 Technology Square
Cambridge, MA 02139
.sp 2
.if ~narrow
.vp 8.5i
.en
.nf l
.fi
.new_font 0
This research was supported in part by the Advanced Research
Projects Agency of the Department of Defense, monitored by the
Office of Naval Research under contract N00014-75-C-0661, and
in part by the National Science Foundation under grant DCR74-21892.
.ls
.end
.if csg_memo>0
.ls 1
.nf c
.new_font 3
Massachusetts Institute of Technology
Laboratory for Computer Science
.new_font 0
(formerly Project MAC)
.sp 1.25i
Computation Structures Group Memo csg_memo-1
.sp 1.25i
.new_font 4
Abstraction Mechanisms in CLU
.new_font 1
.sp
by
.sp
Barbara Liskov
Alan Snyder
Russell Atkinson
Craig Schaffert
.new_font 0
.nf l
.vp 8.25i
.fi
.new_font 0
This research was supported in part by the Advanced Research
Projects Agency of the Department of Defense, monitored by the
Office of Naval Research under contract N00014-75-C-0661, and
in part by the National Science Foundation under grant DCR74-21892.
.nf c
.sp .5i
January 1977
.nf l
.fi
.ls
.end
.
.
.if narrow
.sp 2
.new_font 1
.ll 7i
.ef
.bp
.rs
.vp 3i
.new_font 3
.en
ABSTRACT
.new_font 0
.sp
.ns
.para
CLU is a new programming language designed to support
the use of abstractions in program construction.
Work in programming methodology has led to the realization
that three kinds of abstractions,
procedural, control, and especially data abstractions,
are useful in the programming process.
Of these, only the procedural abstraction
is supported well by conventional languages,
through the procedure or subroutine.
CLU provides, in addition to procedures,
novel linguistic mechanisms that
support the use of data and control abstractions.
.para
This paper provides an introduction to the abstraction mechanisms
in CLU.
By means of programming examples, we illustrate the utility of
the three kinds of abstractions in program construction
and show how CLU programs may be written to use
and implement abstractions.
We also discuss the CLU library, which permits
incremental program development with complete
type-checking performed at compile-time.
.sp
.fi l
Key words and phrases: programming languages, data types,
data abstractions, control abstractions, programming
methodology, separate compilation.
.sp
CR categories: 4.0, 4.12, 4.20, 4.22.
.br
.fi b
.if narrow
.ll
.en
.
.ref All71
Allen, F. E. and Cocke, J.
A catalogue of optimizing transformations.
Rep. RC 3548,
IBM Thomas J. Watson Research Center,
Yorktown Heights, N.@Y., 1971.
.em
.ref All75
Allen, F. E.
A program data flow analysis procedure.
Rep. RC 5278,
IBM Thomas J. Watson Research Center,
Yorktown Heights, N.@Y., 1975.
.em
.ref Atk76
Atkinson, R. R.
Optimization techniques for a structured programming language.
S.M. Thesis,
Dept. of Electrical Engineering and Computer Science,
M.@I.@T., Cambridge, Mass., June 1976.
.em
.ref Dah70
Dahl, O. J., Myhrhaug, B., and Nygaard, K.
The SIMULA 67 common base language.
Publication S-22, Norwegian Computing Center, Oslo, 1970.
.em
.ref DK75
DeRemer, F. and Kron, H.
Programming-in-the-large versus programming-in-the-small.
2Proceedings of International Conference on Reliable Software*,
2SIGPLAN Notices 10*, 6 (June 1975), 114-121.
.em
.ref Dij72
Dijkstra, E. W.
Notes on structured programming.
2Structured Programming,
A.P.I.C. Studies in Data Processing No. 8*,
Academic Press, New York 1972, 1-81.
.em
.ref Guttag
Guttag, J. V., Horowitz, E., and Musser, D. R.
Abstract data types and software validation.
Rep. ISI/RR-76-48, Information Sciences Institute,
University of Southern California, Marina del Rey,
Calif., August 1976.
.em
.ref Hoare72
Hoare, C. A. R.
Proof of correctness of data representations.
2Acta Informatica*, 4 (1972), 271-281.
.em
.ref Knu73
Knuth, D.
2The Art of Computer Programming*, vol. 3.
Addison Wesley, Reading, Mass., 1973.
.em
.ref LCS75
2Laboratory for Computer Science Progress Report 1974-1975*,
Computation Structures Group.
Rep. PR-XII,
Laboratory for Computer Science, M.@I.@T.,
to be published.
.em
.ref Lam71
Lampson, B. W.
Protection.
Proc. Fifth Annual Princeton Conference on Information
Sciences and Systems, Princeton University, 1971, 437-443.
.em
.ref Lis74
Liskov, B. H. and Zilles, S. N.
Programming with abstract data types.
Proc. ACM SIGPLAN Conference on Very High Level Languages,
2SIGPLAN Notices 9*, 4 (April 1974), 50-59.
.em
.ref Lis75
Liskov, B. H. and Zilles, S. N.
Specification techniques for data abstractions.
2IEEE Trans. on Software Engineering*, 2SE-1*,
(1975), 7-19.
.em
.ref Lis76
Liskov, B. H. and Berzins, V.
An appraisal of program specifications.
Computation Structures Group Memo 141,
Laboratory for Computer Science,
M.@I.@T., Cambridge, Mass., July 1976.
.em
.ref McC62
McCarthy, J., et al.
2LISP 1.5 Programmer's Manual*, MIT Press, 1962.
.em
.ref Mor73
Morris, J. H.
Protection in programming languages.
2Comm. ACM 16*, 1 (Jan 1973), 15-21.
.em
.ref Mor74
Morris, J. H.
Toward more flexible type systems.
Proceedings of the Programming Symposium, Paris, April 9-11, 1974,
2Lecture Notes in Computer Science 19*, Springer-Verlag, New York,
377-384.
..
.ref Par71
Parnas, D. L.
Information distribution aspects of design methodology.
Proc. IFIP 1971.
.em
.ref Sch76
Scheifler, R. W.
An analysis of inline substitution for the CLU programming language.
Computation Structures Group Memo 139,
Laboratory for Computer Science,
M.@I.@T., Cambridge, Mass., June 1976.
.em
.ref Spitzen
Spitzen, J. and Wegbreit, B.
The verification and synthesis of data structures.
2Acta Informatica*, 4 (1975), 127-144.
.em
.ref Standish
Standish, T. A.
2Data structures: an axiomatic approach*.
Rep. 2639, Bolt Beranek and Newman, Cambridge,
Mass., 1973.
.em
.ref Thomas
Thomas, J. W.
Module interconnection in programming systems supporting
abstraction.
Rep. CS-16, Computer Science Program, Brown University,
Providence, R.@I., 1976.
.em
.ref Wir71a
Wirth, N.
Program development by stepwise refinement.
2Comm. ACM 14*, 4 (1971), 221-227.
.em
.ref Wir71b
Wirth, N.
The programming language PASCAL.
2Acta Informatica*, 1 (1971), 35-63.
.em
.ref Wul84
Wulf, W. A., London, R., and Shaw, M.
An introduction to the construction and verification
of Alphard programs.
2IEEE Transactions on Software Engineering SE-2*,
(1976), 253-264.
.em
.bp
.if do_refs
.insert_refs
.en
.if narrow
.rs
.sp 3i
.en

212
doc/clu/clup1.r Normal file
View File

@@ -0,0 +1,212 @@
.so clu/clupap.header
.chapter "Introduction"
.para
The motivation for the design of the CLU programming
language was to provide programmers with a tool that would
enhance their effectiveness in constructing programs of
high quality -- programs that are reliable and reasonably
easy to understand, modify, and maintain.
CLU aids programmers
by providing constructs that support
the use of abstractions in program design and implementation.
.para
The quality of software depends primarily on
the programming methodology in use.
The choice of programming language, however, can have a major impact on
the effectiveness of a methodology.
A methodology can be easy
or difficult to apply in a given language, depending on
how well the language constructs match the
structures that the methodology deems desirable.
The presence of constructs that give a concrete form
for the desired structures makes the methodology more understandable.
In addition, a programming language influences the way that
its users think about programming;
matching a language to a methodology increases the likelihood that
the methodology will be used.
.para
CLU has been designed to support a methodology
(similar to
[Dij72,@Wir71a])
in which programs are developed by
means of problem decomposition based on the recognition
of abstractions.
A program is constructed in many
stages.
At each stage, the problem to be solved is
how to implement some abstraction (the initial problem
is to implement the abstract behavior required of the
entire program).
The implementation is developed by envisioning a number
of subsidiary abstractions (abstract objects and
operations) that are useful in the problem domain.
Once the behavior of the abstract objects and operations
has been defined, a program can be written to solve the
original problem; in this program, the abstract objects
and operations are used as primitives.
Now the original
problem has been solved, but new problems have arisen,
namely, how to implement the subsidiary abstractions.
Each of these abstractions is
considered in turn as a new problem; its implementation
may introduce further abstractions.
This process
terminates when all the abstractions introduced at various
stages have been implemented or are present in the
programming language in use.
.para
In this methodology, programs are developed
incrementally, one abstraction at a time.
Further, a distinction is made between an abstraction,
which is a kind of behavior, and a program,
or 2module*, which implements that behavior.
An abstraction isolates
use from implementation: an abstraction can be used
without knowledge of its implementation and implemented
without knowledge of its use.
These aspects of the methodology are supported by the
CLU 2library*, which maintains
information about abstractions
and the CLU modules that implement them.
The library permits separate compilation of
modules with complete type-checking at
compile-time.
.para
To make effective use of the
methodology, it is necessary to understand the kinds
of abstractions that are useful in constructing programs.
In studying this question,
we identified an important kind of abstraction,
the data abstraction, that
had been largely neglected in discussions of programming methodology.
.para
A data abstraction [Hoare72,@Lis74,@Standish]
is used to introduce a new
type of data object that is deemed useful
in the domain of the problem being solved.
At the level of use, the programmer is
concerned with the 2behavior* of these data objects,
what kinds of information can be stored in them and
obtained from them.
The programmer is 2not* concerned
with how the data objects are represented in storage,
nor with the algorithms used to store and access
information in them.
In fact, a data abstraction is
often introduced to delay such implementation
decisions until a later stage of design.
.para
The behavior of the data objects is expressed most
naturally in terms of a set of operations that are meaningful
for those objects.
This set will include operations
to create objects, to obtain information from them,
and possibly to modify them.
For example,
push and pop are among the meaningful operations for stacks,
while meaningful operations for integers include the usual
arithmetic operations.
Thus, a data abstraction consists of a
set of objects and a set of operations
characterizing the behavior of the
objects.
.para
If a data abstraction is to be
understandable at an abstract level,
the behavior of the data objects must be
2completely* characterized by the set of operations.
This property is ensured by making the operations the
2only direct means* of creating and manipulating the objects.
One effect of this restriction
is that, when defining an abstraction,
the programmer must be careful to include a
sufficient set of operations, since every action
he wishes to perform on the objects must be
realized in terms of this set.
.para
We have identified the following requirements that must be
satisfied by a language supporting data abstractions:
.ilist 3
1. A linguistic construct is needed that permits
a data abstraction to be implemented as a unit.
The implementation involves selecting a representation
for the data objects and defining an algorithm for each
operation in terms of that representation.
.next
2. The language must limit access to the
representation to just the operations. This limitation
is necessary to ensure that the operations completely
characterize the behavior of the objects.
.end_list
CLU satisfies these requirements by providing a linguistic construct
called a 2cluster* for implementing data abstractions.
Data abstractions are integrated into the language
through the data type mechanism.
Access to the representation is
controlled by type-checking, which is done at
compile time.
.para
In addition to data abstractions, CLU
supports two other kinds of abstractions:
procedural abstractions and control abstractions.
A procedural abstraction performs a computation on a
set of input objects and produces a set of output objects;
examples of procedural abstractions are sorting an
array and computing a square root.
CLU supports procedural abstractions by means of procedures,
which are similar to procedures in other programming languages.
.para
A control abstraction defines a method
for sequencing arbitrary actions.
All languages provide built-in control abstractions;
examples are the if statement and the while statement.
In addition, however,
CLU allows user definitions of a simple kind of control abstraction.
The method provided is a generalization of the
repetition methods available in many programming
languages.
Frequently the programmer desires to
perform the same action for all the objects in a
collection, such as all
characters in a string or all items in a set.
CLU
provides a linguistic construct called an 2iterator*
for defining how the objects in the
collection are obtained.
The iterator is used in
conjunction with the for statement; the body
of the for statement describes the action to be
taken.
.para
The purpose of this paper is to illustrate
the utility of the three kinds of abstractions
in program construction,
and to provide an informal introduction to CLU.
We do not attempt a complete description of the language;
rather, we concentrate on the constructs that
support abstractions.
The presence of these
constructs constitutes the most important way in
which CLU differs from other languages.
The language closest to CLU is Alphard [Wul84],
which represents a concurrent design effort with goals similar to
our own.
The design of CLU has been influenced by
SIMULA 67 [Dah70], and to a lesser extent by
Pascal [Wir71b] and LISP [McC62].
.para
In the next section we introduce CLU and,
by means of a programming example,
illustrate the use and implementation
of data abstractions.
Section semantics describes the basic semantics of CLU.
In Section more_abstraction, we discuss
control abstractions and more powerful kinds of
data abstractions.
We present the CLU library in Section library.
Section implementation briefly describes
the current implementation of CLU
and discusses efficiency considerations.
.ne 2
Finally, we conclude by discussing
the quality of CLU programs.

595
doc/clu/clup2.r Normal file
View File

@@ -0,0 +1,595 @@
.nd chapter 2-1
.so clu/clupap.header
.
string registers for italic variable names
.
.sr i 2i*
.sr s 2s*
.sr o 2o*
.sr c 2c*
.sr n 2n*
.sr t 2t*
.sr r 2r*
.sr x 2x*
.sr tr 2tr*
.sr w 2w*
.sr wb 2wb*
.sr total 2total*
.sr contents 2contents*
.sr count_words 2count_words*
.sr next_word 2next_word*
.sr wordbag 2wordbag*
.sr wordtree 2wordtree*
.sr wordbags 2wordbags*
.sr wordtrees 2wordtrees*
.sr insert 2insert*
.sr create 2create*
.sr print 2print*
.sr instream 2instream*
.sr instreams 2instreams*
.sr outstream 2outstream*
.sr outstreams 2outstreams*
.
.chapter "An Example of Data Abstraction"
.para
This section introduces the basic data
abstraction mechanism of CLU, the cluster.
By means of an example, we intend to show how
abstractions occur naturally in program design,
and how they are used and implemented in CLU.
In particular, we show how a data abstraction
can be used as structured intermediate storage.
.para
Consider the following problem:
Given some document, we wish to compute,
for each distinct word in the document,
the number of times the word occurs
and its frequency of occurrence as a percentage of the total
number of words.
The document will be
represented as a sequence of characters.
A word is any non-empty sequence of
alphabetic characters.
Adjacent words are
separated by one or more non-alphabetic
characters such as spaces, punctuation, or newline
characters.
In recognizing distinct words, the
difference between upper and lower case letters should
be ignored.
.para
The output is also to be a sequence of characters,
divided into lines.
Successive lines should contain an alphabetical
list of all the distinct words in the document,
one word per line.
Accompanying each word should
be the total number of occurrences and the
.ne 5
frequency of occurrence. For example:
.table
.ta 8 20 28
a 2 3.509%
access 1 1.754%
and 2 3.509%
dots
.rtabs
.end_table
.para
Specifically, we are required to write the
procedure count_words, which takes two arguments:
an instream and an outstream.
The former is the
source of the document to be processed, and the latter
is the destination of the required output.
.ne 5
The form of this procedure will be
.code
count_words = proc (i: instream, o: outstream);
dots
end count_words;
.end_code
Note that count_words does not return any results;
its only effects are modifications of i (reading the entire
document) and of o (printing the required statistics).
.para
2Instream* and outstream are data abstractions.
An instream i contains a sequence of characters.
Of the primitive
operations on instreams, only two will be of interest to us.
2Empty@(i)* returns true if there are no characters available
in i, and returns false otherwise.
2Next@(i)* removes the first character from the sequence
and returns it.
Invoking the next operation on an empty instream is an
error.
.foot
The CLU error handling mechanism is discussed in [LCS75].
.efoot
An outstream also contains a sequence of characters.
The interesting operation on outstreams is
2put_string@(s,@o)*,
which appends the string s to the existing sequence of characters
in o.
.para
Now consider how we might implement count_words.
We begin by deciding how to handle words.
We could define a new abstract data type 2word*.
However, we choose instead to use strings (a primitive
CLU type), with the restriction that only strings of
lower-case alphabetic characters will be used.
.foot
Sometimes it is difficult to decide whether to introduce
a new data abstraction or to use an existing abstraction.
Our decision to use strings to represent words was made
partly to shorten the presentation.
.efoot
.para
Next, we investigate how to scan the document.
Reading a word requires knowledge of the
exact way in which words occur in the input stream.
We choose to isolate this information in a procedural abstraction,
called next_word,
which takes in the instream i and returns the next word
(converted to lower case characters) in the document.
If there are no more words,
next_word must communicate this fact to count_words.
A simple way to indicate that there are no
more words is by returning an ``end of document'' word,
one that is distinct from any other word.
A reasonable choice for the ``end of document'' word is
the empty string.
.para
It is clear that in count_words we must scan the
entire document before we can print our results, and
therefore, we need some receptacle
to retain information about words between these two
actions (scanning and printing).
Recording the
information gained in the scan and organizing it
for easy printing will probably be fairly complex.
Therefore, we will defer such considerations until later
by introducing a data abstraction wordbag with the
appropriate properties.
In particular, wordbag provides
three operations: create, which creates an empty wordbag;
insert, which adds a word to the wordbag; and print, which
prints the desired statistical information about the words
in the wordbag.
.foot
The print operation is not the ideal choice, but a better
solution requires the use of control abstractions.
This solution is presented in Section more_abstraction.
.efoot
.nr count_words current_figure
.para
The implementation of count_words is shown in
Figure count_words.
.begin_figure "The count_words procedure."
.code
count_words = proc (i: instream, o: outstream);
% create an empty wordbag
wb: wordbag := wordbag$create ();
% scan document, adding each word found to wb
w: string := next_word (i);
while w ~= "" do
wordbag$insert (wb, w);
w := next_word (i);
end;
% print the wordbag
wordbag$print (wb, o);
end count_words;
.ns
.end_code
.finish_figure
The ``%'' character starts a comment,
which continues to the end of the line.
The ``~'' character stands for boolean negation.
The notation 2variable:@type* is used
in formal argument lists and declarations
to specify the types of variables;
a declaration may be combined with an assignment
specifying the initial value of the variable.
Boldface is used for reserved words, including the
names of primitive CLU types.
CLU does not permit
redefinition of the primitive types; however,
primitive types are used in the same way as abstract
types.
.para
The count_words procedure declares four variables:
i, o, wb, and w.
The first two denote the instream and
outstream that are passed as arguments to count_words.
The third, wb, denotes the wordbag used to hold
the words read so far,
and the fourth, w, the word
currently being processed.
.para
Operations of a data abstraction are named by
a compound form that specifies both the type and
the operation name. Three examples of operation calls
appear in count_words: 2wordbag$create@()*,
2wordbag$insert@(wb,@w)*
and 2wordbag$print@(wb,@o)*.
The CLU system provides a mechanism that avoids conflicts
between names of abstractions; this mechanism is discussed in
Section library.
However, operations of two different data abstractions may have
the same name;
the compound form serves to resolve this ambiguity.
Although the ambiguity could in most cases be resolved by context,
we have found in using CLU that the compound
form enhances the readability of programs.
.nr next_word current_figure
.para
The implementation of next_word is shown in
Figure next_word.
.begin_figure "The next_word procedure."
.code
next_word = proc (i: instream) returns (string);
c: char := 1' '*;
% scan for first alphabetic character
while ~alpha (c) do
if instream$empty (i)
then return "";
end;
c := instream$next (i);
end;
% accumulate characters in word
w: string := "";
while alpha (c) do
w := string$append (w, c);
if instream$empty (i)
then return w;
end;
c := instream$next (i);
end;
return w; % the non-alphabetic character c is lost
end next_word;
.ns
.end_code
.finish_figure
The 2string$append* operation creates a new string
by appending a character to the characters in the
string argument
(it does 2not* modify the string argument).
Note the use of the instream operations
2next* and 2empty*.
Note also that two additional procedures have been used:
2alpha@(c)*,
which tests whether a character is alphabetic or not,
and 2lower_case@(c)*,
which returns the lower case version of a character.
The implementations of these procedures are not shown in the paper.
.para
Now we must implement the type wordbag.
.ne 5
The cluster will have the form
.code
wordbag = cluster is create, insert, print;
dots
end wordbag;
.end_code
This form expresses the idea that the data abstraction is a set
of operations as well as a set of objects.
The cluster must
provide a representation for objects of the type wordbag and
an implementation for each of the operations.
We are free to choose from the possible representations the
one best suited to our use of the wordbag cluster.
.para
The representation that we choose should allow
reasonably efficient storage of words and easy printing,
in alphabetic order, of the words and associated statistics.
For efficiency in computing the statistics, maintaining
a count of the total number of words in the document
would be helpful.
Since the total number of words in the document is probably
much larger than the number of distinct words, the
representation of a wordbag should contain only one ``item'' for
each distinct word (along with a multiplicity count), rather
than one ``item'' for each occurrence.
This choice of representation requires that, at
each insertion, we check whether the new word is already
present in the wordbag.
We would like a representation that
allows the search for a matching ``item'' and the insertion of a
not-previously-present ``item'' to be efficient.
A binary tree representation [Knu73] fits our requirements nicely.
.para
Thus the main part of the wordbag representation will
consist of a binary tree.
The binary tree is another data abstraction,
wordtree. The data abstraction wordtree
provides operations very similar to those of wordbag:
2create@()* returns an empty wordtree;
2insert@(tr,@w)* returns a wordtree containing all the
words in the wordtree tr plus the additional word w
(the wordtree tr may be modified in the process);
and 2print@(tr,@n,@o)* prints the contents of the
wordtree tr in alphabetic order on outstream o, along with the
number of occurrences and the frequency (based on a total of
n words).
.nr wordbag current_figure
.para
The implementation of wordbag is given in Figure wordbag.
.begin_figure "The wordbag cluster."
.code
wordbag = cluster is
create, % create an empty bag
insert, % insert an element
print; % print contents of bag
rep = record [contents: wordtree, total: int];
create = proc () returns (cvt);
return rep${contents: wordtree$create (), total: 0};
end create;
insert = proc (x: cvt, v: string);
x.contents := wordtree$insert (x.contents, v);
x.total := x.total + 1;
end insert;
print = proc (x: cvt, o: outstream);
wordtree$print (x.contents, x.total, o);
end print;
end wordbag;
.ns
.end_code
.finish_figure
Following the header, we find the definition of the
.ne 3
representation selected for wordbag objects:
.code
rep = record [contents: wordtree, total: int];
.end_code
The reserved type identifier rep indicates that the type
specification to the right of the equal sign is the representing
type for the cluster.
We have defined the representation of a wordbag object to
consist of two pieces: a wordtree,
as explained above, and an integer, which records the total
number of words in the wordbag.
.para
A CLU record is an object with one or more named
components.
For each component name, there is an operation to select
and an operation to set the corresponding component.
The operation 2get_n@(r)* returns the n component
of the record r (this operation is usually
abbreviated 2r.n*).
The operation 2put_n@(r,@x)* makes x the n component
of the record r (this operation is usually
abbreviated 2r.n@*:=2@x*,
by analogy with the assignment statement).
A new record is created by an expression of the form
type${name1: value1, dots}.
.para
There are two different
types associated with any cluster: the abstract
type being defined (wordbag in this case) and the
representation type (the record).
Outside of the cluster,
type-checking will ensure that a wordbag object will always be
treated as such.
In particular, the ability to convert a wordbag object into its
representation is not provided (unless one of the
wordbag operations does so explicitly).
.para
Inside the cluster, however, it is necessary to view
a wordbag object as being of the representation type,
because the implementations of the
operations are defined in terms of the representation.
This change of viewpoint is signalled by having the
reserved word cvt appear as the type of an
argument (as in the insert and print operations).
1Cvt* may also appear as a return type
(as in the create operation);
here it indicates that a returned object
will be changed into an object of abstract type.
Whether cvt appears as the type of an
argument or as a return type,
it stipulates a ``conversion'' of viewpoint
between the external abstract type and the internal representation type.
1Cvt* can be used only within a cluster,
and conversion can be done only between the single abstract
type being defined and the (single) representation type.
.foot
1Cvt* corresponds to Morris' seal and unseal [Mor73],
except that 1cvt* represents a change in viewpoint only;
no computation is required.
.efoot
.para
The procedures in wordbag are very simple.
2Create* builds a new instance of the rep by use of the
.ne 3
record constructor
.code
rep${contents: wordtree$create (), total: 0}
.end_code
Here total is initialized to 0, and contents to the
empty wordtree (by calling the create operation of wordtree).
This rep object is converted to a wordbag object as it
is being returned.
2Insert* and print are implemented directly
in terms of wordtree operations.
.nr wordtree current_figure
.para
The implementation of wordtree is shown in Figure wordtree.
In the wordtree representation, each node
contains a word and the number of times that word has been
inserted into the wordbag, as well as two subtrees.
.begin_page_figure "The wordtree cluster."
.code
wordtree = cluster is
create, % create empty contents
insert, % add item to contents
print; % print contents
node = record [m!value: string, count: int,
(mark!m)lesser: wordtree, greater: wordtree];
rep = oneof [empty: null, non_empty: node];
create = proc () returns (cvt);
return rep$make_empty (nil);
end create;
insert = proc (x: cvt, v: string) returns (cvt);
tagcase x
tag empty:
n: node := node${m!value: v, count: 1,
(mark!m)lesser: wordtree$create (),
(mark!m)greater: wordtree$create ()};
return rep$make_non_empty (n);
tag non_empty (n: node):
if v = n.value
then n.count := n.count + 1;
elseif v < n.value
then n.lesser := wordtree$insert (n.lesser, v);
else n.greater := wordtree$insert (n.greater, v);
end;
return x;
end;
end insert;
print = proc (x: cvt, total: int, o: outstream);
tagcase x
tag empty: ;
tag non_empty (n: node):
wordtree$print (n.lesser, total, o);
print_word (n.value, n.count, total, o);
wordtree$print (n.greater, total, o);
end;
end print;
end wordtree;
.ns
.end_code
.finish_figure
For any
particular node, the words in the ``lesser'' subtree must
alphabetically precede the word in the node, and the words
in the ``greater'' subtree must follow the word in the node.
.ne 4
This information is described by
.code
node = record [m!value: string, count: int,
(mark!m)lesser: wordtree, greater: wordtree];
.end_code
which defines ``node'' to be an
abbreviation for the information following
the equal sign.
(The reserved word rep is used similarly,
as an abbreviation for the representation type.)
.para
Now consider the representation of wordtrees.
A non-empty wordtree can be represented by its top node.
An empty wordtree, however, contains no information.
The ideal type to represent an empty wordtree
is the CLU type null,
which has a single data object nil.
So the representation of a wordtree should
be either a node or nil.
.ne 3
This representation is expressed by
.code
rep = oneof [empty: null, non_empty: node];
.end_code
.para
Just as the record is the basic CLU
mechanism to form an object
that is a collection of other objects,
the oneof is the basic CLU mechanism to form an object
that is ``one of'' a set of alternatives.
Oneof is CLU's method of forming a
discriminated union, and is somewhat similar to
a variant component of a record in Pascal [Wir71b].
.para
An object of the type oneof@[s1:@T1 dots sn:@Tn]
can be thought of as a pair.
The ``tag'' component is an
identifier from the set {s1 dots sn}.
The ``value''
component is an object of the type corresponding to the
tag.
That is, if the tag component is si then the
value is some object of type Ti.
.para
Objects of type oneof@[s1:@T1 dots sn:@Tn]
are created by the operations 2make_si@(x)*, each of
which takes an object x of type Ti
and returns the pair <si,@x>.
Because the type of the value component of a oneof object is not
known at compile-time, allowing direct access
to the value component
could result in a run-time type error (e.g., assigning an object
to a variable of the wrong type). To eliminate this possibility,
.ne 7
we require the use of a special tagcase statement to decompose
a oneof object:
.code
tagcase e
tag s1 (id1: T1): @@@m!statements dots
dots
tag sn (idn: Tn):(mark!m)statements dots
end;
.end_code
This statement evaluates the expression 2e*
to obtain an object of type
oneof@[s1:@T1@dots@sn:@Tn].
If the tag is si,
then the value is assigned to the new variable
idi and the statements following the ith alternative
are executed.
The variable idi is local to those statements.
If, for some reason, we do not need the value,
we can omit the parenthesized variable declaration.
.para
The reader should now know enough to understand
Figure wordtree.
Note, in the create operation, the use
of the construction operation 2make_empty*
of the representation type of wordtree
(the discriminated union oneof@[empty:@null,@non-empty:@node])
to create the empty wordtree.
The tagcase statement is used in both insert and print.
Note that if insert is given an empty wordtree, it creates a
new top node for the returned value,
but if insert is given a non-empty wordtree,
it modifies the given wordtree and returns it.
.foot
It is necessary for insert to return a value in addition to
having a side-effect because, in the case of an empty wordtree
argument, side-effects are not possible. Side-effects are not
possible because of the representation chosen for the empty
wordtree and because of the CLU parameter passing mechanism
(see Section semantics).
.efoot
The insert operation depends on the dynamic
allocation of space for newly-created records (see
Section semantics).
.para
The print operation uses the obvious recursive descent.
It makes use of procedure
2print_word@(w,@c,@t,@o)*, which generates a single line of
output on 2o*, consisting of the word 2w*,
the count 2c*, and the frequency of occurrence
derived from 2c* and 2t*.
The implementation of 2print_word* has been omitted.
.para
We have now completed
our first discussion of the count_words procedure.
We return to this problem in Section more_abstraction,
where we present a superior solution.

181
doc/clu/clup3.r Normal file
View File

@@ -0,0 +1,181 @@
.nd chapter 3-1
.nd current_figure 5
.nd wordbag 3
.so clu/clupap.header
.sr m 2m*
.sr p 2p*
.sr q 2q*
.sr x 2x*
.sr y 2y*
.sr z 2z*
.sr a 2a*
.sr b 2b*
.sr insert 2insert*
.sr increment 2increment*
.chapter "Semantics"
.para
All languages present their users with some model of computation.
This section describes those aspects of CLU semantics that differ
from the common ALGOL-like model.
In particular, we discuss
CLU's notions of objects and variables,
and the definitions of assignment and argument passing that
follow from these notions.
We also discuss type correctness.
.section "Objects and Variables"
.para
The basic elements of CLU semantics are
2objects* and 2variables*.
Objects are the data entities that are created and manipulated
by CLU programs.
Variables are just the names used in a
program to refer to objects.
.para
In CLU, each object has a particular 2type*,
which characterizes its behavior.
A type defines a set of operations
that create and manipulate objects of that type.
An object
may be created and manipulated only via the operations of its type.
.para
An object may 2refer* to objects.
For example,
a record object refers to the objects that are the components
of the record.
This notion is one of logical, not physical, containment.
In particular, it is possible for two distinct record objects to
refer to (or 2share*) the same component object.
In the case of a cyclic structure, it is even possible for an object
to ``contain'' itself.
Thus, it is possible to have recursive data
structure definitions and shared data objects without explicit
reference types.
The 2wordtree* type described in the previous
section is an example of a recursively-defined data structure.
(This notion of object is similar to that in LISP.)
.para
CLU objects exist independently of procedure activations.
Space for objects is allocated from a dynamic storage area
as the result of invoking
constructor operations of certain primitive CLU types.
For example,
the record constructor is used in the implementation of 2wordbag*
(Figure wordbag) to acquire space for new 2wordbag* objects.
In theory, all objects continue to exist forever.
In practice,
the space used by an object may be reclaimed when that object is
no longer accessible to any CLU program.
.foot
An object is accessible if it is denoted by a variable of an active
procedure or is a component of an accessible object.
.efoot
.para
An object may exhibit time-varying behavior.
Such an object, called a 2mutable* object,
has a state which may be modified by certain operations
without changing the identity of the object.
Records are examples of mutable objects.
The record update operations (2put_s (r,@v)*,
written as 2r*.2s*@:=@2v*
in the examples) change the state of record objects and
therefore affect the behavior of subsequent applications of
the select operations (2get_s (r)*, written as 2r*.2s*).
The 2wordbag* and 2wordtree* types are additional examples
of types with mutable objects.
.para
If a mutable object m is shared by two other objects x and y,
then a modification to m made via x will be visible when m is
examined via y.
Communication through
shared mutable objects is most beneficial in the context
of procedure invocation, described below.
.para
Objects that do not exhibit time-varying behavior are called
2immutable* objects, or 2constants*.
Examples of constants are integers, booleans,
characters, and strings.
The value of a constant object can not be modified.
For example,
new strings may be computed from old ones,
but existing strings do not change.
Similarly,
none of the integer operations
modify the integers passed to them as arguments.
.para
Variables are names used in CLU programs to 2denote*
particular objects at execution time.
Unlike variables in many common programming languages,
which 2are* objects that 2contain* values,
CLU variables are simply names
that the programmer uses to refer to objects.
As such, it is possible for two variables to denote
(or 2share*) the same object.
CLU variables are much like those in LISP,
and are similar to pointer variables in other languages.
However, CLU variables are 2not* objects;
they cannot be denoted by other variables or referred to by objects.
Thus, variables are completely private to the procedure
in which they are declared,
and cannot be accessed or modified by any other procedure.
.section "Assignment and Procedure Invocation"
.para
The basic actions in CLU are 2assignment* and
2procedure invocation*.
The assignment primitive 2x*@:=@2E*, where x is a variable
and 2E* is an expression, causes x to denote
the object resulting from the evaluation of 2E*.
For example,
if 2E* is a simple variable y, then the assignment x@:=@y
causes x to denote the object denoted by y.
The object
is 2not* copied; after the assignment is performed, it will be
2shared* by x and y.
Assignment does not affect
the state of any object.
(Recall that 2r*.2s*@:=@2v* is not a true assignment,
but an abbreviation for 2put_s@(r,@v)*.)
.para
Procedure invocation involves passing argument objects
from the caller to the called procedure and returning result
objects from the procedure to the caller.
The formal arguments
of a procedure are considered to be local variables of the procedure,
and are initialized, by assignment, to the objects resulting from the
evaluation of the argument expressions. Thus, argument
objects are shared between the caller and the called procedure.
A procedure may modify mutable argument objects (e.g., records),
but of course it cannot modify immutable ones (e.g., integers).
A procedure has no access to the variables of its caller.
.para
Procedure invocations may be
used directly as statements; those
that return objects may also be used as expressions.
Arbitrary recursive procedures are permitted.
.ne 5
.section "Type Correctness"
.para
Every variable in a CLU module must be declared;
the declaration specifies the type of object
that the variable may denote.
All assignments to a variable must satisfy
the variable's declaration.
Because argument passing is defined
in terms of assignment, the types of actual
argument objects must be consistent with the declarations of the
corresponding formal arguments.
.para
These restrictions, plus the restriction that only the code
in a cluster may use cvt to convert between the abstract
and representation types, ensure that the behavior of an object
is indeed characterized completely by the operations of its type.
For example, the type restrictions ensure that
the only modification possible to a record object that represents
a 2wordbag* (Figure wordbag) is the modification performed by
the insert operation.
.para
Type-checking is performed on a module-by-module basis
at compile-time (it could also be done at run-time).
This checking can catch all type errors -- even those involving
inter-module references -- because the CLU library maintains the
necessary type information for all modules
(see Section 5.)

494
doc/clu/clup4.r Normal file
View File

@@ -0,0 +1,494 @@
.nd chapter 4-1
.nd current_figure 7
.so clu/clupap.header
.
.sr words 2words*
.sr wordbag 2wordbag*
.sr sorted_bag 2sorted_bag*
.sr sorted_bags 2sorted_bags*
.sr wordtree 2wordtree*
.sr tree 2tree*
.sr node 2node*
.sr r 2r*
.sr x 2x*
.sr t 2t*
.sr count_words 2count_words*
.sr count_numeric 2count_numeric*
.sr lt 2lt*
.sr equal 2equal*
.sr print 2print*
.sr string_chars 2string_chars*
.sr create 2create*
.sr insert 2insert*
.sr size 2size*
.sr increasing 2increasing*
.sr s 2s*
.sr n 2n*
.sr index 2index*
.sr limit 2limit*
.sr count 2count*
.sr next_word 2next_word*
.sr elements 2elements*
.sr reverse_elements 2reverse_elements*
.
.
.chapter "More Abstraction Mechanisms"
.para
In this section we continue our discussion of
abstraction mechanisms in CLU.
A generalization of the 2wordbag* abstraction,
called 2sorted_bag*,
is presented as an illustration of parameterized clusters,
which are a means for implementing
more generally applicable data abstractions.
The presentation of 2sorted_bag*
is also used to motivate the introduction of a control
abstraction called an 2iterator*,
which is a mechanism for incrementally generating
the elements of a collection of objects.
Finally, we show an implementation of the sorted_bag
abstraction and illustrate how sorted_bag
can be used in implementing count_words.
.section "Properties of the Sorted_bag Abstraction"
.para
In the count_words procedure given earlier,
a data abstraction called wordbag was used.
A wordbag object is a collection of strings,
each with an associated count.
Strings are inserted into a wordbag object one at a time.
Strings in a wordbag object may be printed in alphabetical order,
each with a count of the number of times it was inserted.
.para
Although wordbag has properties that are specific to the usage
in count_words,
it also has properties in common with a more general abstraction,
sorted_bag.
A bag is similar to a set
(it is sometimes called a multi-set)
except that an item can appear in a bag many times.
For example, if the integer 1 is inserted in the set {1,2},
the result is the set {1,2},
but if 1 is inserted in the bag {1,2},
the result is the bag {1,1,2}.
A sorted_bag is a bag that affords access
to the items it contains
according to an ordering relation on the items.
.para
The concept of a sorted_bag is meaningful not only for strings
but for many types of items.
Therefore, we would like to parameterize the sorted_bag abstraction,
the parameter being the type of item to be collected
in the sorted_bag objects.
.para
Most programming languages provide built-in parameterized
data abstractions.
For example, the concept of an array is a parameterized
data abstraction.
.ne 3
An example of a use of arrays in Pascal is
.code
1array* 1..n 1of* 1integer*
.end_code
These arrays have two parameters,
one specifying the array bounds (1..n)
and one specifying the type of element in the array (integer).
In CLU we provide mechanisms allowing user-defined
data abstractions (like sorted_bag) to be parameterized.
.para
In the sorted_bag abstraction,
not all types of items make sense.
Only types that define a total ordering on their objects
are meaningful,
since the sorted_bag abstraction depends on the presence
of this ordering.
In addition, information about the ordering must be
expressed in a way that is useful for programming.
A natural way to express this information
is by means of operations of the item type.
Therefore, we require that the item type provide
less than and equal operations
(called lt and equal).
.ne 5
This constraint is expressed in the header for sorted_bag:
.code
sorted_bag = cluster [t: type] is create, insert, dots
where t has
lt, equal: proctype (t, t) returns (bool);
.end_code
The item type t is a 2formal parameter* of the sorted_bag
cluster; whenever the sorted_bag abstraction is used,
.ne 3
the item type must be specified as an 2actual parameter*, e.g.,
.code
sorted_bag[string]
.end_code
.para
The information about required operations
informs the programmer about legitimate uses of sorted_bag.
The compiler will check each use of sorted_bag to ensure
that the item type provides the required operations.
The where clause specifies exactly the information
that the compiler can check.
Of course, more is assumed about the item type 2t*
than the presence of
operations with appropriate names and functionalities:
these operations must also define a total ordering on the items.
Although we expect formal and complete specifications
for data abstractions to be included in the CLU library eventually,
we do not include in the CLU language declarations
that the compiler cannot check.
This point is discussed further in Section discussion.
.para
Now that we have decided to define a
sorted_bag abstraction that works for many item types,
we must decide what operations this abstraction provides.
When an abstraction (like wordbag)
is written for a very specific purpose,
it is reasonable to have
some specialized operations.
For a more general abstraction,
the operations should be more generally useful.
.para
The 2print* operation is a case in point.
Printing is only one possible use of the information contained
in a 2sorted_bag*.
It was the only use in the case of 2wordbag*,
so it was reasonable to have a 2print* operation.
However, if 2sorted_bags* are to be generally useful,
there should be some way for the user to obtain
the elements of the 2sorted_bag*; the user can then
perform some action on the elements (for example, print them).
.para
What we would like is an operation on sorted_bags
that makes all of the elements available to the caller
in increasing order.
One possible approach is to map
the elements of a sorted_bag
into a sequence object,
a solution potentially requiring a large amount of space.
A more efficient method is provided by CLU and is discussed below.
This solution computes the sequence
one element at a time, thus saving space.
If only part of the sequence is used
(as in a search for some element),
then execution time can be saved as well.
.section "Control Abstractions"
.para
The purpose of many loops is to perform an action
on some or all of the objects in a collection.
For such loops,
it is often useful to separate the
selection of the next object
from the action performed on that object.
CLU provides a control abstraction that permits
a complete decomposition of the two activities.
The for statement available in many programming languages
provides a limited ability in this direction:
it iterates over ranges of integers.
The CLU for statement
can iterate over collections of any
type of object.
The selection of the next object in the collection
is done by a user-defined 2iterator*.
The iterator
produces the objects in the collection one at a time
(the entire collection need not physically exist);
each object is consumed by the for statement in turn.
.nr rra0 current_figure
.para
Figure rra0 gives an example of a simple iterator
called string_chars, which produces the characters in a string in
the order in which they appear.
.begin_figure "Use and definition of a simple iterator."
.code
count_numeric = proc (s: string) returns (int);
count: int := 0;
for c: char in string_chars (s) do
if char_is_numeric (c)
then count := count + 1;
end;
end;
return count;
end count_numeric;
string_chars = iter (s: string) yields (char);
index: int := 1;
limit: int := string$size (s);
while index <= limit do
yield string$fetch (s, index);
index := index + 1;
end;
end string_chars;
.ns
.end_code
.finish_figure
This iterator uses string operations 2size@(s)*,
which tells how many characters are in the string s,
and 2fetch@(s,@n)*,
which returns the n!th character in the string s
(provided the integer n is greater than zero
and does not exceed the size of the string).
.foot
A while loop is used in the implementation of
string_chars so that the example would be based
on familiar concepts. In actual practice, such a
loop would be written using a for statement invoking
a primitive iterator.
.efoot
.br
.ne 5
.para
The general form of the CLU for statement is
.code
for declarations in iterator-invocation do
body
end;
.end_code
An example of the use of the for statement
occurs in the count_numeric procedure
(see Figure rra0),
which contains a loop
that counts the number of numeric characters in a string.
Note that the details of how the characters are obtained
from the string are entirely contained
in the definition of the iterator.
.para
Iterators work as follows:
A for statement initially invokes an iterator,
passing it some arguments.
Each time a yield statement is executed in the iterator,
the objects yielded
.foot
Zero or more objects may be yielded,
but the number and types of objects yielded each time by an iterator
must agree with the number and types of variables in
a for statement using the iterator.
.efoot
are assigned to the variables declared in the for statement
(following the reserved word for)
in corresponding order, and the body of the for
statement is executed.
Then the iterator is resumed at the statement
following the yield statement,
in the same environment as when the objects were yielded.
When the iterator terminates, by either an implicit
or explicit return, then the invoking for statement
terminates. The iteration may also be prematurely
terminated by a return in the body of the
for statement.
.para
For example, suppose that string_chars is invoked
with the string ``a3''.
The first character yielded is `a'.
At this point within string_chars, index@=@1 and limit@=@2.
Next the body of the for statement is performed.
Since the character `a' is not numeric,
count remains at 0.
Next string_chars is resumed at the statement after the yield
statement, and when resumed, index@=@1 and limit@=@2.
Then index is assigned 2,
and the character `3' is selected from the string and yielded.
Since `3' is numeric, count becomes@1.
Then string_chars is resumed,
with index@=@2 and limit@=@2, and index is incremented,
which causes the while loop to terminate.
The implicit return terminates both the iterator and the
for statement, with control resuming at the statement
after the for statement,
and count@=@1.
.para
While iterators are useful in general,
they are especially valuable in conjunction with data abstractions
that are collections of objects (such as sets, arrays, and
sorted_bags).
Iterators afford users of such abstractions access to all objects
in the collection, without exposing irrelevant details.
Several iterators may be included in a data abstraction.
When the order of obtaining the objects is important,
different iterators may provide different orders.
.section "Implementation and Use of Sorted_bag"
.para
Now we can describe a minimal set of operations
for sorted_bag.
The operations are create, insert, size, and increasing.
2Create*, insert, and size are procedural abstractions
that, respectively,
create a sorted_bag, insert an item into a sorted_bag,
and give the number of items in a sorted_bag.
2Increasing* is a control abstraction
that produces the items in a sorted_bag in increasing order;
each item produced is accompanied by
an integer representing the number of times
the item appears in the sorted_bag.
Note that other operations might also
be useful for sorted_bag,
for example, an iterator yielding the items
in decreasing order.
In general, the definer of a data abstraction
can provide as many operations as seems reasonable.
.para
In Figure current_figure, we give an implementation
of the sorted_bag abstraction.
.begin_figure "The sorted_bag cluster."
.code
sorted_bag = cluster [t: type] is create, insert, size, increasing
where t has equal, lt: proctype (t, t) returns (bool);
rep = record [contents: tree[t], total: int];
create = proc () returns (cvt);
return rep${contents: tree[t]$create (), total: 0};
end create;
insert = proc (sb: cvt, v: t);
sb.contents := tree[t]$insert (sb.contents, v);
sb.total := sb.total + 1;
end insert;
size = proc (sb: cvt) returns (int);
return sb.total;
end size;
increasing = iter (sb: cvt) yields (t, int);
for item: t, count: int
in tree[t]$increasing (sb.contents) do
yield item, count;
end;
end increasing;
end sorted_bag;
.ns
.end_code
.finish_figure
It is implemented using a sorted binary tree,
just as wordbag was implemented.
Thus, a subsidiary abstraction is necessary.
This abstraction, called tree, is a generalization
of the wordtree abstraction (used in Section example),
which has been parameterized to work for all ordered types.
An implementation of tree is given in Figure current_figure.
Notice that both the tree abstraction and the sorted_bag abstraction
place the same constraints on their type parameters.
.begin_page_figure "The tree cluster."
.code
tree = cluster [t: type] is create, insert, increasing
where t has equal, lt: proctype (t, t) returns (bool);
node = record [m!value: t, count: int,
(mark!m)lesser: tree[t], greater: tree[t]];
rep = oneof [empty: null, non_empty: node];
create = proc () returns (cvt);
return rep$make_empty (nil);
end create;
insert = proc (x: cvt, v: t) returns (cvt);
tagcase x
tag empty:
n: node := node${m!value: v, count: 1,
(mark!m)lesser: tree[t]$create (),
(mark!m)greater: tree[t]$create ()};
return rep$make_non_empty (n);
tag non_empty (n: node):
if t$equal (v, n.value)
then n.count := n.count + 1;
elseif t$lt (v, n.value)
then n.lesser := tree[t]$insert (n.lesser, v);
else n.greater := tree[t]$insert (n.greater, v);
end;
return x;
end;
end insert;
increasing = iter (x: cvt) yields (t, int);
tagcase x
tag empty: ;
tag non_empty (n: node):
for item: t, count: int
in tree[t]$increasing (n.lesser) do
yield item, count;
end;
yield n.value, n.count;
for item: t, count: int
in tree[t]$increasing (n.greater) do
yield item, count;
end;
end;
end increasing;
end tree;
.ns
.end_code
.finish_figure
.para
An important feature of the 2sorted_bag*
and 2tree* clusters
is the way that the cluster parameter is used in places
where the type string was used in wordbag and wordtree.
This usage is especially evident in the implementation of tree.
For example, tree has a representation that stores values of
type t: the 2value* component of a node
must be an object of type t.
.para
In the insert operation of tree,
the lt and equal operations of type t are used.
We have used the compound form, e.g. 2t$equal@(v,@n.value)*,
to emphasize that the equal operation of t is being used.
The short form, 2v@=@n.value*, could have been used instead.
.para
The increasing iterator of tree works as follows:
First it yields all items in the current tree
that are less than the item at the top node;
the items are obtained by a recursive use of itself,
passing the 2lesser* subtree as a parameter.
Next it yields the contents of the top node,
and then it yields all items in the current tree
that are greater than the item at the top node
(again by a recursive use of itself).
In this way it performs a complete walk over the tree,
yielding the values at all nodes, in increasing order.
.nr rra1 current_figure
.para
Finally, we show in Figure rra1 how the original
procedure count_words can be implemented in terms of sorted_bag.
.begin_figure "The count_words procedure using iterators."
.code
count_words = proc (i: instream, o: outstream);
wordbag = sorted_bag[string];
% create an empty wordbag
wb: wordbag := wordbag$create ();
% scan document, adding each word found to wb
for word: string in words (i) do
wordbag$insert (wb, word);
end;
% print the wordbag
total: int := wordbag$size (wb);
for w: string, count: int in wordbag$increasing (wb) do
print_word (w, count, total, o);
end;
end count_words;
.ns
.end_code
.finish_figure
.
Note that the count_words procedure now uses 2sorted_bag*[string]
instead of wordbag.
2Sorted_bag*[string] is legitimate, since the type string
provides both lt and equal operations.
Note that two for statements are used in count_words.
The second for statement prints the words
in alphabetic order,
using the increasing iterator of sorted_bag.
.ne 4
The first for statement inserts the words into the sorted_bag;
it uses an iterator
.code
words = iter (i: instream) yields (string);
dots
end words;
.widow 2
.end_code
The definition of words is left as an exercise for the reader.

112
doc/clu/clup5.r Normal file
View File

@@ -0,0 +1,112 @@
.nd chapter 5-1
.nd current_figure 11
.nd wordbag 3
.so clu/clupap.header
.sr p 2p*
.sr q 2q*
.sr x 2x*
.sr y 2y*
.sr z 2z*
.sr a 2a*
.sr b 2b*
.sr insert 2insert*
.chapter "The CLU Library"
.para
So far, we have shown CLU modules as separate pieces of
text, without explaining how they are bound together to form a
program. This section describes the CLU library, which plays a
central role in supporting inter-module references.
.para
The CLU library contains information about
abstractions. The library supports incremental program
development, one abstraction at a time, and, in addition,
makes abstractions that are defined during the construction of
one program available as a basis for subsequent program development.
The information in the library permits the separate
compilation of single modules, with complete type-checking
of all external references (such as procedure
invocations).
.para
The structure of the library derives from the fundamental
distinction between abstractions and implementations.
For each abstraction, there is a 2description
unit* which contains all system-maintained information
about that abstraction. Included in the description unit
are zero or more modules that implement the abstraction.
.foot
Other information that may be stored in the library
includes information about relationships among
abstractions, as might be expressed in a module
interconnection language [DK75,@Thomas].
.efoot
.para
The most important information contained in a description
unit is the abstraction's 2interface specification*, which
is that information needed to type-check uses of the abstraction.
For procedural and control abstractions,
this information consists of the number and types of
parameters, arguments, and output values, plus any constraints
on type parameters (i.e., required operations, as described in
Section 4). For data abstractions,
it includes the number and types of parameters, constraints on
type parameters, and the
name and interface specification of each
operation.
.para
An abstraction is entered in the library by
submitting the interface specification;
no implementations are required.
In fact, a module can be compiled before any implementations
have been provided for the abstractions that it uses;
it is necessary only that interface specifications
have been given for those abstractions.
Ultimately, there can be many implementations
of an abstraction;
each implementation is required to satisfy the
interface specification of the abstraction.
Because all uses and implementations
of an abstraction are checked against the interface
specification, the actual selection
of an implementation can be delayed
until just before (or perhaps during) execution.
We imagine a process of binding together modules
into programs, prior to execution, at which time
this selection would be made.
.para
An important detail of the CLU system is
the method by which CLU modules refer to abstractions.
To avoid problems of name conflicts that can arise in
large systems, the names used by a module to refer to
abstractions can be chosen to suit the programmer's
convenience.
When a module is submitted for
compilation, its external references must be bound to
description units so that type-checking can be
performed. The binding is accomplished by constructing
an 2association list*,
mapping names to description units, which
is passed to the compiler along with the source code when
compiling the module.
The mapping in the association list is stored by the compiler
in the library as part of the module.
A similar process is involved in entering interface
specifications of abstractions, as these will include
references to other (data) abstractions.
.para
When the compiler type-checks a module,
it uses the association list to map the external
names in the module to description units, and then uses
the interface specifications in those description
units to check that the abstractions are used correctly.
The type-correctness of the module thus
depends upon the binding of names to description units
and the interface specifications in those description
units, and could be invalidated if changes to the
binding or the interface specifications were subsequently
made. For this reason, the process of compilation
permanently binds a module to the abstractions it
uses, and the interface description of an abstraction,
'ne 2
once defined, is not allowed to change.
(Of course, a new description unit can be created
to describe a modified abstraction.)

77
doc/clu/clup6.r Normal file
View File

@@ -0,0 +1,77 @@
.nd chapter 6-1
.nd current_figure 1
.nd wordbag 3
.so clu/clupap.header
.sr insert 2insert*
.
.chapter "Implementation"
.para
This section briefly describes the current implementation of CLU
and discusses its efficiency.
.para
The implementation is based on a decision to represent
all CLU objects by 2object descriptors*,
which are fixed-size values containing a type code and some
type-dependent information.
.foot
Object descriptors are similar to capabilities [Lam71].
.efoot
In the case of mutable types, the type-dependent information
is a pointer to a separately-allocated
area containing the state information. For constant
types, the information either directly contains
the value (if the value can be encoded in the
information field, such as for integers, characters,
and booleans) or contains a pointer to separately-allocated
space (as for strings).
The type codes are used by the garbage collector
to determine the physical representation of objects
so that the accessible objects can be traced;
they are also useful for supporting program debugging.
.para
The use of fixed-size object descriptors
allows variables to be fixed-size cells. Assignment
is efficient: the object descriptor resulting
from the evaluation of the expression is simply
copied into the variable. In addition, a single
size for variables facilitates the separate compilation
of modules and allows most of the code of a
parameterized module to be shared among all instantiations
of the module. The actual parameters are made available
to this code by means of a small parameter-dependent
section, which is initialized prior to execution.
.para
Procedure invocation is relatively efficient.
A single program stack is used,
and argument passing is as efficient as assignment.
Iterators are a form of coroutine;
however, their use is sufficiently constrained
that they are implemented using just the program stack.
Using an iterator is therefore only slightly more expensive
than using a procedure.
.para
The data abstraction mechanism is not inherently
expensive. No execution time type-checking is necessary.
Furthermore, the type conversion implied by 1cvt*
is merely a change in the view taken of an object's type,
and does not require any computation.
.para
A number of optimization techniques can be
applied to a collection of modules, if one is
willing to give up the flexibility of separate
compilation. The most effective such optimization is
the inline substitution of procedure (and iterator) bodies
for invocations [Sch76].
The use of data abstractions tends to introduce
extra levels of procedure invocations that perform little or no
computation. As an example, consider the 2wordbag$insert*
operation (Figure wordbag), which merely invokes the
2wordtree$insert* operation and increments a counter.
If data abstractions had not been used, these actions would most
likely have been performed directly by the 2count_words*
procedure. The 2wordbag$insert* operation is thus
a good candidate for being compiled inline.
Once inline substitution has been performed, the increase
in context will enhance the effectiveness of
conventional optimization techniques
[All71,@All75,@Atk76].

151
doc/clu/clup7.r Normal file
View File

@@ -0,0 +1,151 @@
.nd chapter 7-1
.nd current_figure 5
.nd wordbag 3
.so clu/clupap.header
.sr p 2p*
.sr q 2q*
.sr x 2x*
.sr y 2y*
.sr z 2z*
.sr a 2a*
.sr b 2b*
.sr insert 2insert*
.chapter "Discussion"
.para
Our intent in this paper has been to provide an
informal introduction to the abstraction mechanisms in CLU.
By means of programming examples, we have illustrated the
use of data, procedural, and control abstractions, and have
shown how CLU modules are used to implement these
abstractions. We have not attempted to provide a complete
description of CLU, but, in the course of explaining
the examples, most features of the language have appeared.
One important omission is the CLU exception handling mechanism
(which does support abstractions); this mechanism
is described in [LCS75].
.para
In addition to describing constructs
that support abstraction, previous sections have
covered a number of other topics. We have discussed the
semantics of CLU. We have described the organization of the
CLU library and discussed how it supports incremental
program development and separate
compilation and type-checking of modules.
Also, we have described our current
implementation and discussed its efficiency.
.para
In designing CLU, our goal was to simplify the task
of constructing reliable software that is reasonably easy
to understand, modify, and maintain. It seems appropriate,
therefore, to conclude this paper with a discussion of how
CLU contributes to this goal.
.para
The quality of any program depends upon the skill of
the designer. In CLU programs,
this skill is reflected in the choice of abstractions.
In a good design, abstractions will be used
to simplify the connections between modules and to
encapsulate decisions that are likely to change [Par71].
Data abstractions are particularly valuable for these purposes.
For example, through the use of a data abstraction,
modules that share a system data base
rely only on its abstract behavior as
defined by the data base operations. The connections
among these modules are much simpler
than would be possible if they shared knowledge
of the format of the data base and the relationship
among its parts. In addition, the data base abstraction
can be reimplemented without affecting the code of the modules
that use it.
CLU encourages the use of data abstractions,
and thus aids the programmer during program design.
.para
The benefits arising from the use of data
abstractions are based on the constraint, inherent in CLU
and enforced by the CLU compiler, that only the operations
of the abstraction may access the representations of the objects.
This constraint ensures that the distinction made in CLU
between abstractions and implementations
applies to data abstractions as well as to procedural
and control abstractions.
.para
The distinction between abstractions and implementations
eases program modification and maintenance.
Once it has been determined that an abstraction must be
reimplemented, CLU guarantees that the code of
all modules using that
abstraction will be unaffected by the change.
The modules need not be reprogrammed or even recompiled;
only the process of
selecting the implementation of the abstraction must be
redone.
The problem of determining what modules must be
changed is also simplified, because each module has a
well-defined purpose, to implement an abstraction,
and no other module can interfere with that purpose.
.para
Understanding and verification of CLU programs is
made easier
because the distinction between
abstractions and implementations permits this task
to be decomposed.
One module at a time is studied to determine that it
implements its abstraction. This study requires
understanding the behavior of the abstractions
it uses, but it is not necessary to understand the
modules implementing those abstractions. Those
modules can be studied separately.
.para
A promising way to establish the
correctness of a program is by means of a mathematical
proof. For practical reasons, proofs should be
performed (or at least checked) by a verification
system, since the process of constructing
a proof is tedious and error-prone.
Decomposition of the proof is essential for
program proving, which is practical only for small
programs (like CLU modules). Note that when the CLU
compiler does type-checking, it is, in addition
to enforcing the constraint that permits the proof
to be decomposed, also performing a small part of the
actual proof.
.para
We have included as declarations in CLU just
the information that the compiler can check with
reasonable efficiency.
We believe that the other
information required for proofs (specifications and
assertions) should be expressed in a separate
``specification'' language.
The properties of such a language are being
studied [Guttag, Lis75, Lis76, Spitzen].
We intend eventually to add formal specifications to the
CLU system; the library is already organized to
accommodate this addition. At that time various
specification language processors could be added to
the system.
.para
We believe that the constraints imposed by
CLU are essential for practical as well as theoretical
reasons. It is true that data abstractions
can be used in any language by
establishing programming conventions to protect the
representations of objects. However, conventions are no
substitute for enforced constraints. It is inevitable
that the conventions will be violated -- and are likely
to be violated just when they are needed most, in
implementing, maintaining, and modifying large
programs. It is precisely at this time, when the
.ne 3
programming task becomes very difficult, that a
language like CLU will be most valuable and
appreciated.
.chapter "Acknowledgements"
.para
The authors gratefully acknowledge the contributions
made by members of the CLU design group over the
last three years. Several people have made
helpful comments about this paper, including
Toby Bloom, Dorothy Curtis, Mike Hammer,
Eliot Moss, Jerry Saltzer, Bob Scheifler,
and the referees.

117
doc/clu/clupap.header Normal file
View File

@@ -0,0 +1,117 @@
.
header file for clu paper
.
.nd narrow 0
.nd csg_memo 0
.nd started 0
.if ~started
.dv xgp
.fo 0 fonts; 30vr kst
.fo 1 fonts; 31vgb kst
.fo 2 fonts; 30vri kst
.fo 3 fonts; 37vrb kst
.fo 4 fonts; 75vbee kst
.fo 7 fonts; 18fg kst
.tr @
.nr fnfont 7
.nr chapter_starts_page 0
.nr reset_per_page 0
.nr both_sides 1
.sr list_left_margin 500m
.sr list_right_margin 500m
.if narrow
.nr big_font 1
.ls 1
.ll 4.25i
.pl 14i
.sr list_left_margin 200m
.sr list_right_margin 200m
.ef csg_memo==0
.ls 2
.sr left_heading Abstraction Mechanisms in CLU
.sr right_heading date
.ef csg_memo>0
.ls 1.5
.sr left_heading CSG Memo csg_memo-1
.sr right_heading Abstraction Mechanisms in CLU
.en
.sr figure_name Figure \
.sr table_name Table \
.nr immediate_figure 0
.nr tty_table_of_contents 1
.nr verbose 1
.so r/r.macros
.so clu/clukey.r
.nr started 1
.if narrow
.eq begin_page_figure begin_figure
.en
.en
.
number register for section references
.
.nr introduction 1
.nr example 2
.nr semantics 3
.nr more_abstraction 4
.nr library 5
.nr implementation 6
.nr efficiency 6
.nr discussion 7
.
paragraph macro
.
.if ls<150
.de para
.sp
.ne 3
.ti 5
.em
.ef
.de para
.br
.ne 3
.ti 5
.em
.en
.
three dots
.
.sr dots 1...*
.sr th 7th*
.
code brackets - use like table, end_table
.
.de code
.table 2
.if narrow
.ta 8 11 14 17 20 23 26 29 32
.en
.em
.
.de end_code
.rtabs
.end_table
.em
.
mark position
.
.de m
.nr mark hpos
.em
.
widow eliminator
.
.de widow <n>
.lbegin
.nv n \0
.if n<1
.nr n 1
.end
.vx n n!l 2m
.if vtrap<n&vpos>0
.nr n n-vtrap
.ct footer_handler +n!m
.end
.end
.em

24
doc/clu/clupap.r Normal file
View File

@@ -0,0 +1,24 @@
CLU paper
.
set csg_memo to 0 for paper version
set it to csg memo number for csg memo version
.
set narrow to 1 for map version
and insert ;SIZE 14 at the beginning of the XGP file
.
.nr narrow 0
.nr csg_memo 0
.
.so clu/clupap.header
.so clu/clup0.r
.so clu/clup1.r
.so clu/clup2.r
.so clu/clup3.r
.so clu/clup4.r
.so clu/clup5.r
.so clu/clup6.r
.so clu/clup7.r
.if narrow
.ns p
.en
.insert_refs