©2007 Nova Publishers.

The Case for Readable Code
Trygve Reenskaug
Dept. of Informatics, University of Oslo
Abstract
Readable code is the key to correct and maintainable programs. Pure class oriented programming does not scale and tends to
lead to code that is hard to read. Extensive subclassing is an effective obfuscator and should often be replaced with delegation.
A strategy of divide and conquer can be achieved with suitably structured components. This opens a path to readable, object
oriented programs. Pair programming and, even better, peer review are work processes that help getting it right the first time.

Introduction
There is no end to the number of different
programs that can be executed in a computer. A
program may crash or it may go into an infinite
loop. It is all the same; the machine executes the
instructions given to it.
There is almost no end to the number of programs
that will satisfy a given specification and pass the
acceptance tests. But tests only cover a minuscule
number of the potential executions. Untold
glitches and errors may lurk within the untested
parts of the code only to be found by
undisciplined users who run the program in ways
not anticipated by the testers.
There are relatively few programs that will satisfy
a given specification, pass the acceptance tests
with flying colors, and have been read,
understood and accepted by a human reader.
These are the “no surprises” programs that blend
in with the users’ work and that can be adapted to
the inevitable changes in the users’ needs.
The remainder of this comment is about how to
create one of these very desirable programs. This
gives me an opportunity to ride a dear hobbyhorse
of mine: The key to quality programs is that the
code must be readable, and in addition, that it
must actually be read.
I will first discuss why class oriented
programming makes it difficult to write readable
Trygve Reenskaug; commentary

1 of 5

code and look at ways to overcome these
problems. I end this commentary with discussing
some work processes that facilitate the writing of
readable code.

Subclassing is evil
The procedure in procedure-oriented
programming is an ideal unit for independent
reading. There is one entry point, one exit point,
and well-defined calls upon required services.
Compare with the code defining a class with its
methods. There are a potentially large number of
entry points, each having its own exit point.
Required services are invoked indirectly through
expressions that give the links to the service
providers. This can get very complex, and we will
have to restrict ourselves to the subset of all
possible classes where the code is readable and
checkable.
Subclassing is a very powerful feature of object
oriented programming. Common code can be
factored out into a superclass; common changes
can be done in the superclass and apply equally to
all its subclasses. In theory, this practice should be
straight forward, but there is a serious snag.
Subclassing adds another dimension to the already
complex class. It is often necessary to read the
superclass code to understand the behavior
specification. The services provided by the
superclass can be less than obvious and they may
have changed since a previous reading. A
12/10/2007 at 5:28 PM

reviewer can only trust the superclasses if they are
part of a trusted and well-known library.
One or more levels of superclasses may be
evolving as parts of the current project. Any
previous check of a class is invalidated when one
of its superclasses is changed. Do we recheck all
the subclasses? I hear the word “refactoring”, but
semiautomatic refactoring cannot replace code
reading in a world of careful code review.
My conclusion is that subclassing should be
severely restricted because the inevitable
superclass evolution will make us loose control. I
suggest that subclassing can often be replaced by
delegation, thus keeping the number of
inheritance levels within reasonable bounds.
I work at finding a discipline of object oriented
programming that ameliorates the obfuscation
associated with class oriented programming. I
believe we need to
•

replace most of the subclassing with
delegation,

•

enforce an object structure that gives us
readable code through a strategy of divide
and conquer.

Delegation with the
Andersen Representation objects
In his cand.scient. Thesis, Jørn Andersen
discussed how the query operations from
relational algebra can be translated into an object
oriented context [1]. His premise was that there
was a set of encapsulated, black box objects and
that the results of the queries should likewise be
sets of encapsulated, black box objects. In
relational terms, a relation became a set of
objects, a tuple became an object, and an attribute
became a message.
The SELECT operation is simple; it just returns a
subset of the original objects. A JOIN is harder,
the result should appear as instances of a new
class with attributes from both the argument
classes. Andersen’s solution was to introduce a
Representation class. An instance of this class
Trygve Reenskaug; commentary

2 of 5

associates the message names with the objects
that shall handle them. His solution is similar to
the Facade pattern [2], but he utilized the stored
program facility of Smalltalk to make all facade
objects instances of the same, Representation
class. The result was a very powerful delegation
mechanism. A Representation object will
dynamically and automatically extend its interface
when a new handling object is added. A
Representation object will dynamically and
automatically shrink its interface when an object
is removed.
The Andersen Representation object appears to
give dynamic, multiple inheritance in a simple
way. It should be explored as a readable
replacement of uncontrolled subclassing.

Divide and conquer with
components
The Facade pattern and the Andersen
Representation objects are both open constructs;
their handling objects may simultaneously
participate in other constructs. This may lead to
unduly complex structures that make the code
hard to read. I close the constructs by defining a
component as an object that encapsulates a
number of member objects. Being encapsulated, a
component is characterized by its provided and
required interfaces. Components are ideal
building blocks in a strategy of divide and
conquer since they divide the object space into
manageable parts. A reader can check the code for
each provided operation in turn. The code is local
to the component; any required services are
captured in the required interfaces. The
implementations of the required interfaces are in
other components that can be checked separately.
An Andersen Representation can serve as the port
into a component. The member objects can be
structured according to a suitable paradigm in
order to make the code more readable (and thus
reliable). An example is the DCA (DataCollaboration-Algorithm) paradigm discussed in
[3].

12/10/2007 at 5:28 PM

Testing cannot inject quality into an
inferior product
Testing has never been in the forefront of my
attention because no industry has ever been able
to test quality into an inferior product. In
programming, this means that it is too late to save
an inferior program when it gets to the testing
stage. The best is to focus on getting it right the
first time and use testing to check that we have
made no serious blunders.
A minimum requirement for a reasonable test is
that all statements shall have been executed at
least once. Otherwise, a program may pass the
tests while glaring deficiencies remain unnoticed.
This requirement is hard to satisfy in procedural
programming, and I have never been able to
satisfy it with objects. I still work at finding a
discipline of object oriented programming that
facilitates this reasonable level of testing.

Literate programming
Donald Knuth had proposed the notion of literate
programming. The main idea was to treat a
program as a piece of literature in the form of a
textbook. The text could be formatted and printed
like a book, and a filter could extract the code and
compile it. This idea fitted very well with our idea
that a program should primarily be written for
human reader and it looked like a good idea to
develop code and textbook together. So we
extended our multimedia authoring tool with two
new media: Smalltalk class definitions and
Smalltalk method definitions. Our experience
with literate programming was reported at
OOPSLA-89 [4]
A colleague and I worked together on a major
project where we wrote the code directly into a
literate programming document. This combination
of authoring and coding was very inspiring and
great fun. We were highly motivated to write
about a good idea whenever we hit upon it. We
once saw an obvious way to optimize a certain
method. We worked on it for some time before
discovering a catch; the optimization could not
work. There and then we were very motivated to
Trygve Reenskaug; commentary

3 of 5

write a warning to future maintainers directly
below the relevant code in the document. The
next day would have been too late; we were on to
other problems and other solutions.
Literate programming worked beautifully until we
got to a stage where we wanted to refactor the
program. The program structure was easy to
change, but it implied a radical change to the
structure of the book. There was no way we could
spend a great deal of time on restructuring the
book so we ended up with writing appendices and
appendices to appendices that explained what we
had done. The final book became unreadable and
only fit for the dustbin.
The lesson was that the textbook metaphor is not
applicable to program development. A textbook is
written on a stable and well known subject while
a program is under constant evolution. We
abandoned literate programming as being too
rigid for practical programming. Even if we got it
right the first time, it would have failed in the
subsequent maintenance phases of the program’s
life cycle.
Pair programming
Dijkstra is the source of many pregnant maxims
such as program testing can be used to show the
presence of bugs, but never show their absence
and nothing is cheaper than not introducing the
bugs in the first place. [5] This is all well and
good, but easier said than done.
One of the keys to success is to admit the
fallibility of humans and make sure that at least
two people are involved in the creation of a
program. One solution is to let two programmers
work together in front of a common computer. I
have one very successful experience with this
mode of working when a colleague and I
developed the base architecture and core program
of a family of software products. We spent almost
a year on the architecture before getting down to
the concrete programming. We were a very good
team; one being creative and jumping to
conclusions, the other insisting on stopping to
think whenever the code was less than obvious.
12/10/2007 at 5:28 PM

(This effort was also an exercise in literate
programming as described previously). Other
attempts at pair programming have failed because
conflicting personalities made the effort more
competitive than cooperative.
I believe that pair programming can be very
effective under the right conditions. If a fairly
stable team is working on a common body of
programs, the pairing can be varied. All team
members get to know the programs intimately
well and any of them can confidently work on the
inevitable changes and extensions.
Pair programming is still not the ideal solution
because two programmers working closely
together can easily fall into the same trap. Further,
they get intimately acquainted with the intricacies
of the programs so that the code may not be
readable for a future, uninitiated maintainer.

Peer review
I believe it was an article in the Datamation
magazine some time in the sixties that first
brought peer review to my notice. It sounded
great. I had just written a FORTRAN subroutine and
ran around to find somebody who was willing to
read and comment it. I finally persuaded a
colleague to do so.
All the benefits mentioned in the Datamation
article were attained in this first exercise. First,
my colleague pointed out that a certain statement
could be improved by using a FORTRAN feature I
was not aware of. Second, my colleague asked me
to explain the exact meaning of another statement
that he was unfamiliar with. And finally, my
colleague found a bug I would never have found
by blind testing. The program worked beautifully
for N<1000. It failed gracefully for N>1000. But
it crashed for N=1000. Careful reading of the
code might have highlighted the number 1000 as
critical so that it should have a special test. But
such reading would have revealed the bug and the
test would have been superfluous.
The outcome of this first attempt at peer review
was that both my colleague and I learnt something
Trygve Reenskaug; commentary

4 of 5

new in addition to the main result of a bug free
subroutine. All this achieved at the cost of 15
minutes proof reading.
We used peer review in all our work from that day
on. Every subroutine had two comments: One
identified the original programmer and another
identified the reader. The important feature was
that it was the reader who was responsible for the
correctness of the code. In the rare case of a bug,
the original programmer could point at the reader
and say: “your fault!”
When I read my own code, I know what it is
supposed to say and naturally assume that I have
written what I intended. My colleague has no such
pre-conception and he knows perfectly well that I
am fallible. His chance of finding a deficiency is
far better than mine when we read the same code.
I said above that code should primarily be written
for human reader. With peer review, there is an
immediate check the code is indeed readable. We
can be reasonably certain that future maintainers
can read and understand the code since the
reviewer has already done so.

Conclusion
Program testing can never show the absence of
bugs. Indeed, the more bugs we find during
testing, the more bugs remain in the shipped
product. (Because a given test regimen can only
find a certain percentage of all bugs). Contrast
with a competent reviewer who reads all the code
and can reveal bugs, glitches, clumsy code, and
potential traps for a future maintainer.
Effective code reading is only feasible if the code
is partitioned into reasonably independent chunks
and if the remaining dependencies are well
defined. I have tried using peer review in object
oriented programming, but have as yet not
succeeded because I have not been able to
partition the system into reasonable chunks. I
expect that some changes to my programming
method will help:
•

Subclassing must be kept to a minimum
both to reduce system complexity and to
12/10/2007 at 5:28 PM

make it less vulnerable to the inevitable
program evolution.
•

Subclassing can often be replaced by
delegation.

•

Chunking objects into components with a
corresponding chunking of the code is
essential.

•

Literate programming is tempting for an
example educational program, but is too
rigid for general programming.

•

Pair programming is powerful but may not
lead to chunked and readable code.

My pet idea is peer review. May it become an
essential element in the creation of quality
programs.

Scient. Thesis; Department of
informatics, University of Oslo; 1991.
2

Gamma, E,; Helm, R,; Johnson, R;
Vlissides, J. Design Patterns; ISBN 0201-63361-2; Addison-Wesley,
Reading, MA. 1995.

3

See my chapter on Programming with
roles and classes in this book.

4

Reenskaug, T.; Skaar, A. L. An
environment for literate Smalltalk
programming; OOPSLA 1989; ISBN:
0-89791-333-7; ACM Press; New
York, NY 1989; pp. 337–345.

5

For more Dijkstra contributions to
computing, see E. W. Dijkstra
Archive; University of Texas; [web
page]
http://www.cs.utexas.edu/users/EWD/

References
1

Andersen, J. Queries on sets in an
Object-Oriented Database; Cand.

Trygve Reenskaug; commentary

5 of 5

12/10/2007 at 5:28 PM