.fp 1 PA
.fp 2 PI
.fp 3 PB
.ds pf CW
.so /usr/bwk/src/cprog.mac
.if n .ls 2
....ND "April 2, 1981"
....TM 81-11272-12 11173 39199-11
.TR 100
.TL
Why Pascal is Not My Favorite Programming Language
.AU "MH 2C-518" 6021
Brian W. Kernighan
.AI
.MH
.AB
.PP
The programming language Pascal has become the dominant language
of instruction in computer science education.
It has also strongly influenced languages developed subsequently,
in particular Ada.
.PP
Pascal was originally intended primarily as a teaching language,
but it has been more and more often recommended as a language
for serious programming as well,
for example, for system programming tasks and even operating systems.
.PP
Pascal, at least in its standard form, is just plain not suitable for serious programming.
This paper discusses my personal discovery of some of the reasons why.
.AE
.CS 1 0 1 0 0  18
.NH
Genesis
.PP
This paper has its origins in two events \(em
a spate of papers that compare C and Pascal
.[
%T A Comparison of the Programming Languages C and Pascal \(em Part I: Language Concepts
%A Feuer, A. R.
%A N. H. Gehani
%D September 1979
%R Bell Labs internal memorandum
.]
.[
%T A Comparison of the Programming Languages C and Pascal \(em Part II: Program Properties and Programming Domains
%A N. H. Gehani
%A A. R. Feuer
%D February 1980
%R Bell Labs internal memorandum
.]
.[
%T Pascal versus C: A Subjective Comparison
%A P. Mateti
%J Language Design and Programming Methodology Symposium
%O Sydney, Australia
%D September 1979
%I Springer-Verlag
.]
.[
%T A Comparison of Language C and Pascal
%A A. Springer
%R IBM Technical Report G320-2128, Cambridge Scientific Center
%D August 1979
.]
and a personal attempt to rewrite
.ul
Software Tools
.[
kernighan plauger software tools addison
.]
in Pascal.
.PP
Comparing C and Pascal is rather like comparing
a Learjet to a Piper Cub \(em one is meant for getting something done while
the other is meant for learning \(em
so such comparisons tend to be somewhat farfetched.
But the revision of
.ul
Software Tools
seems a more relevant comparison.
The programs therein were originally written
in Ratfor,
a ``structured'' dialect of Fortran implemented by a preprocessor.
Since Ratfor is really Fortran in disguise, it has few of the
assets that Pascal brings \(em
data types more suited to character processing,
data structuring capabilities for better defining
the organization of one's data,
and strong typing to enforce telling the truth
about the data.
.PP
It turned out to be harder than I had expected
to rewrite the programs in Pascal.
This paper is an attempt to distill out of the experience
some lessons about Pascal's suitability for
programming (as distinguished from learning about programming).
It is
.ul
not
a comparison of Pascal with C or Ratfor.
.PP
The programs were first written in that dialect
of Pascal supported by the Pascal interpreter
.ul
pi
provided by the University of California at Berkeley.
The language is close to the nominal standard
of Jensen and Wirth,
.[
jensen wirth pascal report 1978
%O (2nd edition.)
.]
with good diagnostics and careful run-time checking.
Since then, the programs have also been run, unchanged
except for new libraries of primitives, on four other systems:
an interpreter
from the Free University of Amsterdam
(hereinafter referred to as VU, for Vrije Universiteit),
a VAX version of the Berkeley system (a true compiler),
a compiler purveyed by Whitesmiths, Ltd.,
and UCSD Pascal on a Z80.
All but the last of these Pascal systems are written in C.
.PP
Pascal is a much-discussed language.
A recent bibliography
.[
%A David V. Moffat
%T A Categorized Pascal Bibliography
%J SIGPLAN Notices
%V 15
%N 10
%P 63-75
%D October 1980
.]
lists 175 items under the heading of
``discussion, analysis and debate.''
The most often cited papers (well worth reading) are
a strong critique by Habermann
.[
habermann pascal critical
.]
and an equally strong rejoinder by
Lecarme and Desjardins.
.[
%A O. Lecarme
%A P. Desjardins
%T More Comments on the Programming Language Pascal
%J Acta Informatica
%V 4
%P 231-243
%D 1975
.]
The paper by Boom and DeJong
.[
%T A Critical Comparison of Several Programming Language Implementations
%A H. J. Boom
%A E. DeJong
%J Software Practice and Experience
%V 10
%N 6
%D June 1980
%P 435-473
.]
is also good reading.
Wirth's own assessment of Pascal is found in
.[ [
%T An Assessment of the Programming Language Pascal
%A N. Wirth
%J IEEE Transactions on Software Engineering
%V SE-1
%N 2
%D June, 1975
%P 192-198
.]].
I have no desire or ability to summarize the literature;
this paper represents my personal observations
and most of it necessarily duplicates points made by others.
I have tried to organize the rest of the material
around the issues of
.DS
types and scope
control flow
environment
cosmetics
.DE
and within each area more or less in decreasing order
of significance.
.PP
To state my conclusions at the outset:
Pascal may be an admirable language for teaching beginners
how to program; I have no first-hand experience with that.
It was a considerable achievement for 1968.
It has certainly influenced the design of recent languages,
of which Ada is likely to be the most important.
But in its standard form (both current and proposed),
Pascal is not adequate for writing real programs.
It is suitable only for small, self-contained programs
that have only trivial interactions with their environment
and that make no use of any software written by anyone else.
.NH
Types and Scopes
.PP
Pascal is (almost) a strongly typed language.
Roughly speaking, that means that each object
in a program has a well-defined type which implicitly
defines the legal values of and operations on the object.
The language guarantees that it will prohibit illegal values and operations,
by some mixture of compile- and run-time checking.
Of course compilers may not actually do all the checking implied in
the language definition.
Furthermore, strong typing is
.ul
not
to be confused with dimensional analysis.
If one defines 
types
.UL apple
and
.UL orange
with
.P1
type
	apple = integer;
	orange = integer;
.P2
then any arbitrary arithmetic expression involving
.UL apples
and
.UL oranges 
is perfectly legal.
.PP
Strong typing shows up in a variety of ways.
For instance, arguments to functions and procedures are checked
for proper type matching.
Gone is the Fortran freedom to
pass a floating point number into a subroutine that expects an integer;
this I deem a desirable attribute of Pascal,
since it warns of a construction that will certainly cause an error.
.PP
Integer variables may be declared to have an associated range
of legal values, and the compiler and run-time support
ensure that one does not put large integers
into variables that only hold small ones.
This too seems like a service,
although of course run-time checking does exact a penalty.
.PP
Let us move on to 
some problems of type and scope.
.NH 2
The size of an array is part of its type
.PP
If one declares
.P1
var	arr10 : array [1..10] of integer;
	arr20 : array [1..20] of integer;
.P2
then
.UL arr10
and
.UL arr20
are arrays of 10 and 20 integers respectively.
Suppose we want to write a procedure
.UL sort
to sort an integer array.
Because
.UL arr10
and
.UL arr20
have different types,
it is not possible to write a single procedure
that will sort them both.
.PP
The place where this affects
.ul
Software Tools
particularly, and I think programs in general,
is that it makes it difficult indeed
to create a library of routines for doing common, general-purpose operations
like sorting.
.PP
The particular data type most often affected is
.UL array
.UL of
.UL char ,
for in Pascal a string is an array of characters.
Consider writing a function
.UL index(s,c)
that will return the position in the string
.UL s
where the character
.UL c
first occurs, or zero if it does not.
The problem is how to handle the string argument of
.UL index .
The calls
.UL index('hello',c)
and
.UL index('goodbye',c)
cannot both be legal,
since the strings have different lengths.
(I pass over the question of how the end of a constant string
like
.UL 'hello'
can be detected, because it can't.)
.PP
The next try is
.P1
var temp : array [1..10] of char;

temp := 'hello';

n := index(temp,c);
.P2
but the assignment to
.UL temp
is illegal because
.UL \&'hello'
and
.UL temp
are of different lengths.
.PP
The only escape from this infinite regress
is to define a family of routines
with a member for each possible string size,
or to make all strings
(including constant strings like
.UL 'define' )
of the same length.
.PP
The latter approach is the lesser of two great evils.
In
.IT Tools ,
a type called
.UL string
is declared as
.P1
type string = array [1..MAXSTR] of char;
.P2
where
the constant
.UL MAXSTR
is ``big enough,''
and
all strings in all programs are exactly this size.
This is far from ideal,
although it made it possible to get the programs running.
It does
.ul
not
solve the problem of creating true libraries of useful routines.
.PP
There are some situations where it is simply not acceptable
to use the fixed-size array representation.
For example, the
.ul
Tools
program to sort lines of text
operates by fill\%ing up memory with as many lines as will fit;
its running time depends strongly on how full the memory can be packed.
Thus for
.UL sort ,
another representation is used,
a long array of characters and a set of indices into this array:
.P1
type	charbuf = array [1..MAXBUF] of char;
	charindex = array [1..MAXINDEX] of 0..MAXBUF;
.P2
But
the procedures and
functions written to process the fixed-length representation
cannot be used with the variable-length form;
an entirely new set of routines is needed to
copy and compare strings in this representation.
In Fortran or C the same functions could be used for both.
.PP
As suggested above, a constant string is written as
.P1
\&'this is a string'
.P2
and has the type
.UL packed
.UL array
.UL [1..n]
.UL of
.UL char ,
where
.UL n
is the length.
Thus each string literal of different length has a different type.
The only way to write a routine that will print a message and clean up
is to pad all messages out to the same maximum length:
.P1
error('short message                    ');
error('this is a somewhat longer message');
.P2
.PP
Many commercial Pascal compilers provide a
.UL string
data type that explicitly avoids the problem;
.UL string 's
are all taken to be the same type regardless of size.
This solves the problem for this single data type,
but no other.
It also fails to solve secondary problems
like computing the length of a constant string;
another built-in function is the usual solution.
.PP
Pascal enthusiasts
often claim that
to cope with the array-size problem
one merely has to copy some library routine
and fill in the parameters for the program
at hand,
but the defense sounds weak at best:
.[
%A O. Lecarme
%A P. Desjardins
%T ibid
%O p. 239
.]
.QS
``Since the bounds of an array are part of its type
(or, more exactly, of the type of its indexes),
it is impossible to define a procedure or function
which applies to arrays with differing bounds.
Although this restriction may appear to be a severe one,
the experiences we have had with Pascal tend to show that it
tends to occur very infrequently.
[...]
However, the need to bind the size of parametric arrays
is a serious defect in connection with the use of program libraries.''
.QE
.PP
This botch is the biggest single problem with Pascal.
I believe that if it could be fixed,
the language would be an order of magnitude more usable.
The proposed ISO standard for Pascal
.[
%A A. M. Addyman
%T A Draft Proposal for Pascal
%J SIGPLAN Notices
%D April 1980
%V 15
%N 4
%P 1-66
.]
provides such a fix
(``conformant array schemas''),
but the acceptance of this part of the standard is apparently
still in doubt.
.NH 2
There are no static variables and no initialization
.PP
A
.ul
static
variable
(often called an
.ul
own
variable in Algol-speaking countries)
is one that is private to some routine
and retains its value from one call of the routine to the next.
.ul
De facto,
Fortran variables are internal static, except for COMMON;\(dg
.FS
\(dg Strictly speaking, in Fortran 77 one must use
.UL SAVE 
to force the static attribute.
.FE
in C there is a
.UL static
declaration that can be applied to local variables.
.PP
Pascal has no such storage class.
This means that if a Pascal function or procedure intends to remember
a value from one call to another, the variable used
must be external to the function or procedure.
Thus it must be visible to other procedures,
and its name must be unique in the larger scope.
A simple example of the problem is a random number generator:
the value used to compute the current output must be saved
to compute the next one,
so it must be stored in a variable whose lifetime includes
all calls of the random number generator.
In practice, this is typically the outermost block of the program.
Thus the declaration of such a variable is far removed from the place
where it is actually used.
.PP
One example comes from the text formatter described in Chapter 7 of
.ul
Tools.
The variable 
.UL dir
controls the direction from which
excess blanks are inserted during line justification,
to obtain left and right alternately.
In Pascal, the code looks like this:
.P1
program formatter (...);

var
	dir : 0..1;	{ direction to add extra spaces }
	.
	.
	.
procedure justify (...);
begin
	dir := 1 - dir;	{ opposite direction from last time }
	...
end;

	...

begin { main routine of formatter }
	dir := 0;
	...
end;
.P2
The declaration,
initialization
and use
of the variable
.UL dir
are scattered all over the program,
literally hundreds of lines apart.
In C or Fortran,
.UL dir
can be made private to the only routine
that needs to know about it:
.P1
	...
main()
{
	...
}

	...

justify()
{
	static int dir = 0;

	dir = 1 - dir;
	...
}
.P2
.PP
There are of course many other examples of the same problem on a larger scale;
functions for buffered I/O,
storage management,
and symbol tables all spring to mind.
.PP
There are at least two related problems.
Pascal provides no way to 
initialize variables
statically
(i.e., at compile time);
there is nothing analogous to Fortran's
.UL DATA
statement or initializers like
.P1
int dir = 0;
.P2
in C.
This means that a Pascal program must contain explicit assignment statements
to initialize variables
(like the
.P1
dir := 0;
.P2
above).
This code makes the program source text bigger,
and the program itself bigger at run time.
.PP
Furthermore, the lack of initializers exacerbates the problem
of too-large scope caused by the lack of a static storage class.
The time to initialize
things is at the beginning,
so either the main routine itself begins with a lot of initialization code,
or it calls one or more routines to do the initializations.
In either case,
variables to be initialized
must be visible,
which means in effect at the highest level
of the hierarchy.
The result is that any variable that is to be initialized
has global scope.
.PP
The third difficulty is that there is no way
for two routines to share a variable unless
it is declared at or above their least common ancestor.
Fortran 
.UC COMMON
and
C's external static storage class both provide
a way for two routines to cooperate privately,
without sharing information with their ancestors.
.PP
The new standard does not offer static variables, initialization or non-hierarchical communication.
.NH 2
Related program components must be kept separate
.PP
Since the original Pascal was implemented with a one-pass compiler,
the language believes strongly in declaration before use.
In particular, procedures and functions must be declared (body and all)
before they are used.
The result is that a typical Pascal program reads from the bottom up \(em
all the procedures and functions are displayed before any of the code that
calls them,
at all levels.
This is essentially opposite to the order in which
the functions are designed and used.
.PP
To some extent this can be mitigated by a mechanism like the
.UL #include
facility of C and Ratfor:
source files can be included where needed without cluttering
up the program.
.UL #include
is not part of standard Pascal, although the UCB, VU and Whitesmiths compilers
all provide it.
.PP
There is also a
.UL forward
declaration in Pascal that permits separating the
declaration of the function or procedure header from
the body;
it is intended for defining mutually recursive procedures.
When the body is declared later on,
the header on that may contain only the function name,
and must not repeat the information from the first instance.
.PP
A related problem is that Pascal has a strict order
in which it is willing to accept declarations.
Each procedure or function consists of
.P1
.ta .6i
label	\f2label declarations, if any\fP
const	\f2constant declarations, if any\fP
type	\f2type declarations, if any\fP
var	\f2variable declarations, if any\fP
procedure \f2and\fP function \f2declarations, if any\fP
begin
	\f2body of function or procedure\fP
end
.P2
This means that all declarations of one kind (types, for instance)
must be grouped together for the convenience of the compiler,
even when the programmer would like to keep together
things that are logically related
so as to understand the program better.
Since a program has to be presented to the compiler all at once,
it is rarely possible to keep
the declaration, initialization and use of types and variables
close together.
Even some of the most dedicated Pascal supporters agree:
.[
%A J. Welsh
%A W. J. Sneeringer
%A C. A. R. Hoare
%T Ambiguities and Insecurities in Pascal
%J Software Practice and Experience
%V 7
%P 685-696
%D 1977
.]
.QS
``The inability to make such groupings
in structuring large programs
is one of Pascal's most frustrating limitations.''
.QE
.LP
A file inclusion facility helps only a little here.
.PP
The new standard does not relax the requirements
on the order of declarations.
.NH 2
There is no separate compilation
.PP
The ``official'' Pascal language does not provide separate compilation,
and so each implementation decides on its own what to do.
Some (the Berkeley interpreter, for instance) disallow it entirely;
this is closest to the spirit of the language
and matches the letter exactly.
Many others provide a declaration
that specifies that the body of a function is externally defined.
In any case, all such mechanisms are non-standard, and thus done
differently by different systems.
.PP
Theoretically, there is no need for separate compilation \(em
if one's compiler is very fast
(and if the source for all routines is always available
and if one's compiler has a file inclusion facility so that multiple copies
of source are not needed),
recompiling everything is equivalent.
In practice, of course,
compilers are never fast enough
and source is often hidden
and file inclusion is not part of the language,
so changes are time-consuming.
.PP
Some systems permit separate compilation
but do not validate consistency of types across the boundary.
This creates a giant hole in the strong typing.
(Most other languages do no cross-compilation checking either,
so Pascal is not inferior in this respect.)
I have seen at least one paper (mercifully unpublished) that on page
.IT n
castigates C for
failing to check types across separate compilation boundaries
while suggesting on page
.IT n +1
that the way to cope with Pascal
is to compile procedures separately
to avoid type checking.
.PP
The new standard does not offer separate compilation.
.NH 2
Some miscellaneous problems of type and scope
.PP
Most of the following points are minor irritations,
but I have to stick them in somewhere.
.PP
It is not legal to name a non-basic type
as the literal formal parameter of a procedure;
the following is not allowed:
.P1
procedure add10 (var a : array [1..10] of integer);
.P2
Rather, one must invent a type name, make a type declaration,
and declare the formal parameter to be an instance of that type:
.P1
type	a10 = array [1..10] of integer;
\&...
procedure add10 (var a : a10);
.P2
Naturally the type declaration is physically separated
from the procedure that uses it.
The discipline of inventing type names is helpful
for types that are used often,
but it is a distraction for things used only once.
.PP
It is nice to have the declaration
.UL var
for formal parameters of functions and procedures;
the procedure clearly states that it intends to modify
the argument.
But the calling program has no way to declare that a variable
is to be modified \(em
the information is only in one place, while two places would be better.
(Half a loaf is better than none, though \(em
Fortran tells the user nothing about who will do what
to variables.)
.PP
It is also a minor bother that arrays are passed by value
by default \(em the net effect is that every array
parameter is declared
.UL var
by the programmer more or less without thinking.
If the
.UL var
declaration is inadvertently omitted,
the resulting bug is subtle.
.PP
Pascal's
.UL set
construct seems like a good idea,
providing notational convenience and some free type checking.
For example, a set of tests like
.P1
if (c = blank) or (c = tab) or (c = newline) then ...
.P2
can be written rather more clearly and perhaps more efficiently as
.P1
if c in [blank, tab, newline] then ...
.P2
But in practice, set types are not useful for much more than
this, because the size of a set is strongly implementation
dependent (probably because it was so in the original CDC implementation:
59 bits).
For example, it is natural to attempt to write the function
.UL isalphanum(c)
(``is 
.UL c
alphanumeric?'') as
.P1
{ isalphanum(c) -- true if c is letter or digit }
function isalphanum (c : char) : boolean;
begin
	isalphanum := c in ['a'..'z', 'A'..'Z', '0'..'9']
end;
.P2
But in many implementations of Pascal (including the original)
this code fails because sets are just too small.
Accordingly,
sets are generally best left unused if one intends to write
portable programs.
(This specific routine also runs an order of magnitude slower
with sets than with a range test or array reference.)
.NH 2
There is no escape
.PP
There is no way to override the type mechanism when necessary,
nothing analogous to the ``cast'' mechanism in C.
This means that it is not possible to write programs like
storage allocators or I/O systems in Pascal,
because there is no way to talk about the type of object
that they return,
and no way to force such objects into an arbitrary type for
another use.
(Strictly speaking, there is a large hole in the type-checking
near variant records, through which some otherwise illegal 
type mismatches can be obtained.)
.NH
Control Flow
.PP
The control flow deficiencies of Pascal are minor but numerous
\(em
the death of a thousand cuts,
rather than a single blow to a vital spot.
.PP
There is no guaranteed order of evaluation of the logical operators
.UL and
and
.UL or
\(em nothing like
.UL && 
and
.UL ||
in C.
This failing, which is shared with most other languages,
hurts most often in loop control:
.P1
while (i <= XMAX) and (x[i] > 0) do ...
.P2
is extremely unwise Pascal usage,
since there is no way to ensure that 
.UL i
is tested before
.UL x[i]
is.
.PP
By the way, the parentheses in this code are mandatory \(em
the language has only four levels of operator precedence,
with relationals at the bottom.
.PP
There is no
.UL break
statement for exiting loops.
This is consistent with the one entry-one exit philosophy
espoused by proponents of structured programming,
but it does lead to nasty circumlocutions or duplicated code,
particularly when coupled with the inability to control the order in which
logical expressions are evaluated.
Consider this common situation, expressed in C or Ratfor:
.P1
while (getnext(...)) {
	if (something)
		break
	rest of loop
}
.P2
With no
.UL break
statement, the first attempt in Pascal is
.P1
done := false;
while (not done) and (getnext(...)) do
	if something then
		done := true
	else begin
		rest of loop
	end
.P2
But this doesn't work, because there is no way to force
the 
.UL not "" ``
.UL done ''
to be evaluated before the
next call of
.UL getnext .
This leads, after several false starts, to
.P1
done := false;
while not done do begin
	done := getnext(...);
	if something then
		done := true
	else if not done then begin
		rest of loop
	end
end
.P2
Of course recidivists can use a
.UL goto
and a label (numeric only and it has to be declared) to exit a loop.
Otherwise, early exits are a pain,
almost always requiring the invention of a boolean variable
and a certain amount of cunning.
Compare finding the last non-blank in an array in
Ratfor:
.P1
for (i = max; i > 0; i = i - 1)
	if (arr(i) != ' ')
		break
.P2
with Pascal:
.P1
done := false;
i := max;
while (i > 0) and (not done) do
	if arr[i] = ' ' then
		i := i - 1
	else
		done := true;
.P2
.PP
The index of a
.UL for
loop is undefined outside the loop, so
it is not possible to figure out whether one 
went to the end or not.
The increment of a
.UL for
loop can only be
.UL +1
or
.UL -1 ,
a minor restriction.
.PP
There is no 
.UL return
statement,
again for one in-one out reasons.
A function value is returned by setting the value of a pseudo-variable
(as in Fortran), then falling off the end of the function.
This sometimes leads to contortions to make sure that all paths
actually get to the end of the function with the proper value.
There is also no standard way to terminate execution
except by reaching the end of the outermost block,
although
many implementations
provide a
.UL halt
that causes immediate termination.
.PP
The
.UL case
statement is better designed than in C,
.ul
except
that there is no
.UL default
clause and the behavior is undefined 
if the input expression does not match any of the cases.
This crucial omission
renders the 
.UL case
construct almost worthless.
In over 6000 lines of Pascal in
.IT "Software Tools in Pascal" ,
I used it only four times,
although if there had been a
.UL default ,
a
.UL case
would have served in at least a dozen places.
.PP
The new standard offers no relief on any of these points.
.NH
The Environment
.PP
The Pascal run-time environment is relatively sparse,
and there is no extension mechanism except perhaps source-level
libraries in the ``official'' language.
.PP
Pascal's built-in I/O has a deservedly bad reputation.
It believes strongly in record-oriented input
and output.
It also has a look-ahead convention that is hard to
implement properly in an interactive environment.
Basically, the problem is that the I/O system believes that
it must read one record ahead of the record
that is being processed.
In an interactive system, this means that when a program is started,
its first operation is to try to read the terminal for the first
line of input,
before any of the program itself has been executed.
But in the program
.P1
write('Please enter your name: ');
read(name);
\&...
.P2
read-ahead causes the program to hang, waiting
for input before printing the prompt that asks for it.
.PP
It is possible to escape most of the evil effects of
this I/O design by very careful implementation,
but not all Pascal systems do so,
and in any case it is relatively costly.
.PP
The I/O design reflects the original operating system
upon which Pascal was designed;
even Wirth acknowledges that bias,
though not its defects.
.[
%A N. Wirth
%J ibid.
%O p. 196
.]
It is assumed that text files consist of records,
that is, lines of text.
When the last character of a line is read, the 
built-in function
.UL eoln
becomes true;
at that point, one must call
.UL readln
to initiate reading a new line
and reset
.UL eoln .
Similarly, when the last character of the file is read,
the built-in
.UL eof
becomes true.
In both cases,
.UL eoln
and
.UL eof
must be tested before each
.UL read
rather than after.
.PP
Given this, considerable pains must be taken to simulate sensible input.
This implementation of
.UL getc
works for Berkeley and VU I/O systems,
but may not necessarily work for anything else:
.P1
{ getc -- read character from standard input }
function getc (var c : character) : character;
var
	ch : char;
begin
	if eof then
		c := ENDFILE
	else if eoln then begin
		readln;
		c := NEWLINE
	end
	else begin
		read(ch);
		c := ord(ch)
	end;
	getc := c
end;
.P2
The type
.UL character
is not the same as
.UL char ,
since
.UL ENDFILE
and perhaps
.UL NEWLINE
are not legal values
for a
.UL char
variable.
.PP
There is no notion at all of access to a file system except for predefined files
named by (in effect) logical unit number
in the 
.UL program
statement that begins each program.
This apparently reflects the CDC batch system in which
Pascal was originally developed.
A file variable
.P1
var	fv : file of \f2type\fP
.P2
is a very special kind of object \(em
it cannot be assigned to, nor used except by calls to built-in procedures
like
.UL eof ,
.UL eoln ,
.UL read ,
.UL write ,
.UL reset
and
.UL rewrite .
.UL reset "" (
rewinds a file and makes it ready for re-reading;
.UL rewrite
makes a file ready for writing.)
.PP
Most implementations of Pascal provide an escape hatch
to allow access to files by name from the outside environment,
but not conveniently
and not standardly.
For example, many systems permit a filename argument
in calls to
.UL reset
and
.UL rewrite :
.P1
reset(fv, filename);
.P2
But
.UL reset
and
.UL rewrite
are procedures, not functions \(em
there is no status return and
no way to regain control if for some reason
the attempted access fails.
(UCSD provides a compile-time flag
that disables the normal abort.)
And since
.UL fv 's
can not appear in expressions like
.P1
reset(fv, filename);
if fv = failure then ...
.P2
there is no escape in that direction either.
This straitjacket makes it essentially impossible to write programs
that recover from mis-spelled file names, etc.
I never solved it adequately in the
.ul
Tools
revision.
.PP
There is no notion of access to command-line arguments,
again probably reflecting Pascal's batch-processing origins.
Local routines may allow it by adding non-standard procedures to the environment.
.PP
Since it is not possible to write a general-purpose storage allocator
in Pascal
(there being no way to talk about the types that such a function
would return),
the language has a built-in procedure
called
.UL new
that allocates space from a heap.
Only defined types may be allocated,
so it is not possible to allocate, for example,
arrays of arbitrary size to hold character strings.
The pointers returned by
.UL new
may be passed around but not manipulated:
there is no pointer arithmetic.
There is no way to regain control if storage runs out.
.PP
The new standard offers no change in any of these areas.
.NH
Cosmetic Issues
.PP
Most of these issues are irksome to an experienced programmer,
and some are probably a nuisance even to beginners.
All can be lived with.
.PP
Pascal, in common with most other Algol-inspired languages,
uses the semicolon as a statement separator rather than a terminator
(as it is in PL/I and C).
As a result one must have a reasonably sophisticated notion
of what a statement is to put semicolons in properly.
Perhaps more important, if one is serious about using them
in the proper places, a fair amount of nuisance editing is needed.
Consider the first cut at a program:
.P1
if a then
	b;
c;
.P2
But if something must be inserted before
.UL b ,
it no longer needs a semicolon, because it now precedes an
.UL end :
.P1
if a then begin
	b0;
	b
end;
c;
.P2
Now if we add an
.UL else ,
we
.ul
must
remove the semicolon on the
.UL end :
.P1
if a then begin
	b0;
	b
end
else
	d;
c;
.P2
And so on and so on,
with semicolons rippling up and down the program
as it evolves.
.PP
One generally accepted experimental result in programmer psychology
is that semicolon as separator is about ten times
more prone to error than semicolon as terminator.
.[
%A J. D. Gannon
%A J. J. Horning
%T Language Design for Programming Reliability
%J IEEE Trans. Software Engineering
%V SE-1
%N 2
%D June 1975
%P 179-191
.]
(In Ada,
.[
%T Rationale for the Design of the Ada Programming Language
%A J. D. Ichbiah, et al
%J SIGPLAN Notices
%D June 1979
%V 14
%N 6
.]
the most significant language based on Pascal,
semicolon is a terminator.)
Fortunately, in Pascal one can almost always close one's eyes
and get away with a semicolon as a terminator.
The exceptions are in places like declarations, where the separator vs. terminator problem
doesn't seem as serious anyway,
and just before
.UL else ,
which is easy to remember.
.PP
C and Ratfor programmers find
.UL begin
and
.UL end
bulky compared to
.UL {
and
.UL } .
.PP
A function name by itself is a call of that function;
there is no way to distinguish such a function call from a simple variable
except by knowing the names of the functions.
Pascal uses the Fortran trick
of having the function name act like a variable within the function,
except that where in Fortran the function name really is a variable,
and can appear in expressions,
in Pascal, its appearance in an expression
is a recursive invocation:
if
.UL f
is a zero-argument function,
.UL f:=f+1
is a recursive call of
.UL f .
.PP
There is a paucity of operators (probably related to the
paucity of precedence levels).
In particular, there are no bit-manipulation operators
.UC AND , (
.UC OR ,
.UC XOR ,
etc.).
I simply gave up trying to write the following trivial encryption program
in Pascal:
.P1
i := 1;
while getc(c) <> ENDFILE do begin
	putc(xor(c, key[i]));
	i := i mod keylen + 1
end
.P2
because I couldn't write a sensible
.UL xor
function.
The set types help a bit here (so to speak), but not enough;
people who claim that Pascal is a system programming language
have generally overlooked this point.
For example,
.[ [
%A J. Welsh
%A W. J. Sneeringer
%A C. A. R. Hoare
%T ibid
.], p. 685]
.QS
``Pascal is at the present time
[1977]
the best language in the public domain
for purposes of system programming and software implementation.''
.QE
seems a bit naive.
.PP
There is no null string, 
perhaps because Pascal uses the
doubled quote notation to indicate a quote embedded in a string:
.P1
\&'This is a '' character'
.P2
There is no way to put non-graphic symbols into strings.
In fact, non-graphic characters are unpersons in a stronger sense,
since they are not mentioned in any part of the standard language.
Concepts like newlines, tabs, and so on are handled on each system
in an
.ul
ad hoc
manner,
usually by knowing something about the character set
(e.g., ASCII newline has decimal value 10).
.PP
There is no macro processor.
The
.UL const
mechanism for defining manifest constants
takes care of about 95 percent of the uses of
simple
.UL #define 
statements in C,
but more involved ones are hopeless.
It is certainly possible to put a macro preprocessor on a Pascal compiler.
This allowed me to simulate a sensible 
.UL error
procedure as
.P1
#define  error(s)  begin writeln(s); halt end
.P2
.UL halt "" (
in turn might be defined as a branch to the end of
the outermost block.)
Then calls like
.P1
	error('little string');
	error('much bigger string');
.P2
work
since
.UL writeln
(as part of the standard Pascal environment)
can handle strings of any size.
It is unfortunate that there is no way to make this convenience
available to routines in general.
.PP
The
language prohibits expressions in declarations, so it is not possible
to write things like
.P1
const	SIZE = 10;
type		arr = array [1..SIZE+1] of integer;
.P2
or even simpler ones like
.P1
const	SIZE = 10;
		SIZE1 = SIZE + 1;
.P2
.NH
Perspective
.PP
The effort to rewrite the programs in
.ul
Software Tools
started in March, 1980,
and, in fits and starts, lasted until January, 1981.
The final product
.[
%A B. W. Kernighan
%A P. J. Plauger
%T Software Tools in Pascal
%I Addison-Wesley
%D 1981
.]
was published in June, 1981.
During that time I gradually adapted to most of
the superficial problems with Pascal
(cosmetics, the inadequacies of control flow),
and developed imperfect solutions to the significant ones
(array sizes, run-time environment).
.PP
The programs in the book are meant to be complete, well-engineered
programs that do non-trivial tasks.
But they do not have to be efficient,
nor are their interactions with the operating system
very complicated,
so I was able to get by with some pretty kludgy solutions,
ones that simply wouldn't work for real programs.
.PP
There is no significant way in which I found Pascal superior to C,
but there are several places where it is
a clear improvement over Ratfor.
Most obvious by far is recursion:
several programs are much cleaner when written recursively,
notably the pattern-search, quicksort, and expression evaluation.
.PP
Enumeration data types are a good idea.
They simultaneously delimit the range of legal values
and document them.
Records help to group related variables.
I found relatively little use for pointers.
.PP
Boolean variables are nicer than integers for Boolean conditions;
the original Ratfor programs contained some unnatural constructions because
Fortran's logical variables are badly designed.
.PP
Occasionally Pascal's type checking would warn of a slip
of the hand in writing a program;
the run-time checking of values also indicated errors
from time to time, particularly subscript range violations.
.PP
Turning to the negative side, recompiling a large program from scratch to change a single line
of source is extremely tiresome;
separate compilation, with or without type checking,
is mandatory for large programs.
.PP
I derived little benefit from the fact that
characters are part of Pascal and not part of Fortran,
because the Pascal treatment of strings and non-graphics
is so inadequate.
In both languages, it is appallingly clumsy to initialize literal strings
for tables of keywords, error messages, and the like.
.PP
The finished programs are in general about the same number
of source lines as their Ratfor equivalents.
At first this surprised me,
since my preconception was that Pascal is a wordier and less expressive
language.
The real reason seems to be that Pascal permits arbitrary expressions
in places like loop limits and subscripts
where Fortran (that is, portable Fortran 66) does not,
so some useless assignments can be eliminated;
furthermore, the Ratfor programs declare functions while Pascal ones do not.
.sp
.PP
To close, let me summarize the main points in the case against Pascal.
.IP 1.
Since the size of an array is part of its type,
it is not possible to write general-purpose routines,
that is,
to deal with arrays of different sizes.
In particular, string handling is very difficult.
.IP 2.
The lack of static variables, initialization
and a way to communicate non-hierarchically
combine to destroy the ``locality'' of a program \(em
variables require much more scope than they ought to.
.IP 3.
The one-pass nature of the language forces procedures and functions
to be presented in an unnatural order;
the enforced separation of various declarations scatters program
components that logically belong together.
.IP 4.
The lack of separate compilation impedes the development of large programs
and makes the use of libraries impossible.
.IP 5.
The order of logical expression evaluation cannot be controlled,
which leads to convoluted code and extraneous variables.
.IP 6.
The
.UL case
statement is emasculated because there is no default clause.
.IP 7.
The standard I/O is defective.
There is no sensible provision for dealing with files or program arguments
as part of the standard language,
and no extension mechanism.
.IP 8.
The language lacks most of the tools needed for assembling
large programs, most notably file inclusion.
.IP 9.
There is no escape.
.PP
This last point is perhaps the most important.
The language is inadequate but circumscribed,
because there is no way to escape its limitations.
There are no casts to disable the type-checking
when necessary.
There is no way to replace the defective run-time environment
with a sensible one,
unless one controls the compiler that defines the ``standard procedures.''
The language is closed.
.PP
People who use Pascal for serious programming
fall into a fatal trap.
Because the language is so impotent, it must be extended.
But each group extends Pascal in its own direction,
to make it look like whatever language they really want.
Extensions for separate compilation,
Fortran-like
.UC COMMON ,
string data types,
internal static variables,
initialization,
octal numbers,
bit operators,
etc.,
all add to the utility of the language
for one group,
but destroy its portability to others.
.PP
I feel that it is a mistake to use Pascal for anything much
beyond its original target.
In its pure form,
Pascal is a toy language,
suitable for teaching but not for real programming.
.SH
Acknowledgments
.PP
I am grateful to
Al Aho,
Al Feuer,
Narain Gehani,
Bob Martin,
Doug McIlroy,
Rob Pike,
Dennis Ritchie,
Chris Van Wyk
and
Charles Wetherell
for helpful criticisms
of earlier versions of this paper.
.[
$LIST$
.]