Programming Language Concepts: Lecture Notes

Programming Language Concepts:
Lecture Notes
Madhavan Mukund
Chennai Mathematical Institute
92 G N Chetty Road
Chennai 600 017
http://www.cmi.ac.in/madhavan
2
Contents
I Object-oriented programming 7
1 Data encapsulation with classes 9
1.1 Classes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.2 Data encapsulation (public and private) . . . . . . . . . . . . . . . . . . . . 12
1.3 Static components . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
1.4 Constants (the attribute final) . . . . . . . . . . . . . . . . . . . . . . . . . . 16
1.5 Constructors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2 Java Interlude 1
Quick introduction to Java 21
2.1 Structure of a Java program . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.2 Compiling and running Java programs . . . . . . . . . . . . . . . . . . . . . 22
2.3 Basic Java syntax . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.3.1 Scalar variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.3.2 Expressions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
2.3.3 Compound statements . . . . . . . . . . . . . . . . . . . . . . . . . . 24
2.3.4 Final note . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
3 Subclasses and inheritance 29

3.1 Subclasses and superclasses . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
3.2 Inheritance polymorphism . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
3.3 Multiple inheritance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
3.4 The Java class hierarchy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
3.5 Subtyping vs inheritance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
4 Abstract classes 37
4.1 Abstract Classes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
4.2 Generic functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
4.3 Multiple inheritance and interfaces . . . . . . . . . . . . . . . . . . . . . . . 40
5 Interfaces: Applications 43
5.1 Callback functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
5.2 Iterators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
3
5.3 Ports and interaction with state . . . . . . . . . . . . . . . . . . . . . . . . 51
6 Java Interlude 2
Parameter passing, cloning, packages 53
6.1 Parameter passing in Java . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
6.2 Cloning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
6.3 Packages . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
6.4 The protected modifier . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
II Exception handling 59
7 Exception handing 61
7.1 Javas try-catch blocks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
7.2 Cleaning up with finally . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
7.3 Customized exceptions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
7.4 A cute fact about finally . . . . . . . . . . . . . . . . . . . . . . . . . . 65
III Event driven programming 67

8 Event driven programming 69
8.1 Programming graphical user interfaces . . . . . . . . . . . . . . . . . . . . . 69
8.2 Some examples of event driven programming in Java . . . . . . . . . . . . . 71
8.2.1 A button . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
8.2.2 Three buttons . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
8.3 The event queue . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
8.4 Custom events . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
IV Reflection 81
9 Reflection 83
V Concurrent programming 89
10 Concurrent Programming 91
10.1 Race conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
10.2 Protocols based on shared variables . . . . . . . . . . . . . . . . . . . . . . . 94
10.3 Programming primitives for mutual exclusion . . . . . . . . . . . . . . . . . 96
10.4 Monitors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
10.5 Monitors in Java . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
10.6 Java threads . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
4
10.7 Interrupts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
VI Functional programming 109

11 Functional programming in Haskell 111
11.1 Types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
11.2 Defining functions in Haskell . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
11.3 Functions as first class objects . . . . . . . . . . . . . . . . . . . . . . . . . . 115
11.4 Polymorphism . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
11.5 Conditional polymorphism . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
11.6 User defined datatypes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
11.7 Rewriting as a model of computation . . . . . . . . . . . . . . . . . . . . . . 120
11.8 Reduction strategies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
11.9 Outermost reduction and infinite data structures . . . . . . . . . . . . . . . . 123
12 The (untyped) lambda calculus 125

12.1 Syntax . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
12.2 Life without types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
12.3 The rule . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
12.4 Variable capture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
12.5 Encoding arithmetic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128
12.6 One step reduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
12.7 Normal forms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
12.8 -equivalence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
12.9 Church-Rosser property . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
12.10Computability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
12.11Encoding recursive functions in lambda calculus . . . . . . . . . . . . . . . . 135
12.12Fixed points . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
12.13A fixed point combinator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138
12.14Making sense of terms without normal forms . . . . . . . . . . . . . . . . . . 139
13 Introducing types into the lambda calculus 141

13.1 Simply typed lambda calculus . . . . . . . . . . . . . . . . . . . . . . . . . . 141
13.2 Polymorphic typed calculi . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
13.3 Type inference as equation solving . . . . . . . . . . . . . . . . . . . . . . . 145
13.4 Unification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145
13.5 Type inference with shallow types . . . . . . . . . . . . . . . . . . . . . . . . 149
VII Logic programming 153

14 Introduction to logic programming 155
14.1 Facts and rules . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155
5
14.2 Computing in Prolog . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156
14.3 Complex structures in Prolog . . . . . . . . . . . . . . . . . . . . . . . . . . 157
14.4 Turning the question around . . . . . . . . . . . . . . . . . . . . . . . . . . . 158
14.5 Arithmetic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159
14.6 Negation as failure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 160
14.7 Cut . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161
VIII Scripting languages 163

15 Programming in Perl 165
15.1 Scalar datatypes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165
15.2 Arrays . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166
15.3 Control flow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167
15.4 Input/Output . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 168
15.5 Matching and regular expressions . . . . . . . . . . . . . . . . . . . . . . . . 169
15.6 Associative arrays . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171
15.7 Writing functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 172
15.8 Sorting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 172
IX Appendix 175
A Some examples of event-driven programming in Swing 177
A.1 Multicasting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 177
A.2 Checkbox . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 180
6
Part I
Object-oriented programming
7
Chapter 1
Data encapsulation with classes
The starting point of object-oriented programming is to provide a more faithful implemen-

tation of the notion of type in programming. The programming languages we are familiar
with come with standard built-in types that we can assign to variables and values:
For example, in C we can use int, float, double, char, . . . . In Haskell, similarly, we
can use Int, Float, Char, Bool, . . . .
In addition, programming languages provide a way to assign a single name to a collection
of values of the same type. The nature of this collection is determined by the underlying
architecture assumed by the programming language. Since C reflects the standard von
Neumann stored program architecture, we have arrays that provide random (i.e. equal
time) access to each element of the collection, while Haskell, whose semantics is defined with
respect to an idealized functional programming architecture, provides us with lists, which
have sequential access (getting to the ith item takes time proportional to i) but which
come with a number of useful decomposition functions that make inductive definitions easy
to implement.
However, no programming language can expect to predefine all the useful types that we
might need in a particular applicationwe might want to work with stacks or binary search
trees or directed acyclic graphs or . . . Typically, we define these in terms of the existing
primitives available. For instance, in Haskell we might represent a stack as a list or, for
efficiency, as a pair of lists, while in C we might represent a stack as an array or, if we want
to let the stack grow arbitrarily large, as a linked list.
Regardless of the implementation, we expect a stack to be used in a particular way.
Additions to the stack are done using a function push, deletions by a function pop and
the only enquiry that we can make about the state of the stack is the question is-the-stack-
empty. Suppose the stack is defined as
int s[100], top_of_stack = 0;
in C and the stack grows as s[0], s[1], .... When the stack has 10 elements (i.e.,
top_of_stack == 10) we do not want a function to access, say, s[7].
9
In a language like C, this integrity of the datatype stack is difficult to enforce. It can,
in general, be achieved only through disciplined programmingthere is no way to guarantee
that a stack is never misused.
The preceding discussion on stacks applies equally to other kinds of complex datatypes
like binary search trees, directed acyclic graphs etc. In the jargon of programming language
theory, these are all examples of Abstract Data Typesdata structures that have certain
fixed functions to manipulate them and these functions are expected to be rigidly adhered
to as the only means of accessing the underlying data, regardless of the implementation.
1.1 Classes
To facilitate the definition of abstract data types, the programming language Simula (1967)
introduced a concept called a class. (The word class is not important in itself, except that
it has stuck and so is now a standard term in the field.)
A class is a type definition consisting of two parts:
1. How the data is stored in this type.
2. What functions are available to manipulate this data.
For instance, we might represent a stack of integers by the following class definition:
class stack {
int values[100]; /* values are stored in an array */
int tos = 0; /* top of stack, initialize to 0 */
push (int i, ...){ /* push i onto stack */

values[tos] = i;
tos = tos+1; /* Should check that tos < 100!! */
}
int pop (...){ /* pop top of stack and return it */

tos = tos - 1; /* Should check that tos > 0!! */
return values[tos];
}
bool is_empty (...){ /* is the stack empty? */

return (tos == 0); /* yes iff tos is 0 */
}
}
At some places in this hypothetical definition, the function parameters have been left
unspecified and denoted . . . . We will fill in these blanks shortly.
We can use the class name as a new type in the language and define variables of this type
elsewhere in the program, as follows:
10
stack s,t;
These definitions only provide the capability for s and t to denote stacks. No storage is
allocated at this point. They correspond to a pointer declaration of the type int *p in C,
which gives p the ability to point to a memory location holding an integer but does not, in
itself, allocate a valid memory location for p to point to.
In order to get a real stack, we have to ask for it. For instance, we might say:
s = new stack;
t = new stack;
This generates two independent stacks, one named s and the other t. If, instead, we had
written
s = new stack;
t = s;
we get a single stack with two names, s and t, either of which can be used to access the
stack.
The class definition only provides a template for a datatype. The operation new generates
an instance of the template. The word object in object-oriented programming is synonymous
with the phrase instance of a class.
How do we manipulate the objects s and t? Suppose we want to push the value 7 onto
the stack s. We do this by writing:
s.push(7);
Note the difference in style with a conventional language. In C or Haskell, push would
take two arguments, the stack itself and the value to be pushed. In the object-oriented
framework, each instance of the class has its own copy of these functions and is thus an
implicit argument to the function.
Similarly, to retrieve the value at the top of s into a local variable and to query whether
t is empty, we would write the following, respectively:
i = s.top();
if (t.is_empty()) {...}
We can now say what should appear in place of the mysterious . . . in the function defini-
tions in our class stacknothing! In other words, the functions pop and is_empty take no
arguments at allthe only argument they need is the name of the stack on which to operate,
which is implicitly given since these functions will be called with respect to a fixed stack.
Similarly, push needs only one argument: the value to be pushed.
Is there any significant difference between a C style function push(s,i) which takes two
arguments, one of which is the stack, and the object-oriented version s.push(i), where the
11
stack is implicitly passed? Arguably this is only a matter of style and the two approaches
have the same effect.
But, philosophically, the syntax reflects the object-oriented approach: when we create an
object (an instance of a class) we are supposed to visualize the object as a black box with
a display which can show us some information about the box and some slots where we can
insert data. The box is equipped with buttons, one for each function. Pushing a button
invokes the corresponding function, with the input slots used to slip in parameters to the
function, and the value returned shown on the display. Our access to the internal state of
the object, whether to observe it or to modify it, is expected to be only via pushing buttons.
(Note: More jargonthe functions defined in classes are often called methods and the data
items are often called instance variables or instance fields.)
1.2 Data encapsulation (public and private)

What is the status of the instance variables defined within a class? It seems desirable that
these variables should not be accessible externally. For instance, there should be no way to
use the variables s and t of type stack to directly manipulate the contents of tos and the
array value within s and t. For, if these internal variables are accessible externally, we once
again have to deal with the problem of maintaining the integrity of our abstract data type!
One way to achieve total data integrity is to make all internal data variable inaccessible
from outside. Of course, there is almost always a requirement to be able to access the values
of some or all of the variables and also modify or update themfor example, consider the
following definition:
class date {
int day, month, year;
}
where one might reasonably expect to be able to find out what date is currently stored, as
well as, perhaps, to reset the date to a fresh value.
To meet this requirement while keeping internal variables inaccessible, we could write
functions like get_date() and set_date(...) that provide limited, controlled access to the
data stored in the class. (Note: In object-oriented jargon, a method that reads the values of
internal variables, like get_date(), is called an accessor method and a method that updates
the values of internal variables, like set_date(...) is called a mutator method.) Using
appropriate accessor and mutator functions, we can control very precisely the degree to
which the data in the stack can be manipulated externally.
However, probably to accommodate lazy programmers, most object-oriented languages
do permit direct external access to variables. The syntax is similar to that for accessing
members of structures in C (or fields of records in Pascal). For example, the field tos in the
stack s is designated s.tos, so we can write code like:
if (s.tos == 0){ ... }
12
Once we permit this facility, we have to introduce an additional mechanism to prevent
misusethat is, to retain data integrity. This is typically done by adding the qualifiers
public and private to each variable defined in a class. As should be clear, a variable that
is declared to be public is accessible from outside while one that is declared private is not.
One would write
class stack{
private int values[100];
private int tos = 0;
...
}
to ensure that the only way to access the stack is via push, pop and is_empty. Of course,
private only refers to the external world: the functions within the class like push, pop and
is_empty have full access to all the data variables defined in the class.
What about methodsshould all methods be accessible by default? Suppose our lan-
guage has a mechanism to request storage at run-time (like the malloc function in C). We
might then incorporate a method extend_stack into our class and use it when the stack
becomes full:
class stack {
...
push (int i){ /* push i onto stack */
if (stack_full){
extend_stack();
}
... /* Code to add i to stack * /
}
extend_stack(){
... /* Code to get additional space for stack data */
}
...
}
Clearly extend_stack is an implementation detail that is not intended to be revealed

outside the class. It would make sense to extend the use of the attributes public and private
to methods as well, making, in this case, only the methods push, pop and is_empty public
and all other methods, including extend_stack, private.
13
1.3 Static components
A C program is a collection of functions, all at the same level and hence all on an equal
footing. To unambiguously designate where the computation starts, we insist that there is
always a function called main where execution begins.
In a purely object-oriented language, a program is a collection of classes. Every function
and variable that is defined in the program must lie within some class. However, as we
have seen, classes are just templates and their definitions come to life only after they are
instantiated using new. Initially all classes are inactive and there is no way for the program
to get startedeffectively, the program is deadlocked.
So, there is a need to be able to define functions whose existence does not depend on a
class being instantiated.
There is another reason for having such functions. Consider, for example, routine func-
tions like read and write (or scanf and printf, in C terminology), or, for example, math-
ematical functions like sin, sqrt, . . . Clearly, it makes no sense to have to artificially instan-
tiate a class in order to use such functions.
One way to get around this problem without introducing free floating functions that live
outside classes (as happens, for instance, in C++) is to add the qualifier static. A function
that is marked static inside a class is available without having to instantiate the class. We
could have a definition of the form
class IO {
public static ... read(...) { ... }
public static ... write(...) { ... }
...
}
and then invoke these functions from outside as IO.read(...) and IO.write(...).
Static functions also solve the problem of where to startlike C, we can insist that the
collection of classes includes one static function with a fixed name (say, main, like C) and
begin execution by invoking this method.
In addition to static funtions, it also makes sense to have static data variables. This can
be used to define constants, such as:
class Math {
public static double PI = 3.1415927;
public static double E = 2.7182818;
public static double sin(double x) { ... }
...
}
Notice that the designation static is orthogonal to public/private. Does it make sense
to have something that is private static?
14
Though it appears useless at first sight, we could consider a class all of whose instances
needs some magic numbers to do their job, but where these magic numbers need not
be visible externally. Suppose we write a class that computes the annual interest yield on a
fixed deposit. Typically, the rate varies depending on the length of the deposit, but it might
be that all the variable rates are computed from a single base rate. We might then write
class interest-rate {
private static double base_rate = 7.32;
private double deposit-amount;
public double threemonth-yield(){ ... } /* uses base-rate

and deposit-amount */
public double sixmonth-yield(){ ... } /* uses base-rate
public double oneyear-yield(){ ... } /* uses base-rate */
...
}
The idea is that all instances of interest-rate share a single copy of the static variable
base_rate, so that there is no unnecessary duplication of data and also no inconsistency.
Static variables are shared across all instances of a class. Thus, we can also use static
variables to keep track of global information such as how many times a given function is
called across all instances of the class. For example, we could write
class stack {
private int values[100]; /* values are stored in an array */
private int tos = 0; /* top of stack, initialize to 0 */
private static int num_push = 0; /* number of pushes across all

stacks */
push (int i, ...){ /* push i onto stack */

values[tos] = i;
tos = tos+1; /* Should check that tos < 100!! */
num_push++; /* update static variable */
}
...
}
Here again, it makes sense to have num_push restricted to be private because we want it
to be incremented only by the method push within the class stack.
15
We have to be careful when we mix static and non-static entities in the same class. A
non-static function is tied to a specific instance (object) and so can implicitly refer to the
current state of the object as well as to static data (such as the yield functions above).
However, since a static function is shared amongst all instances of a class (and there need
not even be a single instance in existence for the static function to be used), such a function
should not make any reference to any non-static components of the class. We will address
this point in more detail later.
1.4 Constants (the attribute final)

Consider our earlier definition
class Math {
public static double PI = 3.1415927;
...
}
The intention was to make the value Math.PI available everywhere to use in expressions:
e.g.,
area = Math.PI * radius * radius;
However, there is nothing to prevent the following improper update of the public static
field PI:
Math.PI = 3.25;
To forbid this, we attach another attribute, indicating that a value but may not be
altered. Following Java terminology, we use the notation final, so we would modify the
definition above to say:
class Math {
public static final double PI = 3.1415927;
...
}
and achieve what we set out to havea publicly available constant that exists without
instantiating any objects.
Actually, the modifier final has other ramifications. We will see these later, when we
talk about inheritance.
16
1.5 Constructors
When we create an object using new, we get an uninitialized object. We then have to set
the values of the instance variables to sensible values (almost always this has to be done via
appropriate methods, because these variables will be private). It is natural to ask for the
analogue of a declaration of the form:
int i = 10;
which creates a variable i of type int and also, simultaneously, sets its initial state.
In the object-oriented paradigm, this initialization is performed by special methods, called
constructors. A constructor is a method that is called implicitly when the object is created.
It cannot be called after the object has been created.
In Java, a constructor is written as a public method with no return value which takes
the same name as the class. For example, we can define a class Date as follows:
class Date{
private int day, month, year;
public Date(int d, int m, int y){

day = d;
month = m;
year = y;
}
Now, if we write
Date d = new Date(27,1,2003);
d points to a new Date object whose values are initialized with day = 27, month = 1 and
year = 2003.
We may want to have more than one way to construct an object. If no year is supplied,
we might set the field year to the current year, by default.
We can add a second constructor as follows:
public Date(int d, int m)

day = d;
month = m;
year = 2003;
}
Now, if we write
17
Date d1 = new Date(27,1,2002);
Date d2 = new Date(27,1);
d1 calls the first constructor, as before, while d2 calls the second constructor (where year is
set to 2003 by default).
This ability to have multiple constructors with different signatures extends to methods
as well. In Java, the signature of a method is its name plus the types of its arguments. One
can overload method names just like we have overloaded the constructors above. In fact, in
the Java built-in class Arrays, there is a static method sort that sorts arbitrary arrays of
scalars. In other words, we can write:
double[] darr = new double[100];

int[] iarr = new int[500];
...
Arrays.sort(darr); // sorts contents of darr
Arrays.sort(iarr); // sorts contents of iarr
This is done by defining, within the class Arrays, multiple methods named sort with
different signatures, as follows:
class Arrays{
...
public static void sort(double[] a){ // sorts arrays of double[]
...
}
public static void sort(int[] a){ // sorts arrays of int[]
...
}
...
}
When we invoke Arrays.sort with an argument a, the type of a automatically deter-

mines which version of the overloaded method sort is used.
Coming back to constructors, what happens if we have no constructors defined? In this
case, Java provides a default constructor that takes no arguments and sets all instance
variables to some sensible defaults (e.g., an int is set to 0). Suppose we have a class as
follows:
class no_constructor{
private int i;
// some methods below

...
}
18
We would then write something like:
no_constructor n = new no_constructor(); // Note the () after

// the class name
However, if there is at least one constructor defined in the class, the default constructor
is withdrawn. So, if we have the class Date as given above, it is an error to write:
Date d = new Date();
If we want this to work, we must explicitly add a new constructor that has no arguments.
Remember that it is not possible to invoke a constructor later. Though Date(int,int,int)
is a public method in the class, it has a different interpretation. We cannot say;
Date d = new Date(27,1,2003);

...
d.Date(27,1,2003);
One constructor can call another, using the word this. We can rewrite the two construc-
tors in the class Date as follows:
class Date{

day = d;
month = m;
year = y;
}
public Date(int d, int m){

this(d,m,2003);
}
The second constructor invokes the first one by supplying a fixed third argument. In
Java, such an invocation using this must be the first statement in the constructor. We can
reverse the constructors as follows:
class Date{
19
public Date(int d, int m){
day = d;
month = m;
year = 2003;
}

this(d,m); // this sets year to 2003
year = y; // reset year to the value supplied
}
20
Chapter 2
Java Interlude 1
Quick introduction to Java
2.1 Structure of a Java program

A Java program is a collection of classes. Each class is normally written in a separate file
and the name of the file is the name of the class contained in the file, with the extension
.java. Thus, the class stack defined earlier would be stored in a file called stack.java.
As we noted in Chapter 1, we have to fix where the program starts to execute. Java
requires that at least one of the classes has a public static method called main. More
precisely we need a method with the following signature:
public static void main(String[] args)
We have already seen what public and static mean. The word void has the same
connotation as in C this function does not return a value. The argument to main is an
array of strings (well look at strings and arrays in Java in more detail later)this corresponds
to the argv argument to the main function in a C program. We dont need argc because in
Java, unlike in C, it is possible to extract the length of an array passed to a function. Also,
unlike C, the argument to main is compulsory. Of course, the name of the argument is not
important: in place of args we could have used xyz or any other valid identifier.
Here then, is a Java equivalent of the canonical Hello world program.
class helloworld{
public static void main(String[] args){
System.out.println("Hello world!");
}
}
Here println is a static method in the class System.out which takes a string as argument
and prints it out with a trailing newline character.
As noted earlier, we have to save this class in a file with the same name and the extension
.java; that is, helloworld.java.
21
2.2 Compiling and running Java programs
The java compiler is invoked using javac. Though our Java program may consist of several
classes, it is sufficient to invoke javac on the class containing the static method main where
we want the execution to begin. In the simple example above, we would write
javac helloworld.java
Note that we have to supply the full filename, with the extension .java, to javac.
The Java compiler will automatically compile all additional classes that are invoked
(transitively) by the current class being compiled. Each compiled class is stored in a file
with the extension .java replaced by .class. In our simple example, we get a single new
file called helloworld.class.
To get the full effect, try splitting the example into two classes (in two separate files), as
follows, and see what happens when you compile helloworld.java:
class helloworld{
auxclass.printit("Hello world!");
}
}
class auxclass{
public static void printit(String s){
System.out.println(s);
}
}
The Java compiler does not produce executable machine code for the underlying archi-
tecture, unlike, say, gcc. Instead, Java programs execute on a uniform underlying virtual
machine called the Java Virtual Machine (JVM). The compiler produces JVM machine code,
also called JVM bytecode.
To run the program, we need an interpreterthat is, an implementation of the JVM.
This is invoked using the command java. We write
java helloworld
to load the compiled file helloworld.class into the JVM interpreter and execute it. (Note
that we should not give the extension .class in this case, unlike when we invoked the
compiler javac!)
The fact that Java programs run under a uniform artificial environment (the JVM)
on all architectures and operating systems makes it possible for Java to achieve a higher
degree of standardization than most other programming languagesfor instance, there is
no ambiguity about the size of an int, unlike C. Moreover, since the JVM is in complete
22
control, more exhaustive error reporting and handling is possible for runtime errors such as
array bound overflows. Finally, dynamic storage management is significantly simplified by
the incorporation of an automatic garbage collector in the JVM to reclaim freed memory,
removing the peril of memory leaks. Of course, there is a performance penalty, but that
is another story.
2.3 Basic Java syntax

2.3.1 Scalar variables
The basic variable types are similar to C. The basic types are:
int (4 bytes), long (8 bytes), short (2 bytes)

float (4 bytes), double (8 bytes)
char (2 bytes)
boolean
Some points to note.
1. The size of each variable is unambiguously fixed (since all Java programs run on a
uniform Java Virtual Machine, independent of the local architecture).
2. The char type is 2 bytes, not 1 byte. Java uses an encoding called Unicode which
allows all international character sets to be written as char. Unicode is a conservative
extension of ASCII, so if the first byte is 0, the second byte corresponds to the normal
ASCII code. We need not worry about Unicode. As far as we are concerned, what is
important is that character constants are enclosed in single quotes, as usuale.g.,
char c;
c = a;
c = X;
if (c != }) {...}
3. There is an explicit boolean type, unlike C. The two constants of this type are written
true and false. We can write code like:
boolean b, c = false;
b = true;
b = (i == 7);
Another point is that a typo like if (x = 7) ... will be caught as syntactically

illegal since x = 7 returns 7, not the boolean value true.
23
2.3.2 Expressions
Expressions are essentially the same as C, including boolean expressions. As in C, an assign-
ment of the form x = expr returns a value, namely the final value of the expression expr
that is assigned to the variable x. Hence, like C, for checking equality, use == not =, but (as
noted above), stupidities like if (x = 7).... will be caught by the compiler.
2.3.3 Compound statements

Conditional execution
if (condition) ... else ... , same as in C
Loops
while (condition) { ... }
do { ... } while (condition)
for (i = 0; i < n; i++) { ... }
Goto
There is no goto in Java. However, the language has labelled break statements that allow
several levels of nested loops to be terminated in one go. For example:
outer_loop: // this is a loop label

for (i = 0; i < n; i++){
for (j = 0; j < n; j++){
if (a[i][j] == k){
break outer_loop; // exits the outer for loop
}
}
}
Comments
Comments can be C-style, enclosed in /* ... */ (but remember that these are not allowed
to be nested, as in C).
Another possibility is a single line comment starting with // the rest of the line
after // is treated as a comment and ignored. See examples above in the labelled break
explanation.
Strings
Strings in Java are handled using the built-in class String. String variables are declared as
follows:
24
String s,t;
String constants are enclosed in double quotes, as usual. Thus, one can write:
String s = "Hello", t = "world";
Strings are not arrays of characters, so one cannot refer to s[i], t[j] etc. To change
the contents of string s from "Hello" to "Help!" we cannot just say:
s[3] = p; s[4] = !;
What we can do is invoke the method substring in the String class and say something
like:
s = s.substring(0,3) + "p!";
Here, s.substring(0,3) returns the substring of length 3 starting at position 0. The

operator + is overloaded to mean concatenation for strings.
String objects are immutableif a string changes, you get a new String object. Even
though, in principle, the new value of s fits in the same amount of space as the old one, Java
creates a new object with the contents "Help!" and s points to this new object henceforth.
What happens to the old object? It is automatically returned to the system: Java does
runtime garbage collection so we do not have to explicitly return dynamically allocated
storage using something like free(..), unlike in C.
Remember that an expression like
s = "A new piece of text";
is an abbreviation for something like
s = new String("A new piece of text");
The string passed as a parameter to new is converted into the appropriate internal rep-
resentation by the constructor for String class. The point that needs to be noted is that
though String is a built in class like any other and has its own methods etc, there are some
short cuts built-in to Java, such as:
Creating a new String object implicitly using an assignment statement.

Overloading + to mean concatenation, so we can have expressions of type String.
The length of a string s can be obtained by calling the method s.length(). Note that
we dont know (and dont care) how strings are actually represented in the String classfor
instance, we dont have the C headache of remembering that we always need to put a \0
at the end of a string to make it compatible with functions that operate on strings.
For more information on what methods are available in String, look at the documenta-
tion for Java, available on the system.
25
Arrays
Like strings, arrays are objects in Java. A typical declaration looks like this:
int[] a;
a = new int[100];
This can also be done as follows (a bit more C like);
int a[];
a = new int[100];
However, arguably the first definition is clearer in that it clearly declares a to be of type
int[] and not int.
These two lines can be combined as:
int[] a = new int[100];
Observe however, that declaring an array is fundamentally a two step process you
declare that a is an array (without specifying its size) and then create the array with a fixed
size at run time. Of course, you can abandon the current array and create a new one with
a fresh call to new, as follows. (This example also shows that you can genuinely assign the
size of the array at run time.)
public class arraycheck{
public static void main(String[] argv){
int[] a;
int i;
int n;
n = 10;
a = new int[n];
for (i = 0; i < n; i++){

a[i] = i;
}
n = 20;
a = new int[n];
26
for (i = 0; i < n; i++){
a[i] = -i;
}
for (i = 0; i < n; i++){

System.out.print(i);
System.out.print(":");
System.out.println(a[i]);
}
}
}
A very useful feature of arrays in Java is that the length of an array is always available:
the length of an array a is given by a.length. A method such as a sort program that
manipulates an array does not require an additional argument stating the length of the
array to be manipulated, unlike in C.
2.3.4 Final note

Now we know exactly what type the argument to main() is in the following:
class helloworld{
System.out.println("Hello world!");
}
}
Here, args is an array of String. We dont need the equivalent of argc in C because we
can extract it via args.length.
27
28
Chapter 3
Subclasses and inheritance
3.1 Subclasses and superclasses

Consider the following definition:
class Employee
{
private String name;
private double salary;
// Constructors
public Employee(String n, double s){
name = n; salary = s;
}
public Employee(String n){

this(n,500.00);
}
// "mutator" methods
public boolean setName(String s){ ... }
public boolean setSalary(double x){ ... }
// "accessor" methods
public String getName(){ ... }
public double getSalary(){ ... }
// other methods
29
double bonus(float percent){
return (percent/100.0)*salary;
}
}
Suppose we want to define a class of Managers, who are like Employees except that they
have, in addition, a secretary. We can re-use the Employee class as follows:
class Manager extends Employee

{
private String secretary;
public boolean setSecretary(name s){ ... }
public String getSecretary(){ ... }
Objects of type Manager implicitly inherit instance variables and methods from Employee.
Thus, every Manager object has a name and salary in addition to a secretary.
In OO terminology, Manager is referred to as the subclass while Employee is referred to
as the superclass. In general, a subclass specializes the superclass. The terminology makes
sense set-theoretically: every Manager is also an Employee so the set of Managers is a subset
of the set of Employees.
The class Manager cannot see the private fields of Employee! This is to prevent the
data encapsulation from being brokenoften the programmer who creates a subclass is
not the one who wrote the parent class. A constructor for Manager must call Employee
constructor to set name and salary. In a hierarchy of classes, this transfers the responsibility
for construction of higher levels to the parent class, which inductively knows what to do.
We use the word super to call the parent constructor. Thus, we would write:
// Constructors
public Manager(String n, double s, String sn){
super(n,s); // "super" calls the Employee constructor
secretary = sn;
}
Manager can override a method from Employee. For instance, Manager may redefine
bonus as followsnote the use of super to unambiguously use the superclasss bonus in this
definition.
double bonus(float percent){

return 1.5*super.bonus(percent);
}
30
In general, the subclass has greater functionality (more fields and methods) than the
superclass. So, casting up the hierarchy is permitted. For example,
Employee e = new Manager ( ....)
A Manager object can do everything that an Employee object can. The converse is not
possible. What would the reference to getSecretary() mean in the following code?
Manager m = new Employee(...);
System.out.println(m.getSecretary());
3.2 Inheritance polymorphism

The static type of the variable determines what methods are accessible. If we write
Employee e = new Manager ( ....)
we are not permitted to write
System.out.println(e.getSecretary());
even though, at run time, the call to getSecretary() is not a problem.
But, we still have to answer the following. What does
e.bonus(p)
mean?
There are two possibilities:
1. Static: e is an Employee, so take the definition of bonus from class Employee.
2. Dynamic e is declared to be an Employee but is actually a Manager at run time, so
take the definition of bonus from Manager.
The dynamic definition is more useful than the static oneit permits the following type
of generic program.
Employee[] emparray = new Employee[2];
Employee e = new Employee(...);
Manager e = new Manager(...);
emparray[0] = e;
emparray[1] = m;
for (i = 0; i < emparray.length; i++){

System.out.println(emparray[i].bonus(5.0);
}
31
Here, emparray[1] will correctly use bonus as defined in Manager.
This ability to choose the appropriate method at run time, based on the identity of the
object, is called run time polymorphism or inheritance polymorphism.
This is in contrast to overloading based polymorphism such as the multiplicity of sort
methods in the class Arrays, described earlier.
Getting back to the code above, is emparray[1].setSecretary(s) allowed? After all,
we know that it is a Manager, even though it is declared Employee.
As mentioned earlier, if a variable is declared to be of type Employee, only methods that
are accessible in the class Employee may be used.
However, if we are confident that at run time the object is really a Manager, we may ask
for a type upgrade, or a cast. Casting is done using C-like syntax.
((Manager) emparray[1]).setSecretary(s)
At run time, if emparray[1] is not a Manager, the cast will fail (rather than call an
absurd method) and the program will crash. Casting can only be done between compatible
typesthat is, from a class to one of its descendants.
We can check whether emparray[1] is a Manager at run time using the predicate instanceof.
To be safe, we can write:
if (emparray[1] instanceof Manager){

((Manager) emparray[1]).setSecretary(s);
}
This is a trivial example of a property called reflection. Reflection is used in the philo-
sophical sense to refer to a language that can think about itself. Java is rich in reflective
capabilities. We will discuss this again later.
3.3 Multiple inheritance

Is it possible for a class to extend more than one class. For instance, could we have the
following?
class C1{
...
}
class C2{
...
}
class C3 extends C1, C2{

...
}
32
The problem is that what C3 inherits from C1 might conflict with what it inherits from
C2. Suppose we have a method f(..) in both C1 and C2 as follows:
class C1{
...
public int f(int i){ ... }
..
}
class C2{
...
public int f(int i){ ... }
..
}
class C3 extends C1, C2{

... // no definition of "public int f(int i)" in C3
}
Now, if we create an object of type C3 and invoke f(..), it is not clear whether we should
use the definition of f(..) from C1 or C2.
One possibility is to say that C3 can extend C1 and C2 provided it explicitly redefines all
conflicting methods. Another is to say that C3 can extend C1 and C2 provided there are no
conflicting methods. The latter option is available in C++.
3.4 The Java class hierarchy

Java rules out multiple inheritance. In Java, the class hierarchy is tree like.
In fact, not only is the hierarchy tree-like, Java provides a universal superclass called
Object that is defined to be the root of the entire class hierarchy. Every class that is defined
in a Java program implicitly extends the class Object.
The class Object defines some useful methods that are automatically inherited by all
classes. These include the following:
boolean equals(Object o) // defaults to pointer equality
String toString() // converts the values of the

// instance variable to String
The built-in method equals checks whether two variables actually point to the same
physical object in memory. This is rather strong and most classes will redefine this to say
that two objects are the same if their instance variables are the same.
33
The toString() method is useful because Java implicitly invokes it if an object is used as
part of a string concatenation expression. To print the contents of an object o, it is sufficient
to write:
System.out.println(o+"");
Here, o+"" is implicitly converted to o.toString()+"".

Recall that a variable of a higher type can be assigned an object of a lower type (e.g.,
Employee e = new Manager(...)) This compatibility also holds for arguments to methods.
A method that expects an Employee will also accept a Manager.
We can write some generic programs, as follows:
public int find (Object[] objarr, Object o){

int i;
for (i = 0; i < objarr.length(); i++){
if (objarr[i] == o) {return i};
}
return (-1);
}
As mentioned above, we will probably redefine methods like equals to be less restrictive
than just referring to pointer equality. However, we should be careful of the types we use.
Suppose we enhance our class Date to say:
boolean equals(Date d){

return ((this.day == d.day) &&
(this.month == d.month) &&
(this.year == d.year));
}
Before proceeding, we note a couple of facts about this method. Recall that day, month
and year are private fields of class Date. Nevertheless, we are able to refer to these fields
explictly for the incoming Date object d. This is a general featureany Date object can
directly read the private variables of any other Date object. Without this facility, it would
be tedious, or even impossible, to write such an equals(..) method.
Second, note the use of this to unambiguously refer to the object within which equals
is being invoked. The word this can be omitted. If we write, instead,
boolean equals(Date d){

return ((day == d.day) &&
(month == d.month) &&
(year == d.year));
}
34
then, automatically, the unqualified fields day, month and year refer to the fields in the
current object. However, we often use this to remove ambiguity.
Now, what happens when we call our generic find with find(datearr,d), where datearr
and d are declared as Date[] datearr and Date d?
When we call datearr[i].equals(d) in the generic find, we search for a method with
signature
boolean equals(Object o)
because, within find, the parameter to find is an Object. This does not match the revised
method equals in Date, so we end up testing pointer equality after all!
To fix this, we have to rewrite the equals method in Date as:
boolean equals(Object d){
if (d instanceof Date){
return ((this.day == d.day) &&
(this.month == d.month) &&
(this.year == d.year));
}
return(false);
}
Notice that this is another use of the predicate instanceof to examine the actual class
of d at runtime.
In general, the method used is the closest match. If Employee defines
boolean equals(Employee e)
and Manager extends Employee but does not redefine equals, then in the following code
Manager m1 = new Manager(...);
Manager m2 = new Manager(...);
...
if (m1.equals(m2)){ ... }
the call m1.equals(m2) refers to the definition in Employee because the call to a non existent
method
boolean equals(Manager m)
is compatible with both
boolean equals(Employee e) in Employee
and
boolean equals(Object o) in Object.
However, the definition in Employee is a closer match than the one in Object.
35
3.5 Subtyping vs inheritance
In the object-oriented framework, inheritance is usually presented as a feature that goes
hand in hand with subtyping when one organizes abstract datatypes in a hierarchy of classes.
However, the two are orthogonal ideas.
Subtyping refers to compatibility of interfaces. A type B is a subtype of A if every

function that can be invoked on an object of type A can also be invoked on an object
of type B.
Inheritance refers to reuse of implementations. A type B inherits from another type A

if some functions for B are written in terms of functions of A.
In the object-oriented framework, if B is a subclass of A then it is clear that B is a subtype

of A. Very often, B also inherits from Afor example, recall the way the implementation of
bonus() for Manager called on the function bonus() defined in Employee.
However, subtyping and inheritance need not go hand in hand. Consider the data struc-
ture deque, a double-ended queue. A deque supports insertion and deletion at both ends,
so it has four functions insert-front, delete-front, insert-rear and delete-rear. If
we use just insert-rear and delete-front we get a normal queue. On the other hand,
if we use just insert-front and delete-front, we get a stack. In other words, we can
implement queues and stacks in terms of deques, so as datatypes, Stack and Queue inherit
from Deque. On the other hand, neither Stack not Queue are subtypes of Deque since they
do not support all the functions provided by Deque. In fact, in this case, Deque is a subtype
of both Stack and Queue!
In general, therefore, subtyping and inheritance are orthogonal concepts. Since inheri-
tance involves reuse of implementations, we could have an inheritance relationship between
classes that are incomparable in the subtype relationship. In a more accurate sense, this is
what holds between Stack, Queue and Deque. The type Stack supports functions push and
pop that are implemented in terms of the Deque functions insert-front and delete-front.
Similarly, Queue supporst functions insert and delete that are implemented in terms of
the Deque functions insert-rear and delete-front. The function names are different, so
none of these types is comparable in the subtype relation, but Stack and Queue do inherit
from Deque.
36
Chapter 4
Abstract classes
4.1 Abstract Classes

We may use a class only as a collective name for a group of related classes. For instance,
we might have classes such as
class Circle{
private double radius;
...
public double perimeter(){
...
}
...
}
class Square{
private double side;
...
...
}
...
}
class Rectangle{
private double length;
private double width;
...
...
}
37
...
}
and combine these all under a class Shape so that
class Circle extends Shape{...}

class Square extends Shape{...}
class Rectangle extends Shape{...}
We dont actually intend to create objects of the parent type Shape, but we can use
Shape to enforce some common properties of the classes that extend Shape. For example,
we might want to insist that each subclass of Shape define a method with the signature
public double perimeter();
Though this method can be defined to yield a sensible value in each subclass of Shape,
there is no reasonable definition that we can provide for the quantity perimeter() in an
abstract Shape. We could, of course, create a dummy function in Shape, such as
public double perimeter() { return -1.0; }
that returns an absurd value.

This way, each subclass of Shape will definitely be able to access a method of the required
signature, but we want more: we want to insist that the subclass redefines this method to a
sensible value.
The way to do this is to specify the signature of perimeter() in Shape, but say it is
only a template that has to be implemented in any subclass. We declare a method to be a
template by using the word abstract as follows:
public abstract double perimeter();
If a class contains an abstract method, it makes no sense to create an instance of that

class since it is not fully specified. Such a class is itself called abstract. Thus, we have:
abstract class Shape{

...
...
}
If a class has any abstract methods, it must be abstract, but the converse is not true. An
abstract class may have concrete instance variables and methods. For instance, we might
allow each Shape to have an unique identifier and define this field in common as:
38
abstract class Shape{
private String identifier;
public Shape(String s){ // Constructor to set private variable

identifier = s; // Call with "super(s)" from subclass
}
...
...
}
In fact, technically a class may be declared abstract even though it contains no abstract
components!
Let us call a class concrete if it is not abstract: A concrete class is one that we can instan-
tiate as an object. In order for a subclass to be concrete, it must provide an implementation
for all abstract methods that it inherits from above. Otherwise, the subclass also becomes
abstract.
Though we cannot create instances (objects) of an abstract class, we can declare variables
of this class. We can continue to use an array of the higher class to hold a mixed collection
of objects from the lower level, as in the example below:
Shape sarr[] = new Shape[3];
Circle c = new Circle(...); sarr[0] = c;

Square s = new Square(...); sarr[1] = s;
Rectangle r = new Rectangle(...); sarr[2] = r;
for (i = 0; i < 2; i++){

size = sarr[i].perimeter(); // calls the appropriate method
...
}
4.2 Generic functions

We can also use abstract classes to group together classes that are at different, incomparable
points in the class hierarchy but share some common properties. For example, we may want
to group together all classes whose objects can be compared using a linear order. Any such
class can then serve as a valid input to a sort function, using the comparison method defined
within the class whenever the sort function needs to compare two objects.
For instance, we might have a definition such as
abstract class Comparable{
39
public abstract int cmp(Comparable s);
// return -1 if this < s, 0 if this == 0, +1 if this > s
}
Now, we can sort any array of objects that extend the class Comparable:
class Sortfunctions{
public static void quicksort(Comparable[] a){
...
// Usual code for quicksort, except that
// to compare a[i] and a[j] we use a[i].cmp(a[j])
// (Why is this method declared to be static?)
}
...
}
If we want to make the objects of a class sortable using one of the Sortfunctions, we
write
class Myclass extends Comparable{

double size; // some quantity that can be used to compare
// two objects of this class
...
public int cmp(Comparable s){
if (s instanceof Myclass){
// compare this.size and ((Myclass) s).size
// Note the cast to access s.size
...
}
}
}
4.3 Multiple inheritance and interfaces

What if we want to make a class such as Circle (defined earlier) extend Comparable?
Circle already extends Shape and Java does not permit multiple inheritance.
Notice, however, that Comparable is a special kind of abstract class, one that has no
concrete components. In Java, this kind of abstract class can be renamed an interfacethe
word interface is derived from its English meaning and refers to the visible boundary of a
class that implements all the methods specified in the interface.
We can redefine Comparable as follows:
40
interface Comparable{
public abstract int cmp(Comparable s);
// return -1 if this < s, 0 if this == 0, +1 if this > s
}
A class that extends an interface is said to implement it:
class Myclass implements Comparable{

...
public int cmp(Comparable s) { ... }
...
}
In Java, a class may extend only one other class, but may implement any number of
interfaces. Thus, one can have
class Circle extends Shape implements Comparable{

...
public double perimeter(){...}
...
public int cmp(Comparable s){...}
...
}
As we mentioned earlier, extending multiple classes might create a conflict between con-
crete definitions in the multiple parent classes. However, since interfaces have no concrete
components, no such contradiction can occur if a class implements multiple interfaces. At
most, a class may implement two interfaces that both define abstract methods with the
same signature. Since the signature does not indicate anything about the meaning of
the method, any implementation of the method with that signature will be automatically
consistent with both its abstract definitions.
41
42
Chapter 5
Interfaces: Applications
As we have seen, interfaces can be used to tie together disparate classes that share some
common features. For instance, all classes that implement the interface Comparable describe
objects that can be linearly ordered by some internal parameter and hence sorted.
Notice that the property of being Comparable is orthogonal to any other features that
the class may possess. Functions that operate on objects that implement the Comparable
interface, such as the generic sorting functions we described, are not interested in these other
features. Indeed, one can make a stronger statement and say that these functions can only
access the Comparable portion of the objects that are passed to them.
In a more general sense, interfaces provide a limited public view of the overall capabilities
of an object. We look at two examples to illustrate this idea.
5.1 Callback functions

As we shall see later, Java permits parallel execution. One function can invoke another
function to run in a parallel thread. Unlike a conventional function call, where the calling
function is suspended till the called function terminates and returns, in this case the calling
function proceeds to the next statement after the function invocation, while the invoked
function executes in parallel. This parallelism may be actual, such as on a multiprocessor
machine, where each thread is assigned to a separate processor, or virtual, such as on a single
processor, where the threads are interleaved artibrarily to simulate parallel execution.
Suppose now that we have two classes, Myclass and Timer, where an object m of Myclass
can create a Timer object t and invoke a function t.f() to run in parallel with m. Recall that
we have assumed that m is not suspended after invoking t.f(). We would like a mechanism
for t.f() to report back to m when it has completed its work. To do this, t needs to have
a reference back to m. This can be accomplished by passing the reference when we create
t: each instance of Timer is created by an object and the instance of Timer remembers the
identity of the object that created it. For instance, we might have:
class Myclass{
...
43
public void some_function(){
...
Timer t = new Timer(this); // Tell t that this object
// created t
...
t.f(); // Start off t.f() in parallel
...
}
...
}
We need the following complementary structure in Timer:
class Timer implements Runnable{ // Says that Timer can be invoked

// in a parallel thread
private Object owner;
public Timer(Object o){ // Constructor

owner = o; // Remember who created me
}
...
public void f(){
...
o.notify(); // Tell my creator that Im done
}
...
}
The problem with this definition is that the function o.notify() cannot be invoked as
stated without casting o (which is an Object) back to Myclass. Of course, we could fix this
by changing the type of owner from Object to Myclass, but then we would have to write a
different Timer class for each type of owner.
In order to define a generic Timer class, all we need is that the owner have a method
called notify() and that there is a uniform way to cast the owner to a type that can access
this method. This can be achieved via an interface, say
interface Timerowner{
public abstract void notify();
}
Now, we modify Myclass to implement Timerowner:
class Myclass implements Timerowner{

...
44
public void some_function(){
...
Timer t = new Timer(this); // Tell t that this object
// created t
...
t.f(); // Start off t.f() in parallel
...
}
public void notify(){ ... } // Implement the interface

...
}
Finally, we modify Timer to insist that the owner of a Timer object must implement
Timerowner:
class Timer implements Runnable{ // Can be invoked in a

// parallel thread
private Timerowner owner; // Owner must implement Timerowner
public Timer(Timerowner o){ // Constructor

owner = o; // Remember who created me
}
...
public void f(){
...
o.notify(); // Tell my creator that Im done
}
...
}
Even though the timer t gets a reference to its parent object, the parent only exposes
its Timerowner features to t.
5.2 Iterators
A linear list is a generic way of storing Objects of any type. We might define a class to
store generic linear lists:
class Linearlist {
// Implement list as an array with arbitrary upper bound limit
private int limit = 100;
private Object[] data = new Object[limit];
45
private int size; // Current size of the list
public Linearlist(){ // Constructor

size = 0;
}
public void append(Object o){

data[size] = o;
size++;
if (size > data.length){ // Get more space!
Object[] tmp = data; // Save a pointer to current array
limit *= 2;
data = new Object[limit]; // New array of double the size
for (i = 0; i < tmp.length; i++){
data[i] = tmp[i];
}
}
...
}
Alternately, we might modify the internal implementation to use a linked list rather than
an array.
class Node{
private Object data;
private Node next;
public Node(Object o){

data = o;
next = null;
}
}
class Linearlist {
private Node head; // Implement list as a linked list

private int size;
public Linearlist(){ // Constructor

size = 0;
}
public void append(Object o){
46
Node m;
// Locate last node in list and append a new node containing o

for (m = head; m.next != null; m = m.next){}
Node n = new Node(o);
m.next = n;
size++;
}
...
}
Now, what if we want to externally run through a Linearlist from beginning to end?
If we had access to the array data we could write
int i;
for (i = 0; i < data.length; i++){
... // do something with data[i]
}
Alternatively, if we had access to head, we could write
Node m;
for (m = head; m != null; m = m.next}
... // do something with m.data
}
However, we have access to neither. In fact, we do not even know which of the two
implementations is in use! The Linearlist class should add functionality to support such
a list traversal. Abstractly we need a way to do the following:
Start at the beginning of the list;

while (there is a next element){
get the next element;
do something with it
}
We can encapsulate this functionality in an interface called Iterator:
public interface Iterator{

public abstract boolean has_next();
public abstract Object get_next();
}
47
Now, we need to implement Iterator in Linearlist. If we make this a part of the main
class, there will be a single pointer stored within the class corresponding to the temporary
variable i (or m) in the loop we wrote earlier to traverse the list. This means that we will
not be able to simulate a nested loop such as the following:
for (i = 0; i < data.length; i++){

for (j = 0; j < data.length; j++){
... // do something with data[i] and data[j]
}
}
because when we execute the equivalent of the inner loop, we lose the position that we were
in in the outer loop.
A better solution is to allow Linearlist to create, on the fly, a new object that imple-
ments Iterator and return it outside. This object will be an instance of a nested class,
defined within Linearlist. It will have access to private information about the list, even
though it is invoked outside! Here is how we do it.
class Linearlist {
// Implement list as an array with arbitrary upper bound limit
private int limit = 100;
private Object[] data = new Object[limit];
private int size; // Current size of the list
public Linearlist(){..} // Constructor
public void append(Object o){...}
// A private class to implement Iterator

private class Iter implements Iterator{
private int position;
public Iter(){ // Constructor

position = 0; // When an Iterator is created, it points
// to the beginning of the list
}
public boolean has_next(){

return (position < data.length - 1);
}
public Object get_next(){

Object o = data[position];
48
position++;
return o;
}
}
// Export a fresh iterator

public Iterator get_iterator(){
Iter it = new Iter();
return it;
}
...
}
Now, we can traverse the list externally as follows:
Linearlist l = new Linearlist();

...
Object o;
Iterator i = l.get_iterator()
while (i.has_next()){
o = i.get_next();
... // do something with o
}
...
What if Linearlist is implemented as a linked list rather than an array? We just change
the definition of the class Iter within Linearlist.

private Node position;

// Create a dummy node that points to the head of the list
position = new Node(null);
position.next = head;
}

return (position.next != null);
}
49
position = position.next;
Object o = position.data;
return o;
}
}
Externally, the loop that we wrote with get_iterator(), has_next() and get_next()
works as before.
Once again, we have passed a reference to an object but restricted access to the object
using an interface. The Iter object created within Linearlist can (and must) access
private variables within Linearlist, but the reference that is passed to this internal object
is constrained by the Iterator interface, so this capability cannot be exploited externally.
In fact, we can go one step further and even nest the linked list class as a local class within
Linearlist to prevent any accidental misuse. A complete implementation of Linearlist as
a linked list might be as follows.
class Linearlist {
private Node head; // Implement list as a linked list

private int size;
public Linearlist(){...} //Constructor
public void append(Object o){...}
private class Node{

private Object data;
private Node next;
public Node(Object o){

data = o;
next = null;
}
}

private Node position;

// Create a dummy node that points to the head of the list
position = new Node(null);
position.next = head;
50
}

return (position.next != null);
}

position = position.next;
Object o = position.data;
return o;
}
}
// Export a fresh iterator

public Iterator get_iterator(){
return new Iter();
}
...
}
Notice that when we return an iterator, since we do not need to ever use a name for the
iterator locally, we can directly pass the reference generated by new, rather than assigning
it temporarily to a locally defined reference (such as Iter it in the original version) and
returning that local variable.
5.3 Ports and interaction with state

An Iterator is a simple example of a situation where the interaction with an object is
not memoryless. Normally, when an object provides access to a private variable localdata
through an accessor method get_localdata(), each call to get_localdata() is independent
and the value returned does not depend on how many times, if any, get_localdata() has
been invoked earlier.
In contrast, the accessor method get_next() provided by the Iterator interface returns
different values depending on the history of previous calls to the same method. Thus, some
state information has to be maintained by the object to ensure that the value returned by
get_next() is accurate.
In general, the OO paradigm does not provide any direct way of specifying such in-
teractions with state with an object. The word port has been used recently in certain
specification methodologies (e.g., UML) to refer to a point of interaction with an object
that is accompanied by a state. As we have seen with Iterators, ports can be simulated
indirectly by exporting objects from within a class that implement a specified interface.
51
52
Chapter 6
Java Interlude 2
Parameter passing, cloning, packages
6.1 Parameter passing in Java

In Java, scalar variables (i.e. of type int, long, short, float, double, byte, char, boolean)
are always passed to functions by value, like in C. In other words, the value passed to the
function is copied into the local parameter and any change to the local copy has no effect on
the original copy.
Unlike C, there is no way to pass the address of a scalar. We cannot write the equivalent
of the following code in C:
int i,j;
...
swap(&i,&j);
...
void swap(int *p, int *q){

int tmp;
tmp = *p;
*p = *q;
*q = tmp;
}
Objects (i.e., any variable that is not a scalar) are passed by reference, like arrays in C.
Like C, the justification is that objects can be large structures and it would be inefficient to
copy the structure. Notice, however, that (like in C), we can modify the contents of what a
parameter points to, but we cannot modify where it points. If we want to swap the contents
of two objects from the same class, we cannot write:
class Myclass{
53
...
public void swap(Myclass p){ // Swap "this" with p
Myclass tmp;
tmp = p;
p = this;
this = tmp;
}
}
Instead, we must write something like:
class Myclass{
...
public void swap(Myclass p){
Myclass tmp = new Myclass(...); // Make a new tmp object
... // Copy contents of p into tmp
... // Copy contents of this into p
... // Copy contents of tmp back into this
}
}
What about return values? Suppose we enhance the Employee class as follows:
class Employee{
private String name;
private double salary;
private Date joindate;
// Constructor
public Employee(String n, double s, int d, int m, int y){
name = n; salary = s;
joindate = new Date(d,m,y); //
}
...
// "accessor" methods
public Date get_joindate(){ return joindate; }
...
}
where Date is defined as follows:
class Date{
54
day = d;
month = m;
year = y;
}
public void advance(int d){ // Modify day/month/year so that

... // date advances by d days
}
...
}
Now, suppose we use get_joindate() as follows:
Employee e = new Employee(...);

Date d = e.get_joindate();
d.advance(100); // e loses 100 days seniority!
The problem is that get_joindate() returns a reference to its private copy of Date and
this reference can be used to invoke public methods on the Date object, though the Date
object is supposed to be a private component of the Employee e.
The solution is to make a copy of the object before returning it.
class Employee{
...
public Date get_joindate(){
Date tmpdate = new Date(...); // Construct tmpdate using
// current values of joindate
return tmpdate;
}
...
}
Here, copying Date is not difficult because Date is itself a flat object and can be copied
component-wise. If we have to make a copy of Employee, we have to first recursively make
a copy of the nested object Date and then copy the scalars salary and name.
6.2 Cloning
The root class Object in Java provides a method
Object clone(Object o)
55
that makes a bit-wise copy of its argument. Making a bit-wise copy is fine if the the object
being copied consists only of scalars. However, if it has nested objects (like Date inside
Employee), the bitwise copy will copy the reference to the nested object Date so the clone
will accidentally have a reference to a non-cloned internal component. We should, instead,
recursively clone the internal object.
To ensure that a programmer is aware of the perils of using the clone method, Java
requires that any class that invokes clone must implement the interface Cloneable. To use
clone in Employee, we must write
class Employee implements Cloneable{

...
}
(Technically, we also need to redefine clone to be a public method. To understand this,

we need to look at accessibilities other than public and private in Java, which we will do
shortly.)
What does the interface Cloneable specify? Nothing! It is an empty interface, also called
a marker interface that merely allows the method clone() to check if the class in which
it is being invoked has been written by a thoughtful programmer who has understood
the perils of making a bit-wise copy and (hopefully) taken necessary steps to make it work
correctly.
The code for the method clone would contain a check something like the following:
Object clone(Object o){

if (o instanceof Cloneable){
... // go ahead and clone
}else{
... // complain and quit
}
}
6.3 Packages
So far, we have identified two basic modifiers indicating the range within which a variable or
function is accessible: public and private. As we have said already, public data is visible
to all classes and private data is visible only within a class (but to all instances of the same
class, so one object of a class can freely examine private variables of another object of the
same class).
In all this, we have also frequently written definitions without modifiers, such as
class LinearList{
...
}
56
How visible are such definitions?
Java has a separate mechanism to group togther classes using packages. A package,
roughly speaking, is a group of classes that are defined in the same directory. To indicate
that the class XYZ defined in the file XYZ.java belongs to the package named ABC, we add the
line package ABC as the first line in XYZ.java. There is a (somewhat arbitrary) restriction
on package names: all files in package ABC must live in a directory called ABC.
If a file does not carry a package header, its contents implicitly belong the (unique)
anonymous package. All files without a package header in the same directory belong to the
same anonymous package.
Definitions without modifiers are visible within the same package. The class definition
above is public within the package but private outside the package. So, in our toy
programs, all definitions without modifiers behave like public definitions so long as the
definitions lie within the same directory (and thus belong to the anonymous package).
Java packages are arranged in (multiple) directory trees. Java searches for classes using
the variable CLASSPATH, similar to the PATH variable used by the Unix shell to search for
executable files. Suppose CLASSPATH contains the directory /usr/local and under this, we
have a class def that lies in the subdirectory /usr/local/java/xyz/abc. Then, the package
name associated with def corresponds to the directory abc and is given by java.xyz.abc
while the fully qualified class name of def is java.xyz.abc.def. This reflects the path to
reach def relative to the CLASSPATH, with / replaced by ..
The builtin classes in java typically lie in packages of the form java.xyz.abc. To use
classes defined in a package we can use an import statement at the top of the file. For
instance, if we write
import java.math.BigDecimal
then the class BigDecimal defined under .../java/math becomes accessible in the current
program. More generally, we write
import java.math.*
to make all classes defined in .../java/math accessible to the program. Note here that the
* is not recursive, so it does not make available any classes that may lie in subdirectories of
.../java/math.
If we do not use the import statement, we can always explicitly select a class by giving
its fully qualified name (much like giving the full pathname of a Unix command) and write,
for instance,
java.math.BigDecimal d = new java.math.BigDecimal();
instead of
import java.math.*
BigDecimal d = new BigDecimal();
Packages are intended to correspond to natural units of software development so it makes
sense for some definitions to be fully visible within the current project but not outside.
57
6.4 The protected modifier
In addition to public, private and no modifier (implicitly indicating package visibility),
there is another modifier called protected, indicating visibility within the inheritance sub-
tree rooted at a classthat is, the collection of all classes that directly or indirectly extend
the current class.
While the usefulness of protected mode instance variables is doubtful, it does make some
sense to have protected mode functions, that can be used by subclasses that can be trusted
to know what these functions are doing, but not by arbitrary classes outside the subtree.
It is allowed for a subclass to promote a function from protected to public while
redefining it. Note that this is a specific exception to the general rule that says that a
subclass is not allowed to change the visibility of a function when redefining it. (It is also
not permissible for a subclass to change the return type of a function when redefining it.)
One concrete example where this is used is in the function clone() defined in Object.
As we have already noted, clone() makes a bitwise copy, which is not necessarily the correct
way to clone an object, so the clone() method checks that the calling class implements the
marker interface Cloneable() to verify that the programmer actually intends that clone()
be used.
As a further precaution against accidental misuse, clone() is declared to be protected
in Object. This means that by default, even after implementing the Cloneable interface,
only subclasses can invoke clone() on objects of a given class. To make an object fully
cloneable, the class must redefine clone() with the modifier public.
58
Part II
Exception handling
59
Chapter 7
Exception handing
In programming language parlance, an exception is an unexpected event that disrupts the

normal execution of a program. Exactly what constitutes an exception can be a subjective
decision. For instance, an illegal operation such as division by zero or accessing an array
element beyond the bound of the array is almost surely an exception. How about reaching
the end of the file when reading a file? This can be viewed as an interruption of the normal
operation of reading a file. Java takes the stand that reaching the end of file is an exceptional
situation. Most other programming languages would agree that though this is a special
situation, it will generally occur at the end of every read loop and so can be dealt with in
the normal program flow (such as checking that the character read is EOF).
The key problem is to deal with exceptions in a sensible way. The aim is for the program
to be able to recover and take corrective action if possible and abort only if there is no other
option left. Further, it should be possible to report sensible information about the nature of
the exception for diagnostic purposes.
In C, a very rudimentary form of exception reporting is done by examining the return
value of a function such as main(). Often, a return value of int encodes the result of the
computation of the functionminimally, a return value 0 may indicate no error while a
non-zero value might indicate an error. The type of error might be indicated by the actual
non-zero return valuereturning a value 1 might indicate malformed input while returning
a value 2 might indicate the inability to open a file that was need for the program to run.
7.1 Javas try-catch blocks

Java treats an exception as an object with structure. When an error happens, an appropriate
exception object is thrown back to the program that generated the error. The instance
variables of this object contain information about the nature of the exception. At the very
least, there is an error message stored in the object as a string.
If the program is written properly, it can catch the exception object, analyze its contents
to determine the cause of the error and then take corrective action if necessary. This is
known as handling the exception.
61
In Java, all exception objects are subclasses of a class called Throwable. There are two
immediate subclasses of Throwable, called Error and Exception. The class Error covers
exceptional situations that arise because of circumstances beyond the programs controlfor
instance, the kind of error that occurs if a file that needs to be read cannot be found on the
disk. The other subclass, Exception, has one subclass RunTimeException, that captures
standard runtime computational errors such as divide-by-zero and out-of-bounds-access of
an array.
Java supports a try-catch structure to catch exceptions:
try{
... // Code that might generate error
...
}
catch (ExceptionType1 e1){...} // Analyze contents of e1
// Corrective code for ExceptionType1
catch (ExceptionType2 e2){...} // Analyze contents of e1

// Corrective code for ExceptionType2
If an error occurs in the try block, the exception object generated is passed to the
(multiple) catch statements accompanying the try block. The type of the exception object
is checked sequentially against each catch statement till it matches. In the example above,
if the first catch statement were to be rewritten as
catch (Throwable e1){...}
all exception objects would match Throwable and be handled by this piece of code, ignoring
the rest. One should, in general, arrange the catch statements so that the earlier statements
deal with more specific exception types than later catch statements.
The sequence of flow when an exception occurs is as follows:
Abort the code within the try block at the point where the exception occurs.
Transfer control to the first catch block that matches the thrown exception object.
If no such block is found, abort and pass on the exception object to the calling class
(this allows exceptions to be propagated up to a level that can handle them).
If some catch block does match, execute this block and then proceed to the next new
statement after the catch.
62
7.2 Cleaning up with finally
This does not provide a mechanism to do any cleaning up of system resources (e.g., closing
open files) when an exception occurs. The solution is to add a block labelled finally, as
follows.
try{
... // Code that might generate error
...
}
catch (ExceptionType1 e1){...} // Corrective code for ExceptionType1
catch (ExceptionType2 e2){...} // Corrective code for ExceptionType2
finally{
... // Always executed, whether try terminates normally or
... // exceptionally. Use for cleanup statements
}
The semantics of finally is that it is always invoked, whether or not try terminates
normally. If the try block does not raise an exception, after the try block control passes to
the finally block and then to the next statement. If an exception does occur, the finally
block is executed after the appropriate catch block. If no catch block applies, the finally
block is executed before aborting the function and propagating the error back to the caller.
7.3 Customized exceptions

In addition to catching exceptions, Java allows you to generate your own exceptions using
the throw statement. The default Exception objects can store a string, indicating an error
message. In addition, you may add more structure in a user defined subclass of Exception.
If you do not want class LinearList to have negative entries you might write:
class NegativeException extends Exception{
private int error_value; // Stores negative value that

// generated the exception
public NegativeException(String message, int i){ // Constructor

super(message); // Appeal to superclass constructor for message
error_value = i;
}
public int report_error_value(){
63
return error_value;
}
}
Now, inside LinearList you might write:
class LinearList{
...
public add(int i){
...
if (i < 0){
throw new NegativeException("Negative input",i);
}
...
}
Notice that we directly passed the reference returned by new rather than creating a
temporary local variable of this type as follows:
public add(int i){

NegativeException ne;
...
if (i < 0){
ne = new NegativeException("Negative input",i);
throw ne;
}
...
}
When another program uses add(i) in a LinearList, it should be aware of the possi-
bility that a NegativeException may be thrown. Java insists that this exception throwing
capability be advertised in the function definition, as follows:
class LinearList{
...
public add(int i) throws NegativeException{
...
}
...
}
64
Actually, Java relaxes the rule about advertising exceptions so that a program can throw
a built in exception of type Error or RunTimeException without advertising it.
We can now use add(i) externally as:
LinearList l = new LinearList();
...
try{
...
l.add(i);
...
}
catch (NegativeException ne){
System.out.print("Negative input supplied was ");
System.out.print(ne.report_error_value);
}
...
As mentioned earlier, Java takes a rather extreme stand in classifying non-normal
execution as an exception. Reaching the end of file throws an IOException, so the normal
structure of a read loop from a file in Java would be:
try{
while(...){
... // read from a file
}
}
catch (IOException e){
if (e ... EOF) { ...} // if e indicates EOF
...
}
7.4 A cute fact about finally

Normally, we believe the following two pieces of code to be equivalent:
Iteration using for Iteration using while
for (i = 0; i < n; i++){ i = 0;

statement 1; while (i < n){
statement 2; statement 1;
... statement 2;
statement k; ...
} statement k;
i++;
}
65
However, observe what happens if one of the internal statements is continue (or, more
precisely in Java, continue l where l is a loop label).
l: l:
for (i = 0; i < n; i++){ i = 0;
statement 1; while (i < n){
statement 2; statement 1;
... statement 2;
continue l; ...
statement k; continue l
} statement k;
i++;
}
The statement continue l causes the loop l to start its next iteration without executing
any of the statements that follow it in the loop. In the for loop, this still results in the
increment i++ being executed since it is part of the loop definition and not a statement
inside the loop. In the while loop, on the other hand, the increment i++ is skipped along
with statement k so that the loop restarts with the same value of iq.
To fix this, we should use the try-finally construction (without a catch clause) as
follows:
l:
i = 0;
while (i < n){
try{
statement 1;
statement 2;
...
continue l
statement k;
}
finally{
i++;
}
}
Now, the continue l causes the try block to terminate prematurely, but the finally
block is still executed before reentering the loop, so this version does preserve the semantics
of the for.
66
Part III
Event driven programming
67
Chapter 8
Event driven programming
8.1 Programming graphical user interfaces

When programming modern graphical user interfaces, we have to deal with an important new
variety of programming, normally called event driven programming. The operating system
captures external events such as keystrokes and mouse movements and passes them on
to the program we write, which must then react sensibly. For instance, if the mouse is
positioned over a button and then clicked, we must activate the function indicated by the
button.
At a low level, the minimum support that we need is the ability to run a main routine in
parallel with a secondary routine that responds to these events. For instance, if the program
is a web browser, then the main routine will display web pages. Whenever a hyperlink is
selected, the act of displaying the current page is interrupted and the browser initiates a
connection to a new page.
The next question is the form in which the events are passed. At the most basic level, the
events of interest are of the form key a was pressed, the mouse moved to position (x, y)
or the left button of the mouse was pressed. If the program only gets such information
from the underlying system, then the programmer has to do a lot of work to keep track of
the current position of all the graphical objects being displayed so that the events can be
correlated to the position of these objects.
Consider, for instance, how the program would figure out that a button labelled OK has
been selected by the user. First, the program has to remember the current location of
the button on the screen. Next, the program has to track all mouse move events to know
where the mouse is at any point. Suppose the program now gets an event of the form mouse
click, and the currently recorded position of the mouse is (x, y). If (x, y) lies within the
boundary defined by the button labelled OK, then the user has actually clicked on the button
OK and appropriate action must be taken. On the other hand, if (x, y) does not lie within
the current boundaries of the OK button, then some other button or graphical component
has been selected and we then have to figure out which component this is. Or, yet again, it
may be that this mouse click is outside the scope of all the windows being displayed by the
69
current program and can hence be ignored.
Needless to say, programming graphical displays at this low level is very tedious and error-
prone. We need a higher level of support from the run-time environment of the programming
language. The run-time environment should interact with the operating system, receive
low level events such as keystrokes and mouse movements and automatically resolve these
into high level events indicating when a component such as a button has been pressed.
Thus, the programmer does not have to keep track of where a graphical component is or
which component is affected by a particular low level event: implicitly, whenever a graphical
component is selected, it gets a signal and can invoke a function that takes appropriate
action.
This support is available in Java through the Swing package, which in turn relies on
a lower level part of Java called AWT (Abstract Windowing Toolkit). In Swing, we can
directly define graphic objects such as buttons, checkboxes, dialogue boxes, pulldown menus
. . . and specify the size, colour, label . . . of these components. Each of these components is
defined as a builtin class in Swing. In addition, each component is capable of causing some
high level events. For instance, a button can be pressed, or an item can be selected in a
pulldown menu. When an event occurs, it is passed to a prespecified function.
How do we correlate the objects that generate events to those which contain the functions
that respond to these events? Each component that generates events is associated with
a unique collection of functions that its events invoke. This collection is specified as an
interface. Any class that implements this interface is qualified to be a listener for the events
generated by this type of component. The component is then passed a reference to the
listener object so that it knows which object is to be notified when it generates an event.
For instance, suppose we have a class Button that generates a single type of event,
corresponding to the button being pushed. When a button is pushed, we are told that a
function called buttonpush(..) will be invoked in the object listening to the button push.
We handle this as follows:
interface ButtonListener{
public abstract void buttonpush(...);
}
class MyClass implements ButtonListener{

...
public void buttonpush(...){
... // what to do when a button is pushed
}
...
}
...
Button b = new Button();
70
MyClass m = new MyClass();
b.add_listener(m); // Tell b to notify m when it is pushed
The important point is that we need not do anything beyond this. Having set up the
association between the Button b and the ButtonListener m, whenever b is pressed (wher-
ever it might be on the screen), the function m.buttonpush(...) is automatically invoked
by the run-time system.
Why does buttonpush(...) need arguments? The event generated by a Button has
structure, just like, say, an exception or any other signal. This structure includes information
about the source of the event (a ButtonListener may listen to multiple buttons) and other
useful data.
In a sense, the Timer example that we saw in Chapter 5.1 fits in this paradigm. Recall
that we wanted to start off a Timer object in parallel with the current object and have it
report back when it finished. The act of finishing is an event and the function it triggers
is the one we called notify(). In the example, the class that created the Timer passed a
pointer to itself so that the Timer object could notify it. We could instead have passed a
reference to any object that implements the interface Timerowner. Thus, the object we pass
to Timer is a listener that listens to the event generated by the Timer reaching the end of
the function f().
The relationship between event generators and event listeners in Java is very flexible.
Multiple generators can report to the same listener. This is quite naturalfor instance,
if we display a window with three buttons, each of which describes some function to be
performed on the content of the window, it makes sense for the window to listen to all three
buttons and take appropriate action depending on which button is pressed.
More interestingly, the same event can be reported to multiple listeners. This is called
multicasting. A typical example is when you want to close multiple windows with a single
mouse click. For instance, suppose we have opened multiple windows in a browser. When
we exit the browser (one click) all these windows are closed.
This flexibility also means that the connection between event generators and event listen-
ers has to be set up explicitly. If no listener is notified to the event generator, the events that
it generates are lost. In some languages, each component (such as a Button) is automat-
ically associated with a fixed listener object. The advantage is that it is almost automatic
that all events are listened to and not lost. However, the disadvantage is a loss of flexibility
in the ways we can interconnect event generators and event listeners.
8.2 Some examples of event driven programming in

Java
8.2.1 A button
A button is a graphical object that can be clicked. There is only one thing that a user can
do with a buttonclick it. Thus, when a button is clicked, an appropriate listener has to be
71
informed that it was clicked, along with the identity of the button. The built-in button class
in Swing is called JButton. The interface ActionListener describes classes that can listen
to button clicks. When a button is clicked, the method actionPerformed(...) is invoked
in each ActionListener that is associated with this button.
Thus, to construct a button in Java, we need to declare a JButton and provide a class
that listens to it. At the minimum we need the following.
class MyButtons{
private JButton b;
public MyButtons(ActionListener a){

b = new JButton("MyButton"); // Set the label on the button
b.addActionListener(a); // Associate an listener
}
}
class MyListener implements ActionListenerP

public void actionPerformed(ActionEvent evt){
... // What to do when a button
... // is pressed
}
class XYZ{
MyListener l = new MyListener(); // Create an ActionListener l
MyButtons m = new MyButtons(l); // Create a button m, listened to by l
Unfortunately, we have to do a lot more to actually display a button. First, we have

to generate a container to hold the button. Usually, this is done using an object called
a panel the corresponding class in Swing is JPanel. Here is how to define a panel that
contains a single button. The enclosing panel listens to the button click. The import
statements at the top include the appropriate packages from AWT and Swing.
import java.awt.*;
import java.awt.event.*;
import javax.swing.*;
public class ButtonPanel extends JPanel implements ActionListener{
private JButton redButton;
public ButtonPanel(){
redButton = new JButton("Red"); // Create the button
redButton.addActionListener(this); // Make the panel listen
72
// to the button
add(redButton); // Embed the button in the
// panel
}

Color color = Color.red; // Set the background colour
setBackground(color); // to red when the button
repaint(); // is clicked
}
In Java, a panel cannot be displayed directly. We have to embed the panel into a frame
in Swing, the class we need is JFrame. A JFrame can generate seven different types of events
that are handled by objects that implement WindowListener. These events correspond to
actions such as iconizing a window, making a window occupy the full screen, killing the
window, . . . . Each of these actions invokes a different function in WindowListener. It is
important to note that this classification of the seven different types of actions is done auto-
matically by Swingwe do not need to write an elaborate case-analysis in a WindowListener
to decipher the nature of the event.
Here is a simple frame that includes a ButtonPanel. The class ButtonFrame acts as its
own WindowListener. For six of the seven functions, we have provided just a stub. The only
window event that ButtonFrame handles is the one that kills the window. If the window
is killed, the corresponding Java program exits. If we do not write something appropriate
for the method windowClosing(WindowEvent e), when the user kills the window the win-
dow manager will delete the frame but the underlying Java program will continue to run
invisibly!
The only other complication in this example is the use of ContentPane. A JFrame
is a complex object, made up of different conceptual layers. One of these layers is
called the ContentPane. When we add objects to the JFrame, we actually add them to the
ContentPane. Thus, to add a ButtonPanel to a ButtonFrame, we first get access to the
ContentPane of the ButtonFrame and add the ButtonPanel to the ContentPane.
import java.awt.*;
public class ButtonFrame extends JFrame implements WindowListener {

Private Container contentPane;
public ButtonFrame(){
setTitle("ButtonTest");
73
setSize(300, 200);
addWindowListener(this); // ButtonFrame is its
// own listener
contentPane = this.getContentPane(); // ButtonPanel is added
contentPane.add(new ButtonPanel()); // to the contentPane
}
// Seven methods required for implementing WindowListener

// Six out of seven are dummies (stubs)
public void windowClosing(WindowEvent e){ // Exit when window

System.exit(0); // is killed
}
public void windowActivated(WindowEvent e){}
public void windowClosed(WindowEvent e){}
public void windowDeactivated(WindowEvent e){}
public void windowDeiconified(WindowEvent e){}
public void windowIconified(WindowEvent e){}
public void windowOpened(WindowEvent e){}
Finally, here is a main method that creates and displays a ButtonFrame.
import java.awt.*;
public class ButtonTest
{ public static void main(String[] args)

{ JFrame frame = new ButtonFrame();
frame.show();
}
}
74
8.2.2 Three buttons
Suppose we want to have a panel with three buttons, labelled Red, Blue and Yellow, so that
clicking each button results in the background colour changing to the appropriate colour.
We can modify the earlier ButtonPanel to contain three buttons.
The ActionListener associated with each button has to be able to inform the ButtonPanel
to change its background colour. The simplest solution is to make the ButtonPanel itself
the ActionListener for all three buttons. However, we now have a single ActionListener
handling events from multiple sources. In ActionPerformed, we can use the getSource()
method to find out the source of the event and then take appropriate action.
Here is the revised code for ButtonPanel for this example. The code for ButtonFrame
does not change.
import java.awt.*;
private JButton yellowButton; // Panel has three buttons

private JButton blueButton;
yellowButton = new JButton("Yellow");
blueButton = new JButton("Blue");
redButton = new JButton("Red");
yellowButton.addActionListener(this); // ButtonPanel is the

blueButton.addActionListener(this); // listener for all
redButton.addActionListener(this); // three buttons
add(yellowButton);
add(blueButton);
add(redButton);
}

Object source = evt.getSource(); // Find the source of the
// event
Color color = getBackground(); // Get current background
// colour
75
if (source == yellowButton) color = Color.yellow;
else if (source == blueButton) color = Color.blue;
else if (source == redButton) color = Color.red;
setBackground(color);
repaint();
}
}
Some more examples can be found in Appendix A.
8.3 The event queue

Recall that the operating system interacts with the run-time support for the programming
language to pass on events such as keystrokes and mouse movements. These are automat-
ically resolved into high level events in terms of the components being manipulated by the
program, such as buttons, etc.
Events are despatched sequentially by the run-time support system to the program. Thus,
if the user clicks a button such as redraw in a graphics application, until the redrawing is
completed the application cannot respond to other events. However, the user is clearly free
to move the mouse and press some keys while the redrawing is in progress. These events are
stored in an event queue and processed in turn.
One advantage of storing events in a queue is that they can be optimized. For instance, if
there are three mouse movements while the redrawing is in progress, it is sufficient to record
the position of the mouse at the end of the last of the three movements in the queue.
What events go in the queue? We have so far indicated that the program receives only
high level, or semantic, events such as button b was pushed. These are synthesized by
the run-time support system from the low level events that it receives from the operating
system.
However, a program may also like to receive low level events such as mouse movements
and key strokes. Consider a drawing application where the user can draw a line segment
by using the mouse to click both end points. After selecting the first point, it would be
complicated if the user could wander off and do other things before identifying the second
point. In such a situation, after being notified that the first point has been selected, which
is perhaps a high level event, the program can go into a loop where it explicitly consumes
events in the queue until it sees the next mouse click within the drawing area. All irrelevant
intervening events are discarded. Thus, by examining low level events, the program can
grab control of the user and force the user to act in a predictable way.
The low level support for event management in Java is done by a portion of the AWT
system. The AWT system does in fact insert both low level and high level events in the
queue. The low level events are essentially those received from the operating system, but
these may be combined and optimized as indicated earlier. In addition, AWT keeps track
76
of how high level components are affected by these low level events and inserts appropriate
high level events into the queue.
Low level events have listener interfaces, just like high level ones. Thus, normally, a Java
program interacts with the event queue implicitly, by defining appropriate listeners for the
various events of interest. When an event reaches the head of the queue, it activates all the
listeners that exist for it. If there are no listeners active, the event is discarded.
In addition, a Java program can also explicitly interact with the event queue. Java has
a built-in class called EventQueue. An object of this class can hold a pointer to the current
system event queue. This is done as follows.
EventQueue evtq = Toolkit.getDefaultToolkit().getSystemEventQueue();
It is then possible to consume an event from the head of the event queue explicitly, as
follows.
AWTEvent evt = eq.getNextEvent();
(It is also possible to peek at the event at the head of the queue without consuming
it.)
We can use this explicit event queue manipulating facility to write event grabbing code
such as the line drawing application discussed earlier that insists on a second point being
selected after the first point has been placed.
Dually, we can also add events to the queue from within the program by writing something
like
evtq.postEvent(new ActionEvent(.....));
Thus, when one button is pressed, one of the effects could be to internally generate an
event corresponding to another button being pressed. For instance, a button labelled Exit
might generate a frame-closing event for its enclosing frame which, when handled, will shut
down the application.
8.4 Custom events

Just like we can add new types of exceptions in a Java program, we can also define new types
of events. Let us look at how to add a new type of event called TimerEvent to be generated
by a Timer object.
Events in Java (generally) extend the class AWTEvent. Each distinct type of event has a
unique numeric identifier. We can thus write
class TimerEvent extends AWTEvent{
public static final int
77
TIMER_EVENT = AWTEvent.RESERVED_ID_MAX + 5555;
public TimerEvent(Timer t) {
super(t, TIMER_EVENT);
}
}
Here, we use a language defined constant that delimits the numeric identifiers in use for
built-in events to ensure that our new event does not clash with any of the built-in events.
The constructor essentially sets two quantities: the source of the event and the (numeric)
type of the event.
Next, we define an interface that specifies the functions that a listener for TimerEvent
must implement. This might be as follows
interface TimerListener extends EventListener{

public void timeElapsed(TimerEvent evt);
}
Now, we can make the class Timer generate such events by explicitly manipulating the
event queue, as described earlier.
class Timer extends Component implements Runnable{

// Runnable is to ensure that Timer can run concurrently
...
private static EventQueue evtq =
Toolkit.getDefaultToolkit().getSystemEventQueue();
// There is only one copy of the EventQueue, shared by all Timers
public Timer(){
...
}
public void f(){

...
// Generate a TimerEvent when done
TimerEvent te = new TimerEvent(this);
evtq.postEvent(te);
}
}
Of course, we need a class that listens to TimerEvents, so we have
class MyClass implements TimerListener{

...
78
public void timeElapsed(TimerEvent evt){
...
}
}
We need to inform the Timer object of the existence of a specific listener object from
MyClass. For this, we have to add the following in Timer.

...
private TimerListener listener;
public addListener(TimerListener t){

listener = t;
}
...
}
In other words, it is the duty of the Timer object to keep track of the listeners it
has to notify. This is the general Java philosophy, which is why we have methods like
addActionListener(..) for JButtons etc.
Assume now that we have code like the following:
...
Timer t = new Timer();
MyClass m = new MyClass(..);
t.addListener(m);
t.f();
...
At the end of t.f(), how does t know what it has to do to notify m? The answer is that
it does not. This has to be programmed in the definition of Timer. What the AWT system
provides is the name of a fixed function called processEvent that will be called within an
object whenever it generates an AWTEvent. We can now add the following code in Timer.

...
public void processEvent(AWTEvent evt)

{ if (evt instanceof TimerEvent)
{ if (listener != null)
listener.timeElapsed((TimerEvent)evt);
}
else super.processEvent(evt);
79
}
...
}
Now, all the pieces are in place. When a Timer t generates a TimerEvent, AWT calls
the (fixed) method processEvent within t. This provides the explicit code with which to
call any listener associated with t. Notice that instead of one TimerListener, we could
have defined Timer to store an array of listeners and notified each of them in turn, thus
supporting multicasting.
The same general architecture drives the functioning of builtin objects like JButton etc.
Externally, it appears as though the run-time environment knows that when a JButton is
pressed, the correct method to invoke in an ActionListener is actionPerformed. However,
all that the button press event does is to invoke processEvent in the JButton object that
is pressed. The definition of JButton, like the definition that we have just given for Timer,
includes all the necessary information about its listeners and the methods to use to notify
these listeners that the button has been pressed.
80
Part IV
Reflection
81
Chapter 9
Reflection
A programming language is said to be reflective 1 if it can represent (part of its) state as data
and manipulate it. There are two components involved in reflection:
Introspection: The ability of a program to observe and therefore reason about its
own state.
Intercession: The ability of a program to modify its execution state or alter its own
interpretation or meaning.
Constructions such as the following
Employee e = new Manager(...);
...
if (e instanceof Manager){
...
}
are very simple examples of reflection at work. We can, at run time, check the actual type
of an object.
However, notice that in this case we know the type that we want to check in advance
both Employee and Manager are already defined when this code is written. This is important
because the second argument to instanceof must be a fixed name for a class, not a class
variable.
What if we wanted to check whether two pointers are the same at run-time? We would
like to write a function like:
public static boolean classequal(Object o1, Object o2){
...
// return true iff o1 and o2 point to objects of same type
...
}
1
The word reflection is used in the philosophical sense, meaning to think seriously.
83
To implement this using instanceof, we would have to check this for every possible class
that we know whether o1 and o2 are both instances of the same class. This is infeasible for
two reasons:
We could not reasonably expect to enumerate all the classes defined in the system,
even if this set was fixed.
More critically, we would like to use this function in any Java environment, including
those in which classes that are yet to be defined may be loaded!
It turns out that Java allows us to extract the class of an object at runtime, using the
method getClass(), defined in the Java reflection package java.lang.reflect. Thus, we
write the following code to implement the classequal we were looking at earlier.
import java.lang.reflect.*;
class MyReflectionClass{
...
return (o1.getClass() == o2.getClass());
}
}
What values are being compared in the test o1.getClass() == o2.getClass()? In

other words, what does getClass() return? For the purpose of classequal, it would be
sufficient to have a nave return value, such as a String containing the name of the the class.
In fact, getClass() does much more. It returns a representation of the class of the object
on which it is invoked. This means that there is a built in class in Java that can store such
a representation. This built in type is called, somewhat confusingly, Class! To make this
explicit, we can rewrite classequal as follows.
import java.lang.reflect.*;
class MyReflectionClass{
...
Class c1, c2;
c1 = o1.getClass();
c2 = o2.getClass();
return (c1 == c2);
}
}
84
Notice that we are representing dynamic information about the state of the system (the
runtime class of o1 and o2) in terms of data c1 and c2 that can be manipulated by the
program. This process of encoding execution state as data is called reification.
Having got a representation for a class in a Class object, we can do various things with
it. For instance, we can create new instances of this class.
...
Class c = obj.getClass();
Object o = c.newInstance(); // Create a new object of same type as obj
...
Another way of obtaining a class object is to supply the name of the class as a string to
the static function forName defined in Class.
...
String s = "Manager".
Class c = Class.forName(s);
Object o = c.newInstance();
...
We could, of course, simplify this into a single statement as follows:
...
Object o = Class.forName("Manager").newInstance();
...
Once we have a class stored in an object of type Class, we can find out details about
its constructors, methods and fields (instance variables). A constructor, method or field is
itself a complex quantity. For instance, to full describe a constructor we need to provide
the types of arguments. For a method we need the arguments and the return type. For all
three, we need the modifiers (static, private etc). Thus, it makes sense to have additional
classes called Constructor, Method and Field to store information about an individual
constructor, method or field. Given an object c of type Class, we can invoke the functions
getConstructors(), getMethods() and getFields() to obtain the list of constructors,
methods and fields of c. Each of these functions returns an array. So, for instance, we could
write
...
Constructor[] constructors = c.getConstructors();
Method[] methods = c.getMethods();
Field[] fields = c.getFields();
...
85
Not surprisingly, we can now operate on each of Constructor, Method and Field to get
further details of these objects. For instance, the list of arguments to a constructor or a
method is a list of types, or, in other words, an array of Class[]. Thus, we can write
...
Constructor[] constructors = c.getConstructors();
for (int i = 0; i < constructors.length; i++){
Class params[] = constructors[i].getParameterTypes();
..
}
We can also invoke methods and examine/set values of fields.
...
..
Method[] methods = c.getMethods();
Object[] args = { ... } // construct an array of arguments
methods[3].invoke(obj,args); // invoke method stored in methods[3]
// on obj with arguments args
...
Field[] fields = c.getFields();
Object o = fields[2].get(obj); // get the value of fields[2]
// from obj
...
fields[3].set(obj,value); // set the value of fields[3]
// in obj to value
...
Should it be possible to find out about private constructors, methods and fields of a
class? The functions getConstructors(), . . . only return publicly defined values. How-
ever, Java does permit us to access private components using variants of these functions
called getDeclaredConstructors(); getDeclaredMethods() and getDeclaredFields().
Of course, there is an issue of security involved here. In general, Java allows us to find out
about private methods or fields in a class, but the default security level will not allow us to
invoke private methods or access/modify private fields.
The reflective capabilities of Java allow us to write versatile code, such as the Java pro-
gramming environment BlueJ 2 that provides an interactive interface to define Java classes,
create objects, invoke methods on these objects, examine the state of these objects etc. The
important fact is that BlueJ does not have to implement a nested virtual machine to act as
2
Available at http://www.bluej.org
86
a Java interpreter. Using the reflective capabilities of Java, BlueJ can directly perform its
manipulations using the JVM within which it is itself running!
Of course, one might ask for even morethere is no way to directly create or modify new
classes by writing something like
Class c = new Class(....);
Creating class objects is done indirectly by the class loader, which is outside the scope of
this discussion. Nevertheless, Java does provide a rather powerful set of tools for reflective
programming.
87
88
Part V
Concurrent programming
89
Chapter 10
Concurrent Programming
In a distributed computing system, it is natural to regard the nodes as independent processes

that execute concurrently. In such a system, agents communicate with each other to coordi-
nate their activities. Communication may be through synchronization or message-passing.
Synchronization implies a simultaneous two-way exchange of information (like a telephone
call) while message-passing involves one-way transfers that require acknowledgments for the
sender to be sure that a message has reached the recipient (like a normal letter sent by post).
However, even in a single processor system, it is often convenient to regard different
components of a program as executing in parallel. Consider, for instance, an interactive
application like a web browser. When you are downloading a page that takes a long time to
retrieve, you can usually press the Stop button to terminate the download. When program-
ming a browser with such a capability, the most natural model is to regard the download
component and the user-interface component (that reacts to button clicks) as separate pro-
cesses. When the user-interface process detects that the Stop button has been clicked, it
notifies the download process that it can terminate its activity.
On a single processor, the run-time environment will allocate time-slices to each of these
logical processes. These time-slices will be scheduled so that all the concurrent processes
get a chance to execute. However, there is no guarantee about the relative speeds at which
these processes will runone process may get 10 more time-slices than another in a given
interval.
Normally, each concurrent process comes with its own local variables. Thus, when the
run-time environment switches from one process to another, it has to save the state of the
first process and load the suspended state of the second one. Often, however, it is simpler to
assume that all the concurrent processes share the same set of variables. Thus, the processes
interact via a global shared memory. This makes it possible to switch from one process
to another relatively easily, without an elaborate context switch involving all the variables
defined in the processes. In the literature, these kinds of processes that share a global
memory are often called threads (short for threads of execution).
We shall study concurrent programming in the framework of threads that operate with a
global shared memory. In the rest of these notes, the words thread and process will be used
synonymously.
91
10.1 Race conditions
Shared variables provide a simple mechanism for communication between processes. For
instance, in the example of the browser discussed earlier, we could have a boolean variable
that indicates whether the download should be interrupted. This variable is initially false.
When the user-interface thread detects that the Stop button has been clicked, it sets the
variable to true. Periodically, the download thread examines the value of this variable. If it
finds that the variable has been set to true, it aborts the current transfer.
Once we have shared variables, it becomes important to incorporate a mechanism for
consistent update of such variables by concurrent threads. Consider a system in which we
have two threads that update a shared variable n, as follows:
Thread 1 Thread 2
... ...
m = n; k = n;
m++; k++;
n = m; n = k;
... ...
Under normal circumstances, after these two segments have executed, we would expect
the value of n to have been incremented twice, once by each thread. However, because of
time-slicing, the order of execution of the statements may be as follows;
Thread 1: m = n;
Thread 1: m++;
Thread 2: k = n; // k gets the original value of n
Thread 2: k++;
Thread 1: n = m;
Thread 2: n = k; // Same value as that set by Thread 1
In this sequence, the increments performed by the two threads overlap and the final
value of n is only one more than its initial value. This kind of inconsistent update, whose
effect depends on the exact order in which concurrent threads execute, is known as a race
condition.
Here is a more complex example of the same phenomenon. Suppose we have a shared
array
double accounts[100]
that holds the current balance for 100 bank accounts. Suppose we have have two functions
as follows:
92
// transfer "amount" from accounts[source] to accounts[target]
boolean transfer (double amount, int source, int target){
if (accounts[source] < amount){
return false;
}
accounts[source] -= amount;
accounts[target] += amount;
return true;
}
// compute the total balance across all accounts

double audit(){
double balance = 0.00;
for (int i = 0; i < 100; i++){
balance += accounts[i];
}
return balance;
}
Now, suppose we execute these functions in concurrent threads, as follows:
Thread 1 Thread 2
... ...
status = transfer(500.00,7,8); print (audit());
... ...
Suppose that Thread 2 gets access to accounts[7] and accounts[8] after Thread 1 has
debited accounts[7] but before it has credited accounts[8]. The audit will then report
that a sum of 500.00 has been lost from the accounts in the bank, because it uses the
updated value of accounts[7] but the old value of accounts[8].
The situation is even more complicated than this. Even if Thread 1 executes both the
debit and the credit without interruption, it is possible that Thread 2 reads accounts[7]
before the transfer and accounts[8] after the transfer, thereby recording an excess of 500.00
in the total balance.
For these functions to execute concurrently in parallel, both transfer(...) and audit()
should execute from beginning to end without interruption. Stated another way, it should
never be the case that the current control point in one thread is within transfer(...) while
in the other thread it is within audit().
All of these examples can be formulated as instances of the familiar problem of mutual
exclusion to critical regions of program code, which is well-studied, for instance, in the design
of operating systems.
93
10.2 Protocols based on shared variables
One way to solve the mutual exclusion problem is to develop a protocol using auxiliary
shared variables. For instance, we could use a variable turn that takes values 1 and 2 to
indicate whose turn it is to access the critical region, as follows.
Thread 1 Thread 2
... ...
while (turn != 1){ while (turn != 2){
// "Busy" wait // "Busy" wait
} }
// Enter critical section // Enter critical section
... ...
// Leave critical section // Leave critical section
turn = 2; turn = 1;
... ...
Let us assume that turn is initialized to either 1 or 2 (arbitrarily) and there is no other
statement that updates the value turn other than the two assignments shown above. It
is then clear that both Thread 1 and Thread 2 cannot simultaneously enter their critical
sectionsif Thread 1 has entered its critical section, the value of turn is 1 and hence Thread
2 is blocked until Thread 1 exits and sets turn to 2. A symmetric argument holds if Thread
1 tries to enter the critical section while Thread 2 is already in its critical section. If both
threads simultaneously try to access the critical section, exactly one will succeed, based on
the current value of turn.
Notice that this solution does not depend on any atomicity assumptions regarding as-
signing a value to turn or testing the current value of turn.
However, there is one serious flaw with this solution. When Thread 1 executes its critical
section it sets turn to 2. The only way for Thread 1 to be able to reenter the critical section
is for Thread 2 to reset turn to 1. Thus, if only one thread is active, it cannot enter the
critical section more than once. In this case, the thread is said to starve.
Another solution is to maintain two boolean variables, request_1 and request_2, indi-
cating that the corresponding thread wants to enter its critical section.
Thread 1 Thread 2
... ...
request_1 = true; request_2 = true;
while (request_2){ while (request_1)
} }
94
... ...
request_1 = false; request_2 = false;
... ...
Here we assume that request_1 and request_2 are initialized to false and, as before,
these variables are not modified in any other portion of the code. Once again, it is easy to
argue that both threads cannot simultaneously be in their critical sectionswhen Thread 1 is
executing its critical section, for instance, request_1 is true and hence Thread 2 is blocked.
Also, if only one thread is alive, it is still possible for that thread to repeatedly enter and
leave its critical section since the request flag for the other thread will be permanently false.
However, if both threads simultaneously try to access the critical region, they will both set
their request flags to true and there will be a deadlock where each thread is waiting for the
other threads flag to be reset to false.
Peterson [1981] found a clever way to combine these two ideas into a solution that is free
from starvation and deadlock. Petersons solution involves both the integer variable turn
(which takes values 1 or 2) and the boolean variables request_1 and request_2.
Thread 1 Thread 2
... ...
request_1 = true; request_2 = true;
turn = 2; turn = 1;
while (request_2 && while (request_1 &&
turn != 1){ turn != 2){
} }
... ...
request_1 = false; request_2 = false;
... ...
If both threads try to access the critical section simultaneously, the variable turn deter-
mines which process goes through. If only one thread is alive, the request variable of the
other thread is stuck at false and the value of turn is irrelevant. To check that mutual exclu-
sion is guaranteed, suppose Thread 1 is already in its critical section. If Thread 2 then tries
to enter, it first sets turn to 1. While Thread 1 is active, request_1 is true. Thus, Thread
2 has to wait until Thread 1 finishes and turns off request_1. A symmetric argument holds
when Thread 2 is in the critical region and Thread 1 tries to enter.
95
10.3 Programming primitives for mutual exclusion
Petersons algorithm does not generalize easily to a situation where there are more than two
concurrent processes. For n processes, there are other solutions, such as Lamports Bakery
algorithm. However, the problem with these protocols is that, in general, they have to be
designed anew for different situations and checked for correctness.
A better solution is to include support in the programming language for mutual exclu-
sion. The first person to propose such primitive was Edsger Dijkstra, who suggested a new
datatype called a semaphore.
A semaphore is an integer variable that supports an atomic test-and-set operation. More
specifically, if a S is a variable of type semaphore, then two atomic operations are supported
on S: P(S) and V(S). (The letters P and V come from the Dutch words passeren, to pass,
and vrygeven, to release.)
The operation P(S) achieves the following in an atomic manner:
if (S > 0)
decrement S;
else
wait for S to become positive;
The operation V(S) is defined as follows:
if (there are threads waiting for S to become positive)

wake one of them up; //choice is nondeterministic
else
increment S;
Using semaphores, we can now easily program mutual exclusion to critical sections, as
follows:
Thread 1 Thread 2
... ...
P(S); P(S);
... ...
V(S); V(S);
... ...
Mutual exclusion, starvation freedom and deadlock freedom are all guaranteed by the
definition of a semaphore.
One problem with semaphores are that they are rather low-level. The definition of
a semaphore is orthogonal to the identification of the critical region. Thus, we connect
96
critical regions across threads in a somewhat arbitrary fashion by using a shared semaphore.
Correctness requires cooperation from all participating threads in resetting the semaphore.
For instance, if any one thread forgets to execute V(S), all other threads get blocked. Further,
it is not required that each V(S) match a preceding P(S). This could lead to a semaphore
being preloaded to an unsafe value.
10.4 Monitors
Since critical regions arise from the need to permit consistent updates to shared data, it
seems logical that information about conflicting operations on data should be defined along
with the data, using a class-like approach. This is the philosophy underlying monitors, a
higher-level approach to providing programming support for mutual exclusion. Monitors
were proposed by Per Brinch Hansen and Tony Hoare.
Roughly speaking, a monitor is like a class description in an object-oriented language.
At its heart, a monitor consists of a data definition, to which access should be restricted,
along with a list of functions to update this data. The functions in this list are assumed
to all be mutually exclusivethat is, at most one of them can be active at any time. The
monitor guarantees this mutual exclusion by making other function calls wait if one of the
functions is active.
For instance, here is a monitor that handles the bank account example from the previous
lecture:
monitor bank_account{
double accounts[100];

return false;
}
return true;
}

double audit(){
for (int i = 0; i < 100; i++){
}
97
return balance;
}
}
By definition, the functions transfer(...) and audit() are assumed to require mutu-
ally exclusive access to the array accounts[]. Thus, if one process is engaged in a transfer
and a second process wants to perform an audit, the second process is suspended until the
first process finishes.
Thus, there is an implicit external queue associated with the monitor, where processes
that are waiting for access are held. We use the word queue but, in practice, this is just a
set. No assumption is made about the order in which waiting processes will get access to
the monitor.
The capabilities we have described so far for monitors are not as general as, say, semaphores.
Consider for instance the following scenario.
We have three bank accounts, i, j and k and we want to perform the following two
operations concurrently.
transfer(500.00,i,j);
transfer(400.00,j,k);
If the balance in i is greater than 500.00, this pair of transfers should always succeed.
However, if accounts[j] is not at least 400.00 initially, there is a possibility that the second
transfer will be scheduled before the first and will fail.
We could try to get around this by reprogramming the transfer function as follows:

// wait for another transaction to transfer money
// into accounts[source]
}
return true;
}
The problem is that all other processes are blocked out by the monitor while this process
waits. We thus need to augment the monitor definition with the ability to suspend a process
and relinquish the monitor.
A process that suspends itself is waiting for the data in the monitor to change its state.
It does not make sense to place such a process in the same queue as one that is waiting
(unconditionally) to enter the monitor for the first time. Instead, it makes sense to a second
queue with a monitor where suspended processes wait, which we shall call an internal
queue, to distinguish it from the external queue where blocked processes wait.
98
We then need a dual operation to wake up a suspended process when the state changes
in a way that may be of benefit to the suspended processin this example, when another
transfer succeeds.
Using these ideas, we can now rewrite the transfer function as follows, where wait()
suspends a process and notify() wakes up a suspended process.

wait();
}
notify();
return true;
}
We now have to deal with another tricky problem: what happens when a notify() is
issued? The notifying process is itself still in the monitor. Thus, the process that is woken
up cannot directly execute.
At least three different types of notification mechanisms have been identified in the lit-
erature.
The simplest is one called signal and exit. In such a setup, a notifying process must
immediately exit the monitor after notification and hand over control to the process that
has been woken up. This means that notify() must be the last instruction that it executes.
We shall see later that we might use multiple internal queues in a monitor, each of which
is notified independently. With signal and exit, it is not possible to notify multiple queues
because the first call to notify() must exit the monitor.
Another mechanism is called signal and wait. In this setup, the notifying process and
the waiting process swap roles, so the notifying process hands over control of the monitor
and suspends itself. This is not very commonly used.
The last mechanism is signal and continue. Here, the notifying process continues in the
monitor until it completes its normal execution. The woken up process(es) shift from the
internal to the external queues and are in contention to be scheduled when access to the
monitor is relinquished by the notifying process. However, no guarantee is given that one
of the newly awakened processes will be the next to enter the monitorthey may all be
blocked out by another process that was already waiting in the external queue.
When a suspended process resumes execution, there is no guarantee that the condition
it has been waiting for has been achieved. For instance, in the example above, transfer is
waiting for an amount of at least 400 to be transferred into accounts[j]. Any subsequent
successful transfer will potentially notify() this waiting process, including transfers to
accounts other than j. Even if the transfer is into accounts[j], the amount might be
less than the 400 threshold required by the waiting copy.
99
The moral of the story is that a waiting condition should check the state of the data
before proceeding because the data may still be in an undesirable state when it wakes up.
We should rewrite the earlier code as:

while (accounts[source] < amount){
wait();
}
notify();
return true;
}
In other words, the if is replaced by a while so that the first thing done after waking
up from wait() is to check the condition again, before proceeding.
In this example, clearly not every successful transfer is relevant to a waiting process.
The only relevant transfers are those that feed into the account whose balance is low. Thus,
instead of a single internal queue, it makes sense to have multiple queuesin this case, one
queue for each account (in other words, an array of queues). Then, a successful transfer will
notify only the queue pertaining to the relevant account.
This means that we have to extend the monitor definition to include a declaration of the
internal queues (sometimes called condition variables) associated with the monitor. In the
example below, we assume we can declare an array of queues. In practice, this may not be
possiblewe may have to explicitly assign a separate name for each queue.
monitor bank_account{
queue q[100]; // one internal queue for each account

q[source].wait(); // wait in the queue associated with source
}
q[target].notify(); // notify the queue associated with target
return true;
}
100
double audit(){ ...}
If you notify() a queue that is empty, it has no effect. This is more robust than the
V(S) operation of a semaphore that increments S when nobody is waiting on S.
10.5 Monitors in Java

Java implements monitors, but with a single queue. The queue is anonymous (but is really
implicitly named by the object in which it is defined, as we shall see).
Any class definition can be augmented to include monitor like behaviour. We add a
qualifier synchronized to any method that requires exclusive access to the instance variables
of the class. Thus, if we have two or more functions that are declared to be synchronized
and one of them is currently being executed, any attempt to enter another synchronized
method will be blocked and force the thread into the external queue. In Java terminology,
there is a lock associated with the object. Only one thread can own the lock at one time
and only the thread that owns the lock can execute a synchronized method. At the end of
the synchronized method, the thread relinquishes the lock.
Inside a synchronized method, the wait() statement behaves like the normal wait()
statement in a monitor, except that there is only one (unnamed) queue that the process can
wait on. The corresponding wakeup call should be called notify() but in Java, for some
obscure reason, there are two functions notify() and notifyAll(). The function notify()
signals only one of the waiting processes, at random. Instead, notifyAll() signals all waiting
processes (as one would expect) and this should be used by default.
Java monitors use the signal and continue mechanism, so the notifying process can con-
tinue execution after sending a wakeup signal and the processes that are woken up move
to the external queue and are in contention to grab the object lock when it next becomes
available.
The Java implementation of the bank account class looks pretty much like the monitor
description given earlier. One difference is that it can include non-synchronized methods,
which do not require having control of the object lock.
public class bank_account{

public synchronized boolean
transfer (double amount, int source, int target){
101
wait();
}
notifyAll();
return true;
}

public synchronized double audit(){
for (int i = 0; i < 100; i++){
}
return balance;
}
// a non-synchronized method
public double current_balance(int i){
return accounts[i];
}
}
In addition to synchronized methods, Java has a slightly more flexible mechanism. Every
object has a lock, so an arbitrary block of code can be synchronized with respect to any
object, as follows:
public class XYZ{
Object o = new Object();
public int f(){

..
synchronized(o){
...
}
}
public double g(){

..
synchronized(o){
...
102
}
}
Now, f() and g() can both begin to execute in parallel in different threads, but only
one thread at a time can grab the lock associated with o and enter the block enclosed by
synchronized(o). The other will be placed in the equivalent of an external queue for the
object o, while waiting for the object lock.
In addition, there is a separate internal queue associated with o, so one can write.
public class XYZ{
Object o = new Object();
public int f(){

..
synchronized(o){
...
o.wait(); // Wait in the internal queue attached to "o"
...
}
}
public double g(){

..
synchronized(o){
...
o.notifyAll(); // Wake up the internal queue attached to "o"
...
}
}
Notice that we can rewrite completely a synchronized method of the form
public synchronized double h(){

...
}
as an externally unsynchronized method that is internally synchronized on this as follows:
public double h(){

synchronized(this){
...
}
}
103
Also, the anonymous calls wait() and notify()/notifyAll() are actually the normal
object-oriented abbreviations for this.wait() and this.notify()/this.notifyAll().
Actually, wait() throws an exception InterruptedException (that we shall examine
when we look at how to define threads in Java), so more correctly we should encapsulate
calls to wait() (or o.wait()) as follows:
try{
wait();
}
catch (InterruptedException e);
Also, it is a mistake to use wait() or notify()/notifyAll() other than in a synchro-

nized method. This will throw an IllegalMonitorStateException. Similarly, it is an error
to to use o.wait(), o.notify() or o.notifyAll() other than within a block synchronized
on o.
10.6 Java threads

The simplest way to define concurrent threads in Java is to have a class extend the built-in
class Thread. Suppose we have such a class, as follows, in which we define a function run()
as shown:
public class Parallel extends Thread{
private int id;
public Parallel(int i){ // Constructor

id = i;
}
public void run(){

for (int j = 0; j < 100; j++){
System.out.println("My id is "+id);
try{
sleep(1000); // Go to sleep for 1000 ms
}
catch(InterruptedException e){}
}
}
}
Then, we can created objects of type Parallel and start them off in parallel, as follows:
104
public class TestParallel {
Parallel p[] = new Parallel[5];
for (int i = 0; i < 5; i++){

p[i] = new Parallel(i);
p[i].start(); // Start off p[i].run() in concurrent thread
}
The call p[i].start() initiates the function p[i].run() in a separate thread. (Note
that if we write p[i].run() instead of p[i].start() we just execute p[i].run() like any
other function in the current threadthat is, control is transferred to p[i].run() and when
it finishes, control passes back to the caller.)
Notice the function sleep(..) used in run(). This is a static function defined in Thread
that puts the current thread to sleep for the number of milliseconds specified in its argument.
Any thread can put itself to sleep, not just one that extends Thread. In general, one would
have to write Thread.sleep(..) if the current class does not extend Thread. Observe that
sleep(..) throws InterruptedException (like wait(), discussed earlier).
Of course, given the single inheritance mechanism of Java, we cannot always extend
Thread directly. An alternative is to implement the interface Runnable. A class that imple-
ments Runnable is defined in the same way as one that extends Threadwe have to define
a function public void run()... However, to invoke this in parallel, we have to explicitly
create a Thread from the Runnable object by passing the reference to the object to the
Thread constructor. Here is how we would rework the earlier example.
First the Runnable class:
public class Parallel implements Runnable{ // only this line

// has changed
private int id;
public Parallel(int i){ ... } // Constructor
public void run(){ ... }
Then, the class that uses the Runnable class.
public class TestParallel {
105
Parallel p[] = new Parallel[5];
Thread t[] = new Thread[5];
for (int i = 0; i < 5; i++){

p[i] = new Parallel(i);
t[i] = new Thread(p[i]); // Make a thread t[i] from p[i]
t[i].start(); // Start off p[i].run() in concurrent thread
// Note: t[i].start(), not p[i].start()
}
A thread can be in one of four states:
New: When it has been created but not start()ed.
Runnable: When it has been start()ed and is available to be scheduled.

A Runnable thread need not be runningit may be waiting to be scheduled. No
guarantee is made about how scheduling is done but almost all Java implementations
now use time-slicing across running threads.
Blocked: The thread is not available to run at the moment. This could happen for
three reasons:
1. The thread is in the middle of a sleep(..). It will then get unblocked when the
sleep timer expires.
2. The thread has suspended itself using a wait(). It will get unblocked by a
notify() or notfifyAll().
3. The thread is blocked on some input/output. It gets unblocked when the i/o
succeeds.
Dead: This state is reached when the thread terminates.
10.7 Interrupts
A thread can be interrupted by another thread using the function interrupt(). Thus, in
our earlier example, we can write
p[i].interrupt();
106
to interrupt thread p[i]. This raises an InterruptedException, which, as we saw, we
must catch if the thread is in sleep() or wait().
However, if the thread is actually running when the interrupt arrives, no exception is
raised! To identify that an interrupt arrived in such a situation, the thread can check its
interrupt status using the function interrupted(), which returns a boolean and clears the
interrupt status back to false. Thus, if we want to trap interrupts in p[i].run() both at
the sleep(..) and elsewhere, we should rewrite run() as follows:
public void run(){

try{
j = 0;
while(!interrupted() && j < 100){
System.out.println("My id is "+id);
sleep(1000); // Go to sleep for 1000 ms
j++;
}
catch(InterruptedException e){}
}
}
}
It is also possible to check another threads interrupt status using isInterrupted().

Thus t.isInterrupted() returns a boolean. However, it does not clear the interrupt flag,
unlike interrupted(). Actually, interrupted() is a static function that makes use of
isInterrupted() to check the flag of the current thread and then clears its interrupt flag.
Another useful function to spy on the state of another thread is isAlive(). t.isAlive()
returns true if the thread t is Runnable or Blocked and false if t is New or Dead. Notice
that you cannot use isAlive() to distinguish whether a thread is Runnable or Blocked.
For historical reasons, Java also supports a method stop() to kill another thread and
suspend() and resume() to unilaterally suspend and resume a thread (which is not the same
as the wait() and notify()/notifyAll() construction). However, these are not supposed
to be used because they frequently lead to unpredictable conditions (when a thread dies
abruptly) or deadlocks (when suspended threads are not properly resumed).
107
108
Part VI
Functional programming
109
Chapter 11
Functional programming in Haskell
Haskell is a functional programming language with a strong notion of type (like ML, and
unlike the family of languages based on Lisp). Haskell differs from ML in the strategy it uses
for evaluating functional expressionsas we shall see, Haskell uses an outside-in, lazy
approach while ML uses an inside-out, eager approach.
11.1 Types
Haskell has basic built-in scalar types Int, Float, Bool and Char. In addition, given a type
T, you can form a list of elements of type T. The list has type [T]. Also, given types T1, T2,
. . . , Tk, you can form k-tuples of type (T1,T2,...,Tk).
Functions have typesa function whose input type is T1 and output type is T2 has type
T1 -> T2. For instance, the factorial function has type
factorial :: Int -> Int
Note the syntax :: to denote the type of a function.

What about a function such as max, that computes the maximum of two integer inputs?
A natural definition of its type would be
max :: (Int x Int) -> Int
where (Int x Int) denotes the set of all pairs of type Int. In general a function of n
arguments would have type (T1 x T2 x ... x Tn) -> T where Ti is the type of argument
i and T is the return type. However, for this we have to record with each function its arity
that is, the number of arguments that it takes.
A convenient way to bypass this is to assume that each function takes only one argument.
In the case of an n-argument function, the return value after consuming the first argument
is a function of n1 arguments.
For instance, we rewrite the type of max as
max :: Int -> (Int -> Int)
111
This says that max takes as argument an Int and returns a function Int->Int. If we
write
max 2 3
then (max 2) returns a function that we may think of as max_of_2_and where
max_of_2_and :: Int -> Int
max_of_2_and n = n, if n > 2
2, otherwise
In other words, we take a two argument function, freeze one argument and generate a
new one argument function that reads only the second argument and has the frozen first
argument built in to the definition.
To take another example, the function plus that adds two integers would now become
plus :: Int -> (Int -> Int)
and plus 2 3 would work as follows. (plus 2) returns a function plus_2 where
plus_2 :: Int -> Int
plus_2 n = 2+n
Given this convention, we write a function f(x1,x2,...,xn) as f x1 x2 ... xn. This

expression is implicitly bracketed from the left, so we have
f x1 x2 ... xn = (((f x1) x2) ... xn)
Let T1, T2, . . . , Tn be the types of the n arguments of f and let T be the type of the
return value. Then, rather than writing the type of f as (T1 x T2 x ... x Tn) -> T, we
write
f :: T1 -> T2 -> ... -> Tn -> T
To be consistent with the implicit left bracketing of f x1 .. xn, we have to introduce

brackets in the type expression from the right, as follows:
f :: T1 -> (T2 -> ... -> (Tn -> T))
This convention by which all functions are regarding as functions of a single argument is
called currying, after the logician Haskell Curry, who has also lent his name to the Haskell
language.1
1
However, this idea was actually proposed by another logician called Schonfinkel.
112
11.2 Defining functions in Haskell
The most basic way of defining a function in Haskell is to declare what it does. For
example, we can write:
double :: Int -> Int
double n = 2*n
Here, the first line specifies the type of the function and the second line tells us how the
output of double depends on its input. We shall see soon that the definition of double
is computed by treating the equality as a rule for rewriting expressions. We shall also see
that, in Haskell, the type of a function can be inferred automatically, so we can omit the
type when defining a function.
We are not restricted to having single line definitions for functions. We can use multiple
definitions combined with implicit pattern matching. For instance consider the function:
power :: Float -> Int -> Float
power x 0 = 1.0
power x n = x * (power x (n-1))
Here, the first equation is used if the second argument to power is 0. If the second argument
is not 0, the first definition does not match, so we proceed to the second definition. When
multiple definitions are provided, they are scanned in order from top to bottom.
Here is another example of a function specified via multiple definitions, using pattern
matching.
xor :: Bool -> Bool -> Bool
xor True True = False
xor False False = False
xor x y = True
Here, the first two lines explicitly describe two interesting patterns and the last line catches
all combinations that do not match.
Another way to provide multiple definitions is to use conditional guards. For example:
max :: Int -> Int -> Int
max i j | (i >= j) = i
| (i < j) = j
In this definition the vertical bar indicates a choice of definitions and each definition is
preceded by a boolean condition that must be satisfied for that line to have effect. The
boolean guards need not be exhaustive or mutually exclusive. If no guards are true, none of
the definitions are used. If more than one guard is true, the earliest one is used.
It is important to note that all variables used in patterns are substituted independently
we cannot directly match arguments in the pattern by using the same variable for two
arguments to implicitly check that they are the same. For instance, the following will not
work.
113
isequal :: Int -> Int -> Bool
isequal x x = True
isequal y z = False
Instead, we must write

isequal y z | (y == z) = True
| (y /= z) = False
or, more succinctly,

isequal y z = (y == z)
When using conditional guards, the special guard otherwise can be used as a default
value if all other guards fail, as shown in the following example.
max3 i j k | (i >= j) && (i >= k) = i

| (j >= k) = j
| otherwise = k
Basic types can be combined into n-tuplesfor instance:
-- (x,y) :: (Float,Float) represents a point in 2D

distance :: (Float,Float) -> (Float,Float) -> Float
distance (x1,y1) (x2,y2) = sqrt((x2-x1)*(x2-x1) +
(y2-y1)*(y2-y1))
The first line in the definition above is a comment. Observe the use of pattern matching to
directly extract the components of the tuples being passed to distance.
Auxiliary functions can be locally defined using let or where. The two notations are
quite similar in effect, except that let can be nested while where is only allowed at the top
level of a function definition. Formally, let forms part of the syntax of Haskell expressions
while where is part of the syntax of function declarations. A definition using where can
be used across multiple guarded options. Each choice in a guarded set of options is an
independent Haskell expression and a definition using let is restricted to the expression in
which it occurs, so we have to use a separate let for each guarded clause.
distance (x1,y1) (x2,y2) =

let xdistance = x2 - x1
ydistance = y2 - y1
sqr z = z*z in
in sqrt((sqr xdistance) + (sqr ydistance))
114
distance (x1,y1) (x2,y2) = sqrt((sqr xdistance) + (sqr ydistance))
where
xdistance = x2 - x1
ydistance = y2 - y1
sqr z = z*z
11.3 Functions as first class objects

As we have seen, because of currying, a function typically has a return value that is itself a
function (i.e., the value returned has a type of the form X -> Y, where X and Y may themselves
be further function types). Symmetrically, there is nothing to stop us from functions whose
input type is a function type.
Thus, there is no distinction between basic types such as Int, Bool, etc and function
types as far as input and output types of functions are concerned. In the literature, this
property is often referred to as having functions as first class objects. Contrast this to a
language like C or Java where there is no natural way to pass a function as an argument or
as a return value.
A function that takes another function as input is sometimes called a higher order
function. Here is a simple higher order function:
apply :: (Float -> Int) -> Float -> Int

apply f y = (f y)
Thus, apply takes two arguments and applies its first argument to its second argument.
Notice that the types of f and y have to be compatible to permit this.
11.4 Polymorphism
The apply function defined earlier specifically requires the first argument to be a function
with type Float -> Int and, hence, the second argument to be of type Float.
It would be more desirable to be able to write a version of apply whose first argument
has an arbitrary function type T -> T and whose second argument is of a matching type
T, so that apply itself has the type (T -> T) -> T -> T.
This is an example of polymorphism. The word means taking multiple forms and refers
to a function that operates similarly on multiple types. An even more basic example of a
polymorphic function is the identity function that just returns its argument. This function
should have type T -> T, for any type T.
Other examples of polymorphism are familiar functions on lists. For instance, reverse
reorders a list of any type, so reverse has type [T] -> [T] for all types T (recall that for a
type T, the list of that type has type [T]). Similarly, length returns the length of any list,
so length should have type [T] -> Int.
115
This kind of polymorphism is sometimes referred to as structural polymorphism be-
cause of the way it is reflected in functions such as reverse and length, where the function
operates on the structure of the collection rather than on the base type.
One must be careful to distinguish this kind of polymorphism from the ad hoc variety
associated with overloading operators. For instance, in most programming languages, we
write + to denote addition for both integers and floating point numbers. However, since the
underlying representations used for the two kinds of numbers are completely different, we
are actually using the same name (+, in this case) to designate functions that are computed
in a different manner for different base types.
Another example of ad hoc polymorphism is that available in Java, where different meth-
ods can have the same name but different parameter types. Thus, for instance, the class
Arrays seemingly supports a uniform method called sort to sort arrays of arbitrary scalar
types, but this is just a collection of distinct functions with the same namethe correct one
is picked depending on the type of argument passed to the function.
Coming back to Haskell, we can describe polymorphic types using type variables, typically
denoted a, b, c, . . . . Thus we have the following type definitions for the functions discussed
earlier.
identity :: a -> a
apply :: (a -> b) -> a -> b
length :: [a] -> Int
reverse :: [a] -> [a]
Type variables come with implicit universal quantification. Thus, for instance, we read
the type of apply as for every type a and every type b, (a -> b) -> a -> b. Type
variables are uniformly substituted by real types to get a specific instance of the function.
Thus, for instance, when we define a function
plus2 :: Int -> Int

plus2 n = 2+n
and then write
apply plus2 5
this first instantiates both the variables a and b to type Int and uses the concrete definition
apply :: (Int -> Int) -> Int -> Int
to do the job.
116
11.5 Conditional polymorphism
What would be the type of a sort function? The definition
quicksort :: [a] -> [a]
is not accurate, because we cannot sort any list. We need to have a way of unambiguously
comparing items in the list. This is not possible for arbitrary types because, for instance, a
could be a function type and it is not always possible to effectively compare functions.
Consdier a simpler examplememberof that checks whether the value supplied as its first
argument appears in the list supplied as its second argument. This should be of type
memberof :: a -> [a] -> Bool
For this, we have to be able to check whether two elements of type a are equal. Even
this is problematic for arbitrary typesagain, what if a is a function type? Elementary
computability theory tells us that checking whether two functions are equal (i.e., agree on
all inputs) is undecidable.
To get around this, we have to place restrictions on the type of a. In Haskell, this is done
by organizing the collection of all types into type classes (nothing to do with classes as in
object-oriented programming!) A type class in Haskell is just a subset of all possible types.
For instance, the type class Eq is the set of all types for which equality checking is possible.
We can then write the type of memberof as
memberof :: Eq a => a -> [a] -> Bool
This is to be read as memberof is of type a -> [a] -> Bool for any type a that belongs
to the class Eq.
In Haskell, the type class Ord contains all types on which comparisons are possible. So,
we would have
quicksort :: Ord a => [a] -> [a]
Since comparing involves checking equality as well, clearly every type in Ord is also in
Eq.
Recall that in Java, conditional polymorphism was achieved using interfaces. Uncon-
ditional polymorphism, at least for functions that operate on objects, can be achieved by
defining the argument to be of type Object, which is compatible with all other classes in
Java by the definition of the type hierarchy.
117
11.6 User defined datatypes
Haskell permits user defined datatypes. We can define structured types such as trees, stacks,
queues . . . . To do this, we define new names, called constructors, to denote the components
of the new type. In general, a constructor takes arguments. The simplest type of constructor
is a constant, one that takes 0 arguments.
The simplest user defined datatype is an enumerated type, such as
data day = Sun | Mon | Tue | Wed | Thu | Fri | Sat
More interesting are recursive types, such as a binary tree where each node can hold an
integer.
data Btreeint = Nil | Node Int Btreeint Btreeint
This says that a binary tree of integers, Btreeint, is either an empty tree, Nil, or a node
which holds an integer value and contains two nested trees, corresponding to the left and
right children.
Can we define polymorphic datatypes, like polymorphic functions? For instance, can we
define just the structure of a binary tree without fixing the type of value stored at each node?
Yes, we can, using the notion of a type variable, as before.
data Btree a = Nil | Node a (Btree a) (Btree a)
Note here that the complete name of the new type is Btree a not just Btree.
To take another example, we can redefine the builtin type list as follows:
data List a = Nil | Cons a (List a)
The next step is to declare that a user defined datatype belongs to a type class such as
Eq or Ord. Each type class is characterized by a set of functions that each member of the
class must supportanalogous to an interface in Java. For instance, for a type to belong
to Eq, it must support an equality check, denoted ==. The exact implementation of == is
irrelevant. For instance, we might define == on trees to be isomorphism and place Btreeint
in Eq as follows.
instance Eq Btreeint where

Nil == Nil = True
(Node x t1 t2) == (Node y u1 u2) = (x == y) && (t1 == u1)
&& (t2 == u2)
The definition of == for Btreeint follows the normal syntax of a Haskell function
definitionin this example, by cases.
How about making Btree a an element of Eq? We could analogously write
118
instance Eq (Btree a) where
Nil == Nil = True
&& (t2 == u2)
Notice however, that this uses the test x == y for elements of the unspecified type a. In
Btreeint this was not a problem because the type of x and y was known to be Int, which
belongs to Eq. Here, instead, we must make the membership of Btree a in Eq conditional
on the properties of the underlying type a, as follows.
instance Eq a => Eq (Btree a) where

Nil == Nil = True
&& (t2 == u2)
In other words, if a is in Eq, then Btree a is in Eq with == defined as . . .

We have to be a bit careful here. With the definitions we have given, the Haskell inter-
preter responds as follows,
Main> (Node 5 Nil Nil) == (Node 6 Nil Nil)

False
Main> (Node 5 Nil Nil) == (Node 5 Nil Nil)
True
which is as we expect, but, somewhat unexpectedly, it says
Main> Nil == Nil

ERROR - Unresolved overloading
*** Type : Eq a => Bool
*** Expression : Nil == Nil
This is because if we just say Nil, the interpreter cannot instantiate a to a concrete type
and determine whether it belongs to Eq or not. The same happens if we try to ask whether
[] == [] for the empty list that is built in to Haskell. To get around this, we have to
explicitly supply information about the base type of the tree or list as follows.
Main> (Nil::Btree Int) == Nil

True
Main> ([]::[Float]) == []
False
Notice that it is sufficient to specify the base type for [] on one side of the equation.
Interestingly, even if we try, we cannot assert that there are two different instances of [] in
an expression. For instance, if we try the following
119
Prelude> ([]::[Float]) == ([]::[Int])
the response is
ERROR - Type error in application

*** Expression : [] == []
*** Term : []
*** Type : [Float]
*** Does not match : [Int]
11.7 Rewriting as a model of computation

The model of computation that underlies Haskell is rewriting. In a sense, rewriting is at
the heart of all formulations of computabilityany computational devices fundamental
behaviour is to read symbols and generate other symbols, which amounts to rewriting the
input as the output.
Function definitions in Haskell provide the rules for rewriting. Normally, rewriting re-
duces the complexity of an expression, hence reduction is a synonym for rewriting. Here,
complexity is a semantic notion associated with the expression, not a syntactic notion. The
expression that results after reduction is often longer and more complex, syntactically,
than the orignal expression! In the literature, the term rewriting system is usually used to
signify a set of rules that may be applied in either direction, while reduction refers to a
rewriting process that uses the rules in one direction only.
For instance, given the function definition
square: Int -> Int

square n = n*n
we have square 7 ;7*7 ;49, where the symbol ; denotes rewrites to.
The first reduction is user specified, the second is built in.
Reduction continues till a result is obtained. A result is just an expression that cannot
be rewritten further. In Haskell, results are normally constants that look sensiblevalues of
type Int, Float, . . . . However, as we shall see in the lambda calculus, this may not always
be the casewe may have complicated expressions that cannot be further reduced.
What is the reduction sequence for square (4+3)? We have more than one possibility:
square (4+3) ;square 7 ;7*7 ;49
square (4+3) ;(4+3) * (4+3) ;7 * (4+3) ;7*7 ;49
square (4+3) ;(4+3) * (4+3) ;(4+3) * 7 ;7*7 ;49
Is it a coincidence that all three sequences yield the same result? No. In fact, we have
the following theorem.
120
Theorem Let e be a Haskell expression with two reductions e ; f1 ; f2 ; ; fk ; g
and e ; f10 ; f20 ; ; f`0 ; g 0 such that g and g 0 cannot be further reduced. Then,
g = g0.
In general, we need a strategy for reductionthat is, a rule telling us how to apply the
rules! Some questions we might ask are:
1. Are all reduction strategies equivalent with respect to the final result (this property is
called confluence)?
2. Can we tell whether the reduction process will terminate? For example, consider what
happens to the expression factorial -1 with respect to the definition
factorial:: Int -> Int

factorial n = n * (factorial (n-1))
3. Given a choice of reduction strategies, can we always identify the optimal one, in terms
of efficiency? Notice in the reduction of square (4+3) that the first reduction sequence
is shorter than the other two.
11.8 Reduction strategies

We can associate a notion of nesting with Haskell expressions. For instance, if we have
something like f (g x), then the expression (g x) is nested deeper than the expression f.
Thus it makes sense to talk of outer and inner expressions as candidates for reduction.
Two canonical reduction strategies are
Outermost = lazy = do not evaluate arguments till you need to
Innermost = eager = evaluate arguments before evaluating function
The first reduction of square (4+3) above was computed eagerly, while the second and
third were lazy. Haskell follows the strategy of outermost, or lazy, reduction, while ML
follows the strategy of innermost, or eager, reduction.
Here is another example:
power :: Float -> Int -> Float
power x n | (n==0) = 1
| otherwise = x * (power x (n-1))
Consider what happens when we evaluate power (7.0/0) 0.
With outermost reduction, we first try to reduce power. The first rule applies and we get
power (7.0/0) 0 ;1. On the other hand, innermost reduction will try to evaluate (7.0/0)
and generate an error.
Thus, the two reduction strategies are not always equivalent. We have the following
general result.
121
Theorem The outermost strategy for reduction terminates with a value whenever the
innermost strategy does.
The example of power shows that there are expressions where the innermost strategy
does not terminate with a value while the outermost strategy does. We should not assume
that values are always sensible. It is debatable, mathematically, whether power (7.0/0)
0 should evaluate to 1. All we are discussing is the relative power of the two reduction
strategies.
The reduction we saw for square (4+3) suggests that outermost reduction involves
wasteful recomputationin this case, (4+3) is evaluated twice. In fact, Haskell uses an
optimization of outermost reduction called graph reduction. Graph reduction maintains only
one copy of each unreduced expression and a pointer to this copy from each occurrence in
the main expression. Thus, if the same subexpression occurs multiple times in an expression,
the expression will have multiple pointers to the location containing the subexpression. The
first time this subexpression is encountered, it will be reduced. Whenever the subexpression
is encountered subsequently, the pointer will reveal that it has already been reduced.
Theorem Outermost graph reduction is at least as efficient as innermost reduction.
In fact, in some situations outermost reduction can be significantly more efficient, because
it will terminate without performing unnecessary computations. Consider the following
definition of a binary tree which stores values only at the leaves.
data BTree a = Nil | Leaf a | Node BTree a BTree a
We would like to compare the list of values stored at the frontiers of such trees. Let us
define a function that checks whether these lists are the same for two trees.
equaltree :: BTree a -> BTree a -> Bool

equaltree t1 t2 = ((leaflist t1) == (leaflist t2))
leaflist :: BTree a -> [a]

leaflist Nil = []
leaflist (Leaf x) = [x]
leaflist (Node t1 t2) = (leaflist t1) ++ (leaflist t2)
Checking list equality in Haskell is lazy and terminates when the first pair of nonequal ele-
ments is encountered, so the Haskell version of this function will actually construct leaflist
t1 and leaflist t2 only to the extent required to show that they are not the same. Of
course, if they are the same, both lists must be constructed in entirety. In an eager functional
language (like ML), a similar function would first compute leaflist t1 and leaflist t2
in entirety and only then check them for equality, which is more expensive in general.
122
11.9 Outermost reduction and infinite data structures
Outermost reduction permits the definition of infinite data structures. For instance, the list
of all integers starting at n is given by the function
listfrom n = n: (listfrom (n+1))
The output of listfrom m is the infinite list [m, m+1, ... which is denoted [m..] in
Haskell.
We can use infinite lists, for instance, to define in a natural way the Sieve of Eratosthenes
whose output is the (infinite) list of all prime numbers.
sieve (x:xs) = x : (sieve [ y <- xs | mod y x != 0])

primes = sieve [2..]
The function sieve picks up the first number of its input list, constructs a new list from
its input by removing all multiples of the first number and then recursively runs sieve on
this list.
If we work out the reduction for this we get
primes ; sieve [2..]

; 2 : (sieve [ y | y <- [3..] , mod y 2 != 0])
; 2 : (sieve (3 : [y | y <- [4..], mod y 2 != 0])
; 2 : (3 :
(sieve [z | z <- (sieve [y | y <- [4..], mod y 2 != 0])
| mod z 3 != 0])
; 2 : (3 : (5 :
(sieve [w | w <- (sieve [z | z <- (sieve [y | y <- [4..],
mod y 2 != 0]) | mod z 3 != 0]) | mod w 5 != 0])
; ...
Why is this useful? It is often conceptually easier to define a function that returns an
infinite list and extract a finite prefix to get a concrete value.
For instance, from primes we can derive functions such as
nthprime k = primes !! k
that extracts the kth prime number.
123
124
Chapter 12
The (untyped) lambda calculus
The lambda calculus was invented by Alonzo Church around 1930 as an attempt to formalize
a notation for computable functions. It is important to have some basic familiarity with the
lambda calculus because it forms the foundation for the definition of functional programming
languages such as Haskell.
In the conventional set theoretic formulation of mathematics, a function f can be repre-
sented by its graphthat is, a binary relation Rf between the sets domain(f ) and codomain(f )
consisting of all pairs (x, f (x)). Not all binary relations represent functionsfunctions are
single-valued, so if (x, y) Rf and (x, y 0 ) Rf then we must have y = y 0 . In mathematics,
functions are almost always total, so we also have an additional requirement that for each x
in domain(f ), there exists a pair (x, y) Rf . Representing a function in terms of its graph
is called an extensional definitiontwo functions are equal if and only if their graphs are
equal. This is analogous to the definition of equality in set theorytwo sets are equal if and
only if they have the same elements.
From a computational viewpoint, it is not sensible to represent functions merely in terms
of graphs. Clearly, all sorting algorithms define functions that have the same graph. However,
from a computational viewpoint, we would like to distinguish efficient sorting algorithms from
inefficient ones.
What we need is an intensional representation of functionsa representation that cap-
tures not just the output value for each input, but also how this value is arrived at.
With this background, we plunge into the (pure untyped) lambda calculus.
12.1 Syntax
The set of lambda expressions is generated by what we would, in modern day computer
science terminology, call a context-free grammar.
Let Var be a countably infinite set of variables. We define the set of lambda expressions
as follows:
= x | x.M | M M 0
125
where x V ar and M, M 0 .
Thus, the syntax of lambda expressions contains two ways to generate new expressions
from old ones. The construction x.M is called abstraction, while the construction M M 0 is
called application.
The intention is that x.M is an expression representing a function of x whose body (or
rule for computation) is given by M . Thus, x.M abstracts the computation rule described
in M to operate on arbitrary input values x. This is analogous to writing f (x) = M without
having to assign an unnecessary name f to the function (after all, f (x) = M and g(x) = M
are clearly the same function, so the names f and g carry no significance.)
The expression M M 0 is intended to represent the action of applying the function en-
coded by M to the argument M 0 .
12.2 Life without types

At this point, we should remind ourselves that these intended meanings do not stop us from
constructing seemingly nonsensical expressions such as xx.
At first sight, it seems strange that we should try to embark on a syntax for functions in
which there is no notion of type. However, untyped formalisms are familiar to us
Set theory is untyped. Everything is a set. To do arithmetic, we define certain sets to
denote numbers. Functions on numbers apply in principle to all sets but are mean-
ingless in the wrong context.
A computers memory is intrinsically untyped. Consider the operation of dereferencing
a pointer in Cthat is, treat a memory location as a pointer and look up the location
that it points to. This operation is performed without checking if the contents of the
pointer are of the right typethat is, a memory location legally available to this
program. If the contents of the pointer are valid, the operation succeeds. However, if
the pointer is invalid (for instance, if it is uninitialized), the operation fails (sometimes
dramatically, as we all know!)
The point is that we have a syntax for representing data (sets, bit strings) and we have
rules for manipulating these representations. Some representations are more meaningful than
others, but the manipulation rules can be applied uniformly to both the meaningful and the
meaningless representations. All we ask is that if the original data represents a meaningful
value, then, after manipulation, it still represents a meaningful value.
A similar view is taken in the lambda calculusfor instance M M 0 represents application
if the structure of the expression M corresponds to a function. Otherwise, we dont care
what this construction means (e.g., if M M 0 is xx, as describe earlier).
12.3 The rule

The basic rule for computing with the lambda calculus is called and is given by:
126
(x.M )M 0 M {x M 0 }
where M {x M 0 } represents the effect of substituting all free occurrences of x in M
uniformly by M 0 . (We shall formally define what we mean by a free occurrence of x shortly.)
The rule is one that we use unconciously all the time: If f (x) = 2x2 + 3x + 4, then
f (7) = 2 72 + 3 7 + 4 which is just 2x2 + 3x + 4{x 7}.
From the rule it is clear that an application M M 0 is meaningful only if M is of the
form x.M . Thus, though we can write terms like xx, we cannot do much with them.
Effectively, the rule is the only rule we need to set up the lambda calculus. For the
moment, think of it as a rewriting rule for expressions that can be applied to any subterm
in a lambda expression. We will formalize this later.
12.4 Variable capture

An important issue to keep in mind when applying the rule is that the substitution for x
should not capture variables. Consider, for instance, the application (x.(y.xy))y. Using
the rule naively, we get (x.(y.xy))y y.yy However, the y that comes in to y.xy
by the substitution for x is a new y and should not be confused with the y that is used to
define the inner function. Thus, we should redefine the inner function with a fresh variable
before making the substitution, as follows:
(x.(y.xy))y = (x.(z.xz))y z.yz

Clearly, changing the name of the variable defining a function does not affect the definition
there is no difference between the functions f (x) = 2x + 5 and f (z) = 2z + 5.
To formalize this notion we have to define free and bound variables. Intuitively, a variable
is free in an expression until it is bound by an abstraction. Thus, in the expression xy,
both x and y are free. If we abstract y as y.xy, x remains free but y becomes bound. More
formally, we can define two sets F V (M ) (free variables of M ) and BV (M ) (bound variables
of M ) inductively, as follows:
F V (x) = {x}, for any variable x BV (x) = , for any variable x
F V (x.M ) = F V (M ) {x} BV (x.M ) = BV (M ) {x}
F V (M M 0 ) = F V (M ) F V (M 0 ) BV (M M 0 ) = BV (M ) BV (M 0 )
When we apply the rule to transform x.M M 0 into M {x M 0 }, we need to check

that no free variable in M 0 clashes with a bound variable in M . In the example above,
(x.(y.xy))y y.yy,
y is bound in the expression M = y.xy, while y is free in the expression M 0 = y.
127
In the literature, several approaches have been proposed to deal with the variable renam-
ing problem that is required for the rule to work properly. For simplicity, we shall just
assume that we always rename the bound variables in M to avoid capturing free variables
from M 0 .
12.5 Encoding arithmetic

Having set up this rather abstract notation for computable functions, let us see an example
of how to use it.
Before proceeding, we set up some conventions. Implicitly, lambda expressions associate
to the left. Thus, the expresssion M1 M2 M3 denotes ((M1 M2 )M3 ). Also, if we have an
abstraction with multiple variables at the outermost level, such as x.(y.xy), we can pull
out all the variables at the outermost level into a single sequence such as xy.xy.
In set theory, the number n is identified, in a sense, with a set that has nesting level
of depth n. For instance, we start with the emptyset for representing the number 0 and,
inductively, use the set {0, 1, .., n 1} to represent the number n. Thus we have
0 =
1 = {}
2 = {, {}}
3 = {, {}, {, {}}}
...
In the lambda calculus, we encode the number n by the number of times we apply
a function to an argument. Thus, the number n is a lambda expression that takes two
arguments and applies the first argument to the second argument n times. Let hni be an
abbreviation for the lambda term denoting n. Then, we have:
h0i = f x.x
hn + 1i = f x.f (hnif x)
This encoding was proposed by Church and terms of the form hni are hence called the
Church numerals. Concretely, we have
h1i = f x.f (h0if x) = f x.(f ((f x.x)f x))

Note that h0igy (x.x)y y. Thus h0if x x. So, inside the definition of h1i,
we get
h1i = . . . = f x.(f f x.x)f x)) f x.(f x)

| {z }
apply
Since h1igy (x.(gx))y gy, we have
128
h2i = f x.f (h1if x) = f x.(f (f x.(f x)f x)) f x.(f (f x))
| {z }
apply
and
h2igy x.(g(gx))y = g(gy)

Let us write g k y to denote the term g(g(. . . (gy))) with k applications of g to y. Then, it
is not difficult to show inductively that
hni = f x.f (hn1if x) . . . f x.(f n x)

We can now define arithmetic functions such as successor, plus, . . . . For instance, suc-
cessor is just pf x.f (pf x). To see how this works, we observe that
(pf x.f (pf x))hni f x.f (hnif x) f x.f (f n x) = f x.f n+1 x = hn+1i
Thus, if the function representing successor is applied a term of the right type, it yields
the expected answer. We could also apply it to a meaningless term, in which case we get
a meaningless answer!
Here is a definition for plus: pqf x.pf (qf x). We can verify that
(pqf x.pf (qf x))hmihni (qf x.hmif (qf x))hni

(f x.hmif (hnif x))
(f x.hmif (f n x))
(f x.f m (f n x))
= (f x.f m+n x) = hm+ni
Similarly, we define multiplication and exponentiation as follows:
multiplication : pqf x.q(pf )x

exponentiation : pq.(pq)
12.6 One step reduction

Recall that we introduced the rule
(x.M )M 0 M {x M 0 }
and said, informally, that we would permit this rule to be used in all contexts. The rule
is not the only basic rule possible. For instance, observe that the two terms x.(M x) and
M are equivalent with respect to reductionfor any term M 0 , (x.(M x))M 0 M M 0 .
This can be formalized by a rule (normally called ) which says
x.(M x) M
129
Given a set of basic rules such as , , . . . , we can inductively define a one step reduction
that permits any of these basic rules to be used in any context. Let us denote one step
reduction by . We define through inference rules such as the following:
if M x M 0 for some basic rule x such as , , . . . , then M M 0
Following the conventional notation used in logic, we present such a rule in the following
form
M x M 0
M M0
Here is the complete set of rules defining :
M x M 0 0 M M0 M M0 N N0
M M x.M x.M 0 M N M 0 N M N M N 0
Notice that all that we have done is to formalize the fact that we permit the basic
reductions x within any subterm. In the calculations we have seen so far, we have
already informally used this form of applying the rule.
In the discussion that follows, we shall dispense with the rule and assume that the only
basic rule is the rule .
12.7 Normal forms

As we had discussed in the context of Haskell, we are interested in terms that cannot be
reduced. These denote values and are called normal forms.
Formally, we are interested in properties of the relation , the reflexive and transitive
closure of . We have already anticipated some of the questions we would like to ask about
in our discussion of Haskell.
Does every term reduce to a normal form?
Can a term reduce to more than one normal form, depending on the order in which
we perform reductions?
If a term has a normal form, can we always find it?
Consider the term (x.xx)(x.xx). Only one reduction step is possible and this makes
two copies of the argument and returns us to the original term (x.xx)(x.xx). The reduction
process for this term never terminates, so this is an example of a lambda term without a
normal form. In the literature, the term (x.xx)(x.xx) is usually denoted .
Consider now the term (yz.z) that ignores its first argument. If we apply (yz.z) to ,
we have two possible reductions.
(yz.z)((x.xx)(x.xx)) z.z, using outermost reduction.
130
(yz.z)((x.xx)(x.xx)) (yz.z)((x.xx)(x.xx)), using innermost reduction.
Clearly, if we keep making the second choice, we will never reach a normal form, even
though it is always possible to reach a normal form in one step by choosing the first reduction.
To answer the question of multiple normal forms, we will state a more general result
shortly.
First, we define a notion of equivalence based on .
12.8 -equivalence
We define an equivalence on lambda terms, which we shall write , based on the many step
reduction relation . Basically, is the symmetric transitive closure of . It can be
defined inductively using inference rules in the same way that we defined based on x :
M N M N M N, N P
M N N M M P
We can do the same with any reflexive, transitive relation R the symmetric, transitive
R
closure of R defines an equivalence relation .
12.9 Church-Rosser property

Let R be a binary relation that is reflexive and transitive. We say that R has the diamond
property if, whenever X R Y and X R Z, then there is a W such that Y R W and Z R W .
If a relation R has the diamond property, we say that R is Church-Rosser.
R
Theorem (Church-Rosser) Let R be Church-Rosser. Then M N implies there exists
Z, M R Z and N R Z
R
Proof By induction on the definition of .
R
(i) M N because M R N . Then chooze Z N .
R R
(ii) M N because N M . Trivial.
R R R
(iii) M N because M P and P N . By the induction hypothesis, we have Y such
that M R Y and P R Y and Z such that P R Z and N R Z.
Since P R Y , P R Z and R is Church-Rosser, there is a term W such that Y R W
and Z R W .
But then M R W and N R W .
2
We have the following useful Corollary:
131
Corollary Let R be Church-Rosser. Then a term can have at most one R-normal form
(where an R-normal form for M is, as expected, a term N such that M R N and there is
no P such that N R P .)
R R
Proof Suppose N1 and N2 are both normal forms for M . Then, N1 M N2 , so
R
N1 N2 . There must then be a common Z such that N1 R Z and N2 R Z. Since neither
N1 nor N2 can be reduced, it follows that N 1 N2 Z. 2
Recall that our goal is to check whether normal forms in the lambda calculus are unique.
One way to do this is to show that has the Church-Rosser property.
For any relation , if has the Church-Rosser property, then so does . The proof
of this is easy to see pictorially.
M N1

..
.

..
.
N2
If M N1 (the horizontal axis) and M N2 (along the vertical axis), we can break up
both reductions into a sequence of individual steps. From the fact that is Church-Rosser,
we can complete the diamond for each single step in the sequence. In the end, we get a
complete rectangle between N1 and N2 and the corner opposite M is the common term that
we can obtain from N1 and N2 .
Unfortunately, does not have the Church-Rosser property, so we cannot use this strat-
egy to establish that is Church-Rosser! Consider the term (x.xx)((x.x)(x.x)). We
have two reductions possible:
(x.xx)((x.x)(x.x)) ((x.x)(x.x))((x.x)(x.x)) (Outermost)
(x.xx)((x.x)(x.x)) ((x.xx)(x.x) (Innermost)
If we take the second option, we can then in one step get
(x.xx)(x.x) ((x.x)(x.x))
We can reach this term from the first option as well, but it requires two steps!
The solution lies in defining an alternative notion of one step reduction from the basic
beta reduction such that
(i) This new reduction is Church-Rosser.
(ii) Its reflexive, transitive closure is equal to .
132
We define this new reduction as follows.
M M0 M M 0, N N 0 M M 0, N N 0
M M
x.M x.M 0 M N M 0N 0 (x.M )N M 0 {x N 0 }
The last inference rule builds in reduction. The crux of the relation is that it
combines nonoverlapping reductions in one parallel step. For instance, if both M and N
can be reduced to M 0 and N 0 , respectively, then M N can be reduced to M 0 N 0 in one step.
We claim, without proof, that is Church-Rosser and that is the same as .
Assuming the claim, it then follows that is Church-Rosser and so the normal form for a
term in the lambda calculus is unique, if it exists.
12.10 Computability
We now argue that all computable functions can be described in the lambda calculus. As
is usual in computability theory, we restrict our attention to functions on natural numbers.
Recall that we can encode natural numbers in the lambda calculus via the Church numerals
hni f x.f n x. Our goal now is to show that for every computable function f : Nk N we
can find a lambda term hf i that encodes f so that for every k-tuple n1 , n2 , . . . , nk
hf ihn1 ihn2 i . . . hnk i hf (n1 , n2 , . . . , nk )i

In other words, the result of applying the encoding of f to the encodings of the arguments
n1 , n2 , . . . , nk should be the encoding of the result f (n1 , n2 , . . . , nk ). Notice that on the left
hand side we have used currying to break up the k-tuple of arguments to f into a sequence
of k arguments to hf i.
Recursive functions
One of the many formulations of computable functions that arose in the quest to define
effective computability is that of recursive functions, due to Godel. The recursive functions
are defined inductively in terms of a set of initial functions and rules for combining functions,
as follows.
Initial functions
Zero: Z(n) = 0.
Successor : S(n) = n+1.
Projection: ki (n1 , n2 , . . . , nk ) = ni .1
1
Notice that the projection functions are actually form an infinite family, one for each choice of k and
i {1, 2, . . . , k}.
133
Composition
Given f : Nk N and g1 , g2 , . . . , gk : Nh N, the composition of f and g1 , g2 , . . . , gk ,
denoted f (g1 , g2 , . . . , gk ), is the function
f (g1 , g2 , . . . , gk )(n1 , n2 , . . . , nh ) =
f (g1 (n1 , n2 , . . . , nh ), g2 (n1 , n2 , . . . , nh ), . . . , gk (n1 , n2 , . . . , nh ))
For example, the function f (n) = n+2 can be defined as the composition f = S S of the
successor function with itself.
Primitive recursion
Given g : Nk N and h : Nk+2 N, the function f : Nk+1 N is defined by primitive
recursion from g and h if the following equalities hold:
f (0, n1 , n2 , . . . , nk ) = g(n1 , n2 , . . . , nk )
f (n+1, n1 , . . . , nk ) = h(n, f (n, n1 , n2 , . . . , nk ), n1 , . . . , nk )
Here are some examples of primitive recursive definitions.
The function plus(n, m) = n+m can be defined by primitive recursion from g = 11

and h = S 32 . In other words,
plus(0, n) = g(n) = 11 (n) = n
3
plus(m+1, n) = h(m, plus(m, n), n) = S 2 (m, plus(m, n), n) = S(plus(m, n))
The function times(n, m) = n m can be defined by primitive recursion from g = Z
and h = plus (33 , 32 ).
Minimalization
Given g : Nk+1 N, the function f : Nk N is defined by minimalization from g if
f (n1 , n2 , . . . , nk ) = n.(g(n, n1 , n2 . . . , nk ) = 0)
where is the minimalization operator : n.P (n) returns the least natural number n such
that P (n) holds. If P (n) does not hold for any n, then the result is undefined.
In modern algorithmic notation, f can be computed by a while loop of the form
n := 0;
while (g(n,n1,n2,...,nk) != 0) {n := n+1};
return n;
For example, consider the function log2 n defined as follows: log2 n = k, if 2k = n. If n is not
a power of 2, log2 n is undefined. The function log2 can be defined by minimalization from
the function g(k, n) = 2k n.2
2
Technically, we have to be a bit more careful here. Since we are dealing only with functions over N,
134
Recursive functions
The set of recursive functions is the smallest set which contains the initial functions and is
closed with respect to composition, primitive recursion, and minimalization.
If we exclude minimalization from this construction, we get the set of primitive recur-
sive functions, all of which are total functions. Minimalization is the only operator which
introduces partial functions (or, from a computational standpoint, non-termination).
12.11 Encoding recursive functions in lambda calculus

Recall that we have already fixed the following definitions:
hni f x.(f n x).
succ nf x.(f (nf x)) such that succhni hn + 1i.
The function succ is the encoding hS i for the initial function Successor. We can define
encodings for the other initial functions Zero and Projection as follows:
hZ i x.(gy.y).
hik i x1 x2 . . . xk .xi .
Similarly, the composition of functions is easily defined in the lambda calculus. What
remains is to encode primitive recursion and minimalization. We begin by defining encodings
for constructing pairs and projecting out the components of a pair.
pair xyz.(zxy).
fst p.(p(xy.x)).
snd p.(p(xy.y)).
Observe that fst(pair a b) (p.(p(xy.x)))((xyz.(zxy)) a b) (p.(p(xy.x)))

(z.(zab))(xy.x) (xy.x)ab (y.a)b a. Similarly, we can show that snd (pair a b)
b.
We now show how to encode primitive recursion. For notational simplicity, we restrict
ourselves to the case where the function f being defined by primitive recursion has just one
argument. Thus, the definition of primitive recursion becomes
f (0) = g
f (n+1) = h(n, f (n))
mn would be defined to be 0 if m < n, so we have to modify the definition of g(k, n) appropriately to take
care of this.
135
Observe that g is just a constantthat is, some fixed number k. Inductively assume that
h is a function defined by the lambda expression hhiin other words, for all numbers m and
n, hhihmihni hg(m, n)i. The aim is to come up with a lambda expression hf i defining
f , such that for all n, hf ihni hf (n)i.
Consider the function t(n) = (n, f (n)). Then, t can be defined by primitive recursion as
follows:
t(0) = (0, f (0)) = (0, g)

t(n+1) = (n+1, f (n+1)) = (n+1, h(n, f (n)))
= (succ(fst(t(n))), h(fst(t(n)), sec(t(n))))
This is a simpler form of primitive recursion called iterationthink of it as repeatedly
iterating a step function starting with (0, g). (In programming language parlance, this is
equivalent to translating a top-down recursive definition into an equivalent bottom-up iter-
ative one).
The step function is lambda-definable by the following expression:
step x.pair (succ(fst x))(hhi(fst x)(sec x)).

We can check that step (pair hnihf (n)i) pair hn+1ihf (n+1)i.
step (pair hni hf (n)i) pair (succ(fst(pair hni hf (n)i)))

(hhi(fst(pair hni hf (n)i))(sec(pair hni hf (n)i)))
pair (succ hni) (hhihnihf (n)i)
pair (succ hni) hh(n, f (n))i
pair hn+1i hh(n, f (n))i
pair hn+1i hf (n+1)i.
The encoding for t is as follows:
hti y.y step (pair h0i hgi)

We can check that for all n, htihni pair hni hf (n)i.
htih0i h0i step (pair h0i hgi) pair h0i hgi

htihn+1i hn+1i step (pair h0i hgi)
step (hni step (pair h0i hgi))
step (pair hni hf (n)i)(by ind. hyp.)
pair hn+1i hf (n+1)i(by what has been proved above)
Since f (n) is sec(t(n)), hf i is the following expression:
hf i y.sec(htiy)
It is immediate that for all n, hf ihni hf (n)i.
136
Collapsing all the steps we have seen above, we can write down an expression PR that
constructs a primitive recursive definition from the functions g and h:
PR hgy.sec(y(x.pair (succ (fst x)) (h(fst x)) (sec x)))(pair h0i g))
We could also have used recursive definitions of lambda experssions to encode primitive
recursion. Recursive definitions can also be used to define minimalization. It turns out that
every recursive definition in the lambda calculus can be solved by finding its fixed point.
12.12 Fixed points

We begin our discussion of fixed point by providing an encoding for boolean values true (tt)
and false (ff) and conditional expressions. We begin with the conditional. We seek a term
bxy. E that will take three arguments, return x if b (the conditional expression) is true and
y if b is false.
We claim that the expression bxy. bxy with the definitions of tt and ff given by
htti xy.x and hffi xy.y do the job. (Incidentally, note that hffi is the same as h0i).
Clearly
(bxy.bxy)httif g xy.((xy.x)xy)f g
y.((xy.x)f y)g
(xy.x)f g
(y.f )g
f
and, similarly,
(bxy.bxy)hffif g xy.((xy.y)xy)f g
y.((xy.y)f y)g
(xy.y)f g
(y.y)g
g
Thus bxy.bxy gives us an encoding of the construct if-then-else. We are now close
to being able to write recursive function definitions in a syntax similar to languages like
Haskell. For instance, we could aim to write
factorial hni = if (iszero(hni)) then 1

else mult hni(factorial (pred hni))
where, of course, we still have to encode the predicate iszero and the arithmetic function
pred (predecessor).
Assuming we can do this, how do we convert such a recursive definition into a lambda
term?
137
From recursive functional definitions to lambda terms
Let F = x1 x2 . . . xn E be a recursive definition of F where E contains not only the variables
x1 , x2 , . . . , xn and also the expression F . How do we transform such a definition into an
equivalent lambda term?
To do this, we choose a new variable f and convert E to E by replacing every occurrence
of F in E by (f f ). That is, if E is of the form F F then E is (f f ) (f f ) .
Now, write
G = f x1 x2 . . . xn .E
= f x1 x2 . . . xn . (f f ) cdots(f f )
Then
GG = x1 x2 . . . xn . (GG) (GG)
This means that GG satisfies the defining equation for F . We can therefore write F =
GG, where G = f x1 x2 . . . xn .E .
Encoding minimalization
Now that we know how to convert recursive definitions into lambda terms, we can encode
minimalization as follows. Assume that we want to define a function f : Nk N by
minimalization from a function g : Nk+1 N for which we already have an encoding hgi.
We begin with a recursive definition of a term F as follows.
F = nx1 x2 . . . xk . if iszero(hgi n x1 x2 . . . xk ) then n

else F (succ n) x1 x2 . . . xk
Let F be the lambda term corresponding to F after unravelling the recursive definition.
The encoding hf i of f is then F h0i.
All that remains is to provide an encoding for the predicate iszero.
Encoding of iszero
12.13 A fixed point combinator

Consider the recursive definition F = x.x(F x). By the previous trick for unravelling
recursive definitions, we can find a lambda term for F as follows.
F = GG, where G = f x.x(f f x)

= f x.x(f x.x(f f x)f x.x(f f x)x)
Notice that F X = X(F X) for any lambda term X, by the definition of F . For any term
Z, a fixed point of Z is a term M such that ZM = M . Clearly, if we set M F Z, we obtain
a fixed point for Z. Notice that it does not matter what Z isany lambda term Z has a
fixed point F Z where F is the function we have just constructed. This fixed point operator
is due to Turing and is traditionally denoted .
138
12.14 Making sense of terms without normal forms
Recall that for terms with normal forms, the normal form denotes the value of the term.
What about terms without normal forms? Are all terms without normal forms equally
meaningless?
Suppose we want an equivalence on lambda terms such that:
1. (xM )N M {x N }that is, the equivalence induced by the reduction.
2. If M and N do not have normal forms, then M N .
3. Functions that are equated by yield equivalent results for the same arguments. That
is, if M N then for all R, M R N R.
Then, we can show that is the trivial relation the equates all termsin other words,
M N for all M , N !
To see this, consider the function F defined by
F xb = if b then x else (F xb)

As we have seen above, we can plug in our definition for if-then-else and then unravel
this recursive definition to yield a lambda term for F .
It turns out that
F = GG, where G = f xb.( if b then x else (f f xb)

Now, consider F Xhtti and F Xhffi, where htti and hffi are the encodings of the truth
values true and false.
F Xhtti if hT i then X else (F Xhtti) X.
F Xhffi if hF i then X else (F Xhffi) F Xhffi.
Now, suppose we have an equivalence as described above. Then
FZ (xb.(if b then x else (F xb))Z

(b.(if b then Z else (F Zb)))
(b.(if b then Z else G), where G = if b then Z else (f ZB)
(b.(if b then Z else (if b then Z else G))
...
Since F Z does not terminate for any Z, we have F X F Y for all X and Y . But, since
F X F Y implies that F XM F Y M for all M (by Condition 3 in the definition of ).
But then we have F Xhtti F Y htti. Since F Zhtti Z for all Z, this means that X Y
for all X and Y !
139
140
Chapter 13
Introducing types into the lambda

calculus
So far, we have been looking at the untyped lambda calculus. What happens if we introduce
types into the calculus?
13.1 Simply typed lambda calculus

In the simply typed lambda calculus, we begin with a syntax for types. Since we have no
constants in our set of lambda terms, we fix a single type constant 0 that represents an
atomic type. The syntax of types is then given by the grammar
t := 0 | t t
Thus, examples of types are 0, 0 0, (0 0) 0 0. As we have seen in Haskell, the
operator associates to the right.
We now build up typed terms inductively in such a way that only well typed terms are
generated.
First, for each type s, we fix a separate set of variables. Then, we define by mutual
induction the set s of lambda terms of each type s as follows:
1. For each s, every variable of type s is in s .
2. If M is a term in st and N is a term in s then (M N ) is a term in t .
3. If M is a term in t and x is a variable of type s then (x.M ) is a term in st .
Notice that every term we construct has a well defined type. Terms like (f f ) that gave
us some apparently contradictory behaviour earlier cannot be assigned a type (why?) and
are thus not legal terms of this calculus.
The rule for the simply typed lambda calculus is the usual one, with a check for type
consistency:
141
(x.M )N M {x N } provided x and N are of the same type s
and M is of type s t
It turns out that the corresponding reduction relation for this calculus is Church-
Rosser. In fact, this calculus has a much stronger propertyit is strongly normalizing.
We say that a term is normalizing if it has a normal form. Further, at term is strongly
normalizing if every reduction sequence that the term admits leads to a normal form. We
have seen examples of non-normalizing and non-strongly normalizing terms in the untyped
calculus. For instance
(x.xx)(x.xx) is not normalizing, and
(yz.z)((x.xx)(x.xx)) is not strongly normalizing.
We say that a lambda calculus is strongly normalizing if every term in the calculus
is strongly normalizing. It turns out that the simply typed lambda calculus is strongly
normalizing. We will not prove this result, but one intuitive explanation is that every
application of reduction strictly reduces the type complexity of the lambda term, so there
is a natural limitation on the length of a reduction from a fixed starting term.
We have observed that the syntax of the simply typed lambda calculus permits only well-
typed terms. What if we ask the converse questiongiven an arbitrary term in the untyped
lambda calculus, is it well-typed? That is, is it a valid term in the simply typed lambda
calculus? For instance, a term of the form (f f ) cannot be assigned a sensible type and is
hence clearly not definable in the simply typed calculus. It turns out that this problem is
decidable.
In fact, if M is a typable term, then we can find a unique type s for M such that every
possible type for M is an instance of s (that is, can be obtained by replacing some of the 0s
in s by another type). Such a type is called the principal type scheme for M.
Observe also that the encodings hmi for natural numbers that we have provided are well
typed. Hence, we can carry over much of our arithmetic from the untyped calculus to the
simply typed calculus.
13.2 Polymorphic typed calculi

The simply typed lambda calculus has only explicit types. It cannot describe polymorphic
functions as found, say, in Haskell. To remove this limitation, Girard and Reynolds indepen-
dently invented what is called the second-order polymorphic lambda calculus (also known
as Girards System F ).
We begin by extending the syntax of types to include type variables, that we shall denote
a,b . . . . We assume that our lambda calculus includes constants in addition to variable, so
we also permit an arbitrary set of type constants, typically denoted i, j, . . . .
The set of type schemes is given by the following grammar:
s ::= a | i | s s | a.s
142
The terms of the second order polymorphic lambda calculus are then given by
1. Every variable and constant is a term.
2. If M and N are terms, so is (M N ).
3. If M is a term, x is a variable and s is a type scheme, then (x s.M ) is a term.
4. If M is a term and s is a type scheme, (M s) is a term.
5. If M is a term and a is a type variable, then (a.M ) is a term.
The last two rules allow us to define functions whose types are defined in terms of type
variables (like polymorphic functions in Haskell). Rule 5 says that we can make a type
variable a parameter of an expression to get type abstraction, just as we make a normal
variable into a parameter in lambda abstraction. Rule 4 permits us to apply a type
abstraction to a concrete type.
In this syntax, a polymorphic identity function would be written as
a.x a.x
We have two rules, one for lambda abstraction and one for type abstraction.
(x s.M )N M {x N }.
(a.M )s M {a s}.
It turns out that this calculus is also strongly normalizing. However, it is not known
whether the type checking problem is decidablethat is, given a term, can it be assigned a
sensible type. This process of assigning a type to a term without explicit type information
is called type inference.
What is a sensible type? There is a natural way in which the type of a complex
expression can be deduced from the types assigned to its constituent parts. One way to
formalize this is to define a relation A ` M : s where A is list {xi : ti } of type assumptions
for variables. We read this relation as follows: under the assumptions in A, the expression
M has type s.
Clearly, for a variable x,
A ` x : s iff (x : s) belongs to A
We use the following notation to extend lists of assumptions. Let A be a list of assumptions.
Then, A + {x : s} is the list in which all variables other than x carry the same assumptions
as in A and any assumption for x in A, if such an assumption exists, is overridden by the
new assumption x : s.
We can now present the type inference rules for complex terms as follows:
143
A + {x : s} ` M : t
A ` (x s.M ) : s t
A ` M : s t, A ` N : s
A ` (M N ) : t
A`M :s
A ` (a.M ) : a.s
A ` M : a.s
A ` M t : s{a t}
For instance, we can derive a type for our polymorphic identity function as follows:
x:a`x:a
` (x a.x) : a a
` (a.x a.x) : a.a a
As we remarked earlier, the decidability of type inference for this calculus is openthat is,
we do not have an algorithm (nor can we prove the nonexistence of an algorithm) to decide
whether a given expression can be assigned a type that is consistent with the inference system
described above.
How is it then that type checking is possible for languages like Haskell? The answer is
that Haskell, ML and other functional programming languages use a restricted version of
the polymorphic types defined here.
In Haskell, all type variable are universally quantified at the top level. Thus, if we write
a Haskell type such as
map :: (a -> b) -> [a] -> [b]
we really mean the type
map :: a, b. (a b) [a] [b]
This implicit top level universal quantification of all type variables is known as shallow typing.
In contrast, the second order polymorphic calculus, we can permits deep typing, with more
complicated type expressions such as
a. [(b. a b) a a]
Note here that the b quantifier inside the expression cannot be pulled out to the top without
altering the meaning of the type.
144
13.3 Type inference as equation solving
One way to tackle the type inference problem is to represent it as a problem of solving
a system of equations involving type variables. Consider, for instance, how we intuitively
assign a type to the following function definition in Haskell:
twice f x = f (f x)
We begin by observing that twice is a function of two arguments so, generically, we have
twice :: a -> (b -> c) for some unknown types a, b and c. We then observe that the
arguments f and x occur as a function application f x in the right hand side, so we have the
further requirement that a, the type of f, must itself be of the form d -> e and the type of
x must be compatible with d -> e, so b must be the same as d. Then, we see that the term
(f x) itself appears as an argument to f so the output type e of f must be compatible with
its input type d. Finally, since the output of twice is f (f x), we must have c compatible
with e. We can represent this reasoning in terms of the following set of type equations.
a = d -> e (because f is a function)

b = d (because f is applied to x
e = d (because f is applied to (f x)
c = e (because output of twice is f (f x)
If we solve these equations, we come to the conclusion that all four type variables b, c,
d and e must in fact be equal to each other and a is of the form b -> b (or c -> c or . . . ).
However, there is no further restriction on these variables, so we arrive at the conclusion
that
twice :: (b -> b) -> b -> b
Here, the variable b is universally quantified. Notice that by leaving the interpretation of b
open, we arrived a most general type for twice.
If we have an expression involving twice, such as twice f 5, we can deduce that since
5 is of type Int, the type variable b in the type of twice must be uniformly assigned the
value Int. Thus, in twice f 5, the (unknown) function f must have type Int -> Int and
the overall type of the expression must be Int.
13.4 Unification
In general, for type inference we need a systematic method to perform the intuitive compu-
tation described above. The solution we seek to our system of equations is a special type of
function, called a substitution, that maps each variable to an expression in such a way that
all the equations are satisfied. In order to find the most general solution to a system of
equations involving variables, we have to identify a substitution with least constraints.
145
This problem is called unification and the least constrained substitution map that we
seek is called the most general unifier or mgu. In 1965, J.A. Robinson described an algorithm
that solves the unification problem. We present a version of the algorithm below.
In general, the unification problem is one of solving equations involving terms over a
fixed signature. A signature consists of a collection of function symbols, each of a fixed arity.
Constants can be regarded as functions with arity 0. In addition we have variables that
can stand for arbitrary terms. Informally, a term is a well-formed expression consisting of
function symbols and variables. Formally, terms can be defined inductively as follows:
Every variable is a term.
If f is a k-ary function symbol in the signature and t1 , t2 , . . . , tk are terms, then

f (t1 , t2 , . . . , tk ) are terms.
In the discussion that follows, we use lower case letters like a, b, f , p, . . . to denote
function symbols and capital letters like X, Y , . . . to denote variables. An instance of the
unification problem is a system of equations, like the following:
f (X) = f (f (a))
g(Y ) = g(Z)
To solve this system, we have to assign values to the variables X, Y and Z. The values
we assign are terms and, as we mentioned earlier, an assignment of terms to variables is
called a substitution. For instance, we could assign X the term f (a), Y the term g(a) and Z
the term g(a). We write such a substitution as {X f (a), Y g(a), Z g(a)}. We can
give this substitution a name, for instance .
Given a term t and a substitution , the effect of applying to t is usually written t
(note that this convention reverses the usual notation for function application). The term t
is the result obtained by uniformly mapping the variables in t according to the assignments
dictated by . For instance, in the example above, if = {X f (a), Y g(a), Z g(a)}
and t = f (X), then t = f (f (a)). We can easily verify that f (X) = f (f (a)) and g(Y ) =
g(Z), so is a solution to the unification problem given above.
Substitutions are applied in parallel, simultaneously. For instance, if we have an expres-
sion of the form g(p(X), q(f (Y ))) and we apply the substitution = {X Y, Y f (a)}
to this expression, we get a new expression g(p(Y ), q(f (f (a)))). The point to note is that
subexpression p(Y ) obtained by the assignment X Y is not transformed transitively into
p(f (a)) by the assignment Y f (a) in .
Coming back to the example above, the substitution = {X f (a), Y g(a), Z
g(a)} is not the only solution to the set of equations
f (X) = f (f (a))
g(Y ) = g(Z)
Another, simpler, solution is 0 = {X f (a), Y a, Z a}. Yet another solution is
00 = {X f (a), Y Z}. The last solution is the least constrained substitution. Any
146
other substitution that solves the problem can be broken upto into two steps, the first of
which is the substitution 00 . For instance, is the same as 00 followed by the substitution
{Y g(a)}. A solution that has this property is called a most general unifier for the
unification problem.
Not all unification problems admit a solution. For instance, if we have an equation of
the form p(Z) = q(f (Y )) with different outermost function symbols, no substitution for Z
and Y can make these terms equal.
We also have a problem with equations of the form X = f (X). Since X occurs in f (X),
any substitution for X will also apply to the nested X within f (X) and the two terms cannot
be made equal.
It turns out that any set of equations in which these two problems do not occur admits
a most general unifier.
A unification algorithm
The algorithm to find a most general unifier proceeds as follows.
We begin with a set of equations {tl1 = tr1 , tl2 = tr2 , . . . , tlm = trm }.
Perform the following transformations on the set of equations so long as any one of
them is applicable.
1. Transform t = X, where t is not a variable, to X = t.

2. Erase any equation of the form X = X.
3. Let t = t0 be an equation where neither t nor t0 is a variable. Then, t = f (. . .)
and t0 = f 0 (. . .) for some function symbol f and f 0 .
If f 6= f 0 , terminatethe set of equations is not unifiable.
Otherwise, t = f (t1 , t2 , . . . , tk ) and t0 = f (t01 , t02 , . . . , t0k ), where f is of arity k.
Replace the equation
f (t1 , t2 , . . . , tk ) = f (t01 , t02 , . . . , t0k )

by k new equations
t1 = t01 , t2 = t02 , . . . , tk = t0k

4. Let X = t be an equation such that X occurs in t. Terminatethe set of equations
is not unifiable.
5. Let X = t be an equation such that X does not occur in t and X also occurs in
some other equation. Transform the set of equations by replacing all occurrence
of X in other equations by t.
147
We claim that this algorithm terminates with a most general unifier. Termination follows
from the fact that the first four transformation rules can be used only a finite number of
times without using rule 5. Let n be the number of distinct variables in the original set of
m equations. Rule 5 can be used at most one time for each variable. Thus, overall, the five
transformations can be applied only a finite number of times.
When no rules apply, every equation is of the from X = t and each variable X that
appears on the left hand side of an equation does not appear anywhere else in the set of
equations. Thus, the resulting set of equations defines a substitution
{X1 t1 , X2 t2 , . . . , Xn tn }
We have to argue that this is a most general unifier. First, we argue that it is a unifying
substitution. This follows from the fact that each of the transformations in the algorithm
preserves the set of unifiers. Arguing that it is an mgu is more complicated and we omit the
proof.
Here is an example of how the algorithm works. We start with the equations
g(Y ) = X, f (X, h(X), Y ) = f (g(Z), W, Z)

Applying rule 1 to the first equation and rule 3 to the second equation, we get
X = g(Y ), X = g(Z), h(X) = W, Y = Z

Using rule 5 on the second equation, we replace X by g(Z) everywhere to get
g(Z) = g(Y ), X = g(Z), h(g(Z)) = W, Y = Z

Applying rule 3 to the first equation we get
Z = Y, X = g(Z), h(g(Z)) = W, Y = Z
Using rule 5 on the last equation to replace Y by Z and eliminating the resulting equation
Z = Z by rule 2, we get
X = g(Z), h(g(Z)) = W, Y = Z
Finally, we reverse the second equation by rule 1 to get
X = g(Z), W = h(g(Z)), Y = Z
No further rules apply, so the most general unifier is {X g(Z), W h(g(Z)), Y Z}.
The algorithm we have described is different from Robinsons original algorithm, whose
correctness is more difficult to establish. The most efficient algorithms for unification work
in linear time.
148
13.5 Type inference with shallow types
Consider the basic lambda calculus, enhanced with constants. We have some built-in types
i, j, k, . . . and a set of constants Ci for each built-in type i. For instance, if i is the type Char,
Ci is the set of character constants. The syntax of lambda terms is then
= c | x | x.M | M N
where c Ci for some built-in type i, x is a variable and M, N .
To infer types for these terms, we can set up equations involving type variables inductively
as follows. Let M be a lambda term. Then:
If M is a constant c Ci , the type of M is i.
If M is a variable x, assign a fresh type variable to x.
If M is of the form x.M 0 , assign M the type for fresh type variables and
. Inductively, let be the type of x in M 0 . Then, the input type of x.M 0 should
match the type of x, so add the equation = to the set of type equations.
If M is of the form M 0 N 0 , we must inductively have assigned M 0 a type of the form

and N 0 a type of the form . Assign M the type and add the equation =
to the set of type equations to enforce that this application is well-typed.
With this syntax, suppose we want to write a function equivalent to the following Haskell
definition.
applypair f x y = (f x,f y)
In this definition, we might like to permit f to be a polymorphic function and x and y

to be of different types. For instance, we might ask for the following expression to be well
typed, where id is the identity function id z = z.
applypair id 7 c = (id 7, id c) = (7,c)
However, if we try to assign a type to this function, we find that we have the following
set of constraints, which cannot be unified.
id :: a -> a
7 :: Int
c :: Char
a = Int (from id 7)
a = Char (from id c)
149
In fact, the type that Haskell assigns to applypair is (a -> b) -> b -> b -> (b,b).
To see why this is so, let us look how this function would be defined in the version of the
lambda calculus we have just defined. The corresponding expression is
f xy.pair (f x)(f y) where pair xyz.(zxy)

As the type inference rules we have provided suggest, when we pass a function to this term
for the argument f with type , the value of will have to be unified with the types of
both x and y because of the expressions (f x) and (f y) in the body of the expression. This
then forces x and y to have the same type.
One way to permit the definition of the richer version of applypair is to extend the
syntax of lambda terms to allow local definitions as follows.
= Ci | x | x.M | M N | let f = e in M
Here f = e in let . . . in . . . provides a local declaration. We can now define the lambda
term for applypair id as
let f = z.z in xy.pair (f x)(f y)

which translates into Haskell as
applypair x y = (f x,f y) where f z = z
or, equivalently,
applypair x y = let f z = z in (f x,f y)
What is the type inference rule for let . . . in . . . ? Here is a first attempt at formulating
the rule.
Let M = let f = e in M 0 where f inductively has type t. Let {, , . . .} be the set of

type variables that occur in t. Then for each instance of f in M 0 , make copies of these
variables. Thus, the first instance of f in M 0 will be assigned type t with the variables
modified to 1 , 1 , . . ., the second instance will be assigned type t with the variables
modified to 2 , 2 , . . ., and so on.
With this rule, we find that the two occurrences of f in let f = z.z in xy.pair (f x)(f y)
have types 1 1 and 2 2 . The expression f x unifies the type variable 1 with the
type of x while the expression f y unifies the type variable 2 with the type of y. However,
since 1 and 2 are different variables, this does not force the type of x to match the type
of y.
Notice that from the point of view of -reduction, the terms let f = e in x.M and
(f x.M )e are equivalent. But, as we have seen, the type inference rules for the two expres-
sions are quite different.
150
In a more extreme example, one form may be typable while the other is not. Consider the
expressions (I.(II))(x.x) and let I = x.x in (II). Here, the first form cannot be typed
because no self-application f f can be typed. However, in the let . . . in . . . version, the type
of x.x is and this is copied as 1 1 and 2 2 for the copies of I in the body
of the function. We can then unify 1 with 2 2 to derive that the first I in II has the
type (2 2 ) (2 2 ) and, overall, II has type (2 2 ).
There is a subtlety that we have overlooked in our type inference rule for let . . . in . . . .
Consider the function
applypair2 w x y = ((tag x),(tag y))

where
tag = pair w
pair s t = (s,t)
Intuitively, it is clear that the type of applypair2 is a -> b -> c -> ((a,b),(a,c))
because applypair2 constructs the pair of pairs ((w,x),(w,y)) from its three inputs w, x
and y.
However, if we apply our type inference strategy, we begin with the types:
applypair2 :: a -> b -> c -> (d,e)

pair :: f -> g -> (f,g)
tag :: h -> (i,h)
Using our variable copying rule for let . . . in . . . , we get that the two instances of tag in
applypair2 have type d = h1 -> (i1,h1) and e = h2 -> (i2,h2). The application tag
x gives us the constraint h1 = b and the application tag y gives us the constraint h2 = c.
The application pair w gives us the constraint a = i, but this constraint is not propagated
to the copies i1 and i2 that we have made of i when expanding the type of applypair2.
Thus, we get the type applypair2 :: a -> b -> c -> ((i1,b),(i2,c)) instead of the
correct type a -> b -> c -> ((a,b),(a,c)).
Clearly, the problem lies in the copies that we made of the type variable i when instan-
tiating the types for the two copies of tag. The distinction between the variables h and i
that appear in the type of tag is that the variable i is unified with the type of one of the
arguments to the main function within which tag is defined. In the literature, variables like
h that are free within the local definition are called generic variables. The corrected type
inference rule for let . . . in . . . requires that copies be made only for generic variables.
Let M = let f = e in M 0 where f inductively has type t. Let {, , . . .} be the set of

generic type variables that occur in t. Then for each instance of f in M 0 , make copies
of these generic variables. Thus, the first instance of f in M 0 will be assigned type t
with the generic variables modified to 1 , 1 , . . ., the second instance will be assigned
type t with the generic variables modified to 2 , 2 , . . ., and so on. The non-generic
type variables in t retain their identity in all instances of f .
151
If we apply this new rule when assigning a type to applypair2, we observe that the type
variable i should not be duplicated. This results in the type a -> b -> c -> ((i,h1),(i,h2))
being assigned to applypair2 after copying the generic variables of the type of tag into the
two instances of tag that occur in the body of applypair2. The constraints a = i, b = h1
and c = h2 then result in the correct type a -> b -> c -> ((a,b),(a,c)) being assigned
to applypair2.
The type inference strategy that we have sketched here can be efficiently implemented
using a set of inference rules in the same style as those that we described for the second-order
polymorphic typed lambda calculus in an earlier section. This algorithm was first described
by Milner (he called it Algorithm W) and was implemented in the programming language
ML and, subsequently, in other typed functional languages like Haskell. The important fact
is that for these languages, type inference is decidable, so the programmer can write function
definitions without providing explicit type information. The type inference system can then
recognize whether the expression is well-typed. Freeing the programmer from the burden
of providing explicit types makes these strongly typed functional languages significantly
easier and more attractive to use. In a sense, the combination of let . . . in . . . definitions
in the lambda calculus and shallow types seems to provide an optimum combination of
expressiveness and useability in the typed lambda calculus.
152
Part VII
Logic programming
153
Chapter 14
Introduction to logic programming
Logic programming is a probably more accurately called programming with relations, in

contrast to functional programming, which is programming with functions.
The idea is that a function f with n arguments can always be written as a relation Rf
with n+1 arguments:
f (x1 , x2 , . . . , xn ) = y iff (x1 , x2 , . . . , xn , y) Rf

Having gone from functions to relations, we can generalize our objectives and give rules
to compute not only those special relations that correspond to functions (where the last
component of each tuple is uniquely fixed by the others), but more general relations as well.
We will describe logic programming using a syntax that is based on Prolog, the most
commonly used programming language in this paradigm.
14.1 Facts and rules

We start with two types of simple entities, variables and constants. The convention is that
variables start with a capital letterX, Y, Name, . . . , while constants start with a lower case
letterball, node, graph, a, b, . . . .
Constants are, in general, uninterpretedthat is, they are not further organized into
types like Char and Float with underlying meaning. However, for convenience, we will
allow ourselves to use natural numbers as a special type of constant and compute arithmetic
expressions over these numbers.
A Prolog program defines a relation through facts and rules. For instance, consider a
directed graph with five nodes.
1
@
@R
@
5H
Y
H 2
6 HH
H
H?
4 3
155
We can represent its edge relation using the following facts.
edge(3,4).
edge(4,5).
edge(5,1).
edge(1,2).
edge(3,5).
edge(3,2).
Each fact just lists out explicitly a concrete tuple that belongs to the relation edge.
We can now define another relation path using a set of rules, as follows:
path(X,Y) :- edge(X,Y).
path(X,Y) :- edge(X,Z), path(Z,Y).
These rules are to be read as follows:
Rule 1 For all X,Y, (X,Y) belongs to path if (X,Y) belongs to edge.
Rule 2 For all X,Y, (X,Y) belongs to path if there exists Z such that (X,Z) belongs to edge
and (Z,Y) belongs to path.
Thus, each rule is of the form Conclusion if Premise1 and Premise2 . . . and Premisen .
This special type of logical formula is called a Horn Clause. In Prolog, the if is written as
:- and the ands on the right hand side are written as commas. The left hand side of a rule
is called the goal.
Notice the implicit quantification of variablesvariables that appear in the goal are uni-
versally quantified (for example, X and Y above) while variables that appear in the premises
are existentially quantified (for example, Z above).
14.2 Computing in Prolog

To start off a computation in Prolog, we ask a query. For instance, we may ask, at the
Prolog interpreter prompt ?- whether there is a path from 3 to 1, as follows
?- path(3,1).
To answer this, Prolog will try to use the rules given above and see if it can find a
satisfying assignment for all the variables that answers this query. In this example, it scans
the rules for path(X,Y) from top to bottom and proceeds as follows:
(3,1) is not in edge, so skip Rule 1.
Rule 2 generates two subgoals. Does there exists Z such that
156
(3,Z) belongs to edge and
(Z,1) belongs to path.
Prolog tries to satisfy the first subgoal by scanning the edge relation from top to bottom
and finds edge(3,4), so Z is set to 4.
The second goal then becomes path(4,1). Since (4,1) is not in edge, this again
generates two subgoals: Does there exist Z (a new variable, not the old Z) such that
(4,Z) belongs to edge and

(Z,1) belongs to path.
Once again, to satisfy the first goal we scan edge from the top and find edge(4,5), so
Z is set to 5.
The second goal then becomes path(5,1). Now, (5,1) does belong to edge, so
path(5,1) is true.
We can now backtrack and report successboth the subgoals for path(4,1) have
succeeded, path(4,1) is true.
Backtracking once more, both the subgoals for the original goal have succeeded, so
path(3,1) is true.
Essentially, Prolog does a depth first search through all possible satisfying assignments
to the variables in the premises in order to satisfy the goal. In this case, it found a path of
length three, even though there is a shorter path of length two. This is because of the way
that the facts in edge are arranged. If we had placed the fact edge(3,5) before edge(3,4), Z
would have been set to 5 in the first search itself. (What would happen if we place edge(3,2)
before both edge(3,4) and edge(3,5)?)
14.3 Complex structures in Prolog

In general, we can represent arbitrary structures in Prolog by using nested terms. For
instance, we can represent a Pascal record or a C struct of the form:
personal_data{
name : amit
date_of_birth{
year : 1980
month : 5
day : 30
}
}
157
by a Prolog term
personal_data(name(amit),date_of_birth(year(1980),month(5),day(30)))
The list type is built in to Prolog. Analogous to list decomposition by the expression
(head:tail) in Haskell, we can decompose a list in Prolog as [Head|Tail]. For instance,
we can write a function to check if an item belongs to a list as follows:
member(X,[X|T]).
member(X,[Y|T]) :- member(X,T).
In this function, for the first rule to succeed, X must match the head of the list. Formally,
this is done using unification, which we have seen in our discussion of type inference. The
important thing to note is that unification permits implicit pattern matching of more than
one part of the goal, a feature that we noted was not permitted in patterns as used in Haskell
definitions. The second rule can be invoked only if the first rule fails, in which case we are
guaranteed that the head of the list does not match X, so X and Y are different. In this case,
we make a recursive call.
Notice that the names T in the first rule and Y in the second rule are not really used.
Prolog allows the use of a dummy variable _ in such cases.
member(X,[X|_]).
member(X,[_|T]) :- member(X,T).
All other variables are uniformly substituted in a ruleif X appears in many places, all
Xs will be replaced by the same value. However, for convenience, the dont care variable
_ is not subject to this limitation, so each occurrence of _ in a rule can take on a different
value. Thus, we can use the same dont care variable _ in many places in the same rule
without implicitly asserting that all these positions acquire the same value. This is similar
to the use of _ in Haskell patterns.
14.4 Turning the question around

Since Prolog deals with relations and not functions, we can invoke inverse computations in a
natural way, unlike functional programming. For instance, in our first example, we can ask
?- edge(3,X).
This asks for all values of X that satisfy the relation edge. If we respond with ; to each
answer provided, Prolog will exhaustively list out all possible values that X can take, in the
order in which they are defined. The response would look something like
158
X=4;
X=5;
X=2;
No.
The final No. indicates that there are no further ways of satisfying the goal. Likewise
?- path(3,Y).
will exhaustively list out all values for Y that make this relational assertion true.
14.5 Arithmetic
A first attempt to write a program to compute the length of a list would look as follows.
length([],0).
length([H,T],N) :- length(T,M), N = M+1.
If we run this program, we get a somewhat surprising result.
?- length([1,2,3,4],N).
N=0+1+1+1+1
This is because = signifies unification, not equality. Thus, when the base case is reached
that is, length(T,M) with T equal to the empty listM is assigned 0. The phrase N=M+1
unifies N with M+1 and sets N to the expression 0+1. As the program pops out of each level
of recursion, an extra +1 is added, resulting in the final expression that we saw above.
To match arithmetic expressions based on their value, Prolog provides the connective is.
Thus, length should be defined as follows.
length([],0).
length([H,T],N) :- length(T,M), N is M+1.
Now, the query length([1,2,3,4],N) behaves as we would expect.
?- length([1,2,3,4],N).
N=4
Here is a more convoluted version of length.
length(L,N) :- auxlength(L,0,N).
auxlength([],N,N).
auxlength({H,T],M,N) :- auxlength(T,M1,N), M1 is M+1.
159
Consider what happens when we ask for
?- length([0,1,2],N)
The goals generated are
auxlength([0,1,2],0,N) -> auxlength([1,2],1,N) ->

auxlength([2],2,N) -> auxlength([],3,N) -> auxlength([],3,3)
so Prolog reports back that
N = 3.
Observe the role played by unificationthe variable N is being propagated through all the
subgoals and the final substitution works its way back to the original query.
14.6 Negation as failure

In Prolog, we can write a premise of the form not(R(X,Y,Z)). This becomes true if Prolog is
not able to justify R(X,Y,Z). In other words, for each relation Prolog uses a closed world
assumptionthe rules and facts exhaustively determine all the members of the relation and
anything that is not computable as being part of the relation is out of the relation. This
approach is called negation as failure.
Negation has some unexpected effects. For instance, consider the following query.
?- not(X=5).
No.
The No is justified as follows. To check not(X=5), Prolog first tries the goal X=5. Since X
can be unified with 5, this goal succeeds. Therefore, not(X=5) fails.
Now, consider a query to extract from a list Ls all elements that are not equal to 1. We
could write it two ways:
?- not(X=1), member(X,Ls).
?- member(X,Ls), not(X=1).
In the first version, not(X=1) always fails, as we saw above, because X is not bound to
any value initially. This results in the query failing. The second query, however, does the
job that we want because X is first bound to an element of Ls before being tested for X=1.
160
14.7 Cut
Prolog provides a somewhat controversial mechanism to control backtracking. If we insert
an exclamation mark ! in the set of premises, backtracking will not cross the exclamation
mark.
For instance, if we rewrote path above as
path(X,Y) :- edge(X,Z),!, path(Z,Y).
then for each X, it would try path(Z,Y) only once, for the first edge of the form edge(X,Z).
If there is no path using this value of Z it will not backtrack to another edge involving Xthe
overall search just fails.
Cut is useful for optimizing Prolog code to avoid unnecessary backtracking. Suppose
we want to program the ternary predicate ifthenelse(B,X,Y) that evaluates X if B is true
(that is, if the goal B succeeds) and evaluates Y otherwise. A first attempt would be to write.
ifthenelse(B,X,Y) :- B,X.
ifthenelse(B,X,Y) :- not(B),Y.
We have to look at Prolog backtracking a little more to explain this code. If the first rule
fails for a goal, Prolog will try alternative rules before backtracking along the first rule.
Thus, if we omit the not(B) in the second rule, we have a problem. If B succeeds and X
fails, Prolog backtracking would directly try to satisfy Y. By placing a guard not(B) on the
second clause, we ensure that this does not happen. However, evaluating this guard involves
an unnecessary computation because, if we reach this stage, we have already evaluated B so
we already know the result not(B).
A second source of inefficiency in this definition lies in the first rule. Suppose that B
succeeds and X fails. We should then conclude that the ifthenelse(B,X,Y) fails. However,
if there are multiple ways of satisfying B, Prolog backtracking will retry X for all of these
options before abandoning the original goal!
We can fix both these inefficiencies by introducing a cut in the first rule.
ifthenelse(B,X,Y) :- B,!,X.
ifthenelse(B,X,Y) :- Y.
The cut has two effects. First, as we mentioned earlier, it prevents backtracking to B.
Second, it discards all future alternative rules for this goal. Thus, even if X fails, the cut
ensures that Y is not tried. On the other hand, if B fails, which is before the cut, normal
backtracking takes Prolog to the second rule and tries Y without wastefully computing not(B)
as in the previous version.
In more complex context, it is quite difficult to predict the portion of the search tree that
is pruned by such a cut, so this feature is to be used with caution.
161
162
Part VIII
Scripting languages
163
Chapter 15
Programming in Perl
For the most part, we have focussed on features in programming languages that add structure
and discipline to the process of programming. The fundamental idea has always been to
achieve suitable degree of abstraction. To ensure the integrity of the abstraction, we have
seen different approaches to defining a strong notion of type combined with the flexibility to
define polymorphic functions.
At the other end of the spectrum we have scripting languages that impose minimal
programming discipline but are ideal for rapid deployment of programs that attack routine
tasks. One of the most successful scripting languages is Perl, an imperative language whose
control flow is along the lines of C, but with several interesting features that make it ideal
for text processing.
15.1 Scalar datatypes

Values associated with scalar variables in Perl are either numbers or character strings. Num-
bers correspond to double precision floating point numbers. Variable names begin with the
special character $. For instance, we could have the following assignments in a Perl program.
$num1 = 7.3;
$str1 = "Hello123";
$str2 = "456World";
Variables (and their types) do not have to be declared. Moreover, the type of a variable
changes dynamically, according to the type of expression that is used to assign it a value. We
use normal arithmetic operators to combine numeric values. The symbol . denotes string
concatenation. If a numeric value is used in a string expression, it is converted automatically
to the string representation of the number. Conversely, if a string is used in a numeric
expression, it is converted to a number based on the contents of the string. If the string begins
with a number, the resulting value is that number. Otherwise, the value is 0. Continuing
our example above, we have the following (text after # is treated as a comment in Perl).
165
$num3 = $num1 + $str1; # $num3 is now 7.3 + 0 = 7.3
$num4 = $num1 + $str2; # $num4 is now 7.3 + 456 = 463.3
$str2 = $num4 . $str1; # $str2 is now "463.3Hello123"
$str1 = $num3 + $str2; # $str1 is now the number 470.6
$num1 = $num3 . "*" . $num4; # $num1 is now the string "7.3*463.3"
By default, Perl values are made available to the program the moment they are introduced
and have global scope and are hence visible throughout the program. Uninitialized variables
evaluate to 0 or "" depending on whether they are used as numbers or strings. The scope
of a variable can be restricted to the block in which it is defined by using the word my when
it is first used.
my $a = "foo";
if ($some_condition) {
my $b = "bar";
print $a; # prints "foo"
print $b; # prints "bar"
}
print $a; # prints "foo"
print $b; # prints nothing; $b has fallen out of scope
15.2 Arrays
Perl has an array type, which is more like a list in Haskell because it can grow and shrink
dynamically. However, unlike arrays in Java or C and lists in Haskell, Perl arrays can contain
a mixture of numeric and string values. Collectively, a Perl array variable starts with the
symbol @, thus distinguishing it from a scalar. An array can be explicitly written using the
notation (element1,element2,...,elementk).
Perl supports array like notation to access individual elements of an arrayarray indices
run from 0, as in C or Java. The element at position $i in array @list is referred to as
$list[$i], not @list[$i]. A useful mnemonic is that an element of an array is a scalar,
so we use $, not @.
The [...] notation can be used after the @ name of the array to denote a sublist. In
this form, we provide a list of the indices to be picked out as a sublist. This list of indices
can be in any order and can have repetitions.
@list = ("zero",1,"two",3,4); # Note the mixed types
$val1 = $list[3]; # $val1 is now 3
$list[4] = "four"; # @list is now ("zero",1,"two",3,"four")
@list2 = @list[4]; # @list2 is ("four")
@list3 = @list[4,1,3,4,0]; # @list3 is ("four",1,3,"four",0)
A key fact about Perl lists is that they are always flatnested lists are automatically
flattened out. Thus, we dont need to write functions such as append to combine lists.
166
@list1 = (1,"two");
@list2 = (3,4);
@list = (@list1,@list2); # @list is (1,"two",3,4)
$list = ("zero",@list); # @list is now ("zero",1,"two",3,4)
The example above shows that we can use a list variable on the left hand side of an
assignment. We can also combine scalar variables into a list to achieve a multiple parallel
assignment. If we use a list variable as part of the multiple assignment on the left hand side
it soaks up all the remaining values.
($first,$second) = ($list[0],$list[1]);
($first,@rest) = @list; # $first is $list[0]
# @rest is @list[1,2,...]
($first,@rest,$last) = @list; # $first is $list[0]
# @rest is @list[1,2,...]
# $last gets no value
($this,$that) = ($that,$this); # Swap two values!
Perl makes a distinction between list context and scalar context. The same expression
often yields different values depending on the context in which it is placed. One example of
this is that a list variable @list, when used in scalar context, evaluates to the length of the
list.
@n = @list; # Copies @list into @n

$n = @list; # $n is the length of @list
$m = $list[@list-1]; # $m is the last element of @list
Perl has several builtin functions to manipulate arrays. For instance, reverse @list
reverses a list. The function shift @list removes and returns the leftmost element of
@list. The function pop @list removes and returns the last element of @list. The function
push(@list,$value) adds $value at the end of @list while unshift(@list,$value) adds
$value at the beginning of @list.
15.3 Control flow

Perl supports the normal branching and looping constructs found in C or Java:
if (condition) {statement-block} else {statement-block}
while(condition) {statement-block}
for (init; condition; step) {statement-block}
167
The only thing to remember is that in Perl the braces around the statement blocks are
compulsory, even they contain only one statement.
When evaluating the boolean condition in these statements, the number 0, the strings
"0" and "" and the empty list () are all treated as false. All other values are true.
Because Perl does not have explicit declarations of numeric and string variables, there
are separate comparison operators for numeric and string values. The symbolic operators <,
<=, >, >=, == and != compare their arguments as numeric values while string comparison is
denoted by the operators lt, le, gt, ge, eq and ne.
In addition to the basic compound statements above, Perl supports an abbreviation elsif
for mulitply branching if statements, a version of if called unless and a version of while
called until, both of which invert the sense of the condition (that is, the statement block is
executed if the condition is false).
if (condition) {statement-block} elsif {statement-block}
elsif {statement-block} ... else {statement-block}
unless (condition) {statement-block}
until (condition) {statement-block}
For stepping through a list, Perl provides a special loop called foreach.
foreach $value (@list){ # $value is assigned to each item in
print $value, "\n"; # @list in turn
}
15.4 Input/Output
Files can be opened for input or output using the open() function. As in languages like C,
the open() statement associates a filehandle with a filename and a mode for the file, namely
input, output or append.
Here are some examples. In these examples the function die exits after printing a diag-
nostic. It is separated by an or whose semantics is that the second part of the or is executed
only if the first part fails.
open(INFILE, "input.txt") or die "Cant open input.txt";
# Open in read mode -- could also write open(INFILE,"<input.txt")
open(OUTFILE, ">output.txt") or die "Cant open output.txt";
# > indicates open in (over)write mode
open(LOGFILE, ">>my.log") or die "Cant open logfile";
# >> indicates open in append mode
Reading a file is achieved using the <> operator on a filehandle. In scalar context it reads
a single line from the filehandle, while in list context it reads the whole file in, assigning each
line to an element of the list:
168
$line = <INFILE>; # Read one line from INFILE into $line
@lines = <INFILE>; # Read all lines from INFILE into list @lines
The <> operator is most often seen in a while loop:
while ($line = <INFILE>) { # assigns each line in turn to $line

print "Just read in this line:", $line;
}
The example above also introduces the print statement. The simplest form of the print
statement takes a list of string arguments separated by commas. print can also take an
optional first argument specifying which filehandle to print to:
print "Hello, world";

print "Hello ", $name,"\n";
print OUTFILE $record;
print LOGFILE $logmessage;
To close a file, we call the function close with the corresponding filehandle.
15.5 Matching and regular expressions

We mentioned in the beginning that Perl is well suited to text manipulation. A particularly
useful feature in this regard is a powerful matching operator, that extends naturally to a
search-and-replace operator.
The match operator is =~ which allows us to match a string variable against a pattern,
usually delimited by /. For instance, to print all lines from a file that contain the string CMI,
we would write:
while ($line = <INFILE>) {

if ($line =~ /CMI/){
print $line;
}
}
More generally, the pattern could be a regular expression. The syntax for regular expres-
sions is similar to that in text editors like vi or emacs or in the grep command of Unix.
In a regular expression, a character stands for itself. A sequence of characters in square
brackets stands for a choice of characters. Thus, the pattern /[Cc][Mm][Ii]/ would match
any combination of lower and upper case letters that make up the string cmifor instance,
CMI, CmI, . . . . The character . is special and matches any character, so, for instance, the
pattern /[Cc].i/ would match any three letter pattern beginning with C or c and ending
with i. We can specify a case-insensitive search by appending the modifier i at the end of
the pattern.
169
if ($line =~ /CMI/i){ # Same as ($line =~ /[Cc][Mm][Ii]/)
Perl provides some special abbreviations for commonly used choices of alternatives. The
expression \w (for word ) represents any of the characters _,a,. . . ,z,A,. . . ,Z,0,. . . ,9. The ex-
pression \d (for digit) represents 0,. . . ,9, while \s represents a whitespace character (space,
tab or newline).
Repetition is described using * (zero or more repetitions), + (one or more repetitions) and
? (zero or one repetitions). For instance the expression \d+ matches a nonempty sequence
of digits, while \s*a\s* matches a single a along with all its surrounding white space, if
any. More controlled repetition is given by the syntax {m,n}, which specifies between m and
n repetitions. Thus \d{6,8} matches a sequence of 6 to 8 digits.
A close relative of the match operator is the search and replace operator, which is given
by =~ s/pattern/replacement/. For instance, we can replace each tab (\t) in $line by a
single space by writing
$line =~ s/\t/ /;
More precisely, this replaces the first tab in $line by a space. To replace all tabs we have
to add the modifier g at the end, as follows.
$line =~ s/\t/ /g;
Often, we need to reuse the portion that was matched in the search pattern in the
replacement string. Suppose that we have a file with lines of the form
phone-number name
which we wish to read and print out in the form
name phone-number
If we match each line against the pattern /\d+\s*\w.*/, then the portion \d+ would
match the phone number, the portion \s* would match all spaces between the phone num-
ber and the first part of the name (which could have many parts) and the portion \w.*
would match the rest of the line, containing all parts of the name. We are interested in
reproducing the phone number and the name, corresponding to the first and third groups
of the pattern, in the output. To do this, we group the portions that we want to capture
within parentheses and then use \1, \2, . . . to recover each of the captured portions. In
particular, if $line contains a line of the form phone-number name, to modify it to the new
form name phone-number we could write
$line =~ s/(\d+)\s*(\w.*)/\2 \1/; # \1 is what \d+ matches,
# \2 is what \w.* matches
One thing to remember is that if we assigned a value to $line using the <> operator,
then it would initially have a trailing newline character. In the search and replace that we
wrote above, this newline character would get included in the pattern \2, so the output
would have a new line between the name and the phone number. The function chomp $line
removes the trailing newline from $line, if it exists, and should always be used to strip off
unwanted newlines when reading data from a file.
170
15.6 Associative arrays
In addition to regular arrays indexed by position, Perl has another collective type called an
associative array, or a hash. A hash is a set of pairs of the form (key,value). The keys of a
hash are strings. Collectively, the name of a hash starts with a %, unlike $ for scalars and @
for arrays. To access an element of a hash we use braces rather than square brackets. The
function keys extracts the keys from a hash table. Here is an example.
$phone{"CMI"} = 28157814;
$phone{"IMSc"} = 22541856;
$phone{"Home"} = 24404396;
foreach $k (keys %phone){
print $k, ":", $phone{$k}, "\n";
}
The output generated by this will be something like
CMI : 28157814
Home : 24404396
IMSc : 22541856
Observe that the sequence in which keys listed out the keys of %phone is not the same
as the sequence in which the keys were added to the hash. This usually happens in a Perl
hash, due to internal optimization.
Here is a typical use of a hash. We read data about students marks in a course from
a file in which each line is of the form rollnumber marks. The file records marks from
multiple evaluations (tests, assignments, . . . ) so each rollnumber appears more than once,
in general. We split each line of input using the builtin function split that splits a string
along whitespace boundaries and returns a list whose elements are the constituents of the
string after splitting. We accumulate the marks for each student in a hash and print out a
summary at the end.
while ($line = <MARKS>){
($rollno, $marks) = split $line; # split $line on whitespace boundary
$eval{$rollno} += $marks; # Add $marks to $eval{$rollno}
}
foreach $k (keys %eval){

print $k, ":", $eval{$k}, "\n";
}
The function split can also take another argument, which is a pattern describing the
delimiter between fields. For instance, the following line splits a line in which values are
separated by commas, where we want to also ignore the spaces before and after commas.
@values = split /\s*,\s*/,$line;
171
15.7 Writing functions
A function in Perl begins with the keyword sub followed by a name. The interesting feature
of functions in Perl is that there is no explicit declaration of the list of arguments to the
function. When a function is invoked, the list of arguments, if any, is made available to the
function in a specially named list, @_. The function can then shift through @_ to extract
all the values that are passed to it.
For instance, here is a function that returns the maximum of exactly two values:
sub max {
my ($a,$b,@rest) = @_; # Get first arguments from @_ in $a, $b
if ($a > $b) {return $a};
return $b;
}
This function does not check whether @_ does in fact have two valid values. No matching
is done while invoking a function, since there is no template associated with a function. Thus,
the function max defined above could be invoked, legally, as max(), or max(1) or max($x,$y)
or max(@list) or . . . . Here is a more generic version of max, that works properly provided
it has at least one argument. Recall that for an array @list, the value @list in scalar context
gives the length of the array.
sub max {
my ($tmp); # Declare $tmp as a local variable
unless (@_) return 0; # Return 0 if input list is empty
my $max = shift @_; # Set $max to first value in @_

while (@_){ # Compare $max to each element in @_
$tmp = shift @_;
if ($tmp > $max){$max = $tmp;}
}
return $max;
}
15.8 Sorting
An extremely useful builtin function in Perl is sort, which sorts an array. The default
behaviour of sort is to use alphabetic sort. So, for instance, to read a list of phone numbers
where each line is of the form name:phone number and print them out in alphabetical order
we could write:
172
while ($line = <PHONE>){
($name,$phone) = split /:/,$line;
$phonehash{$name} = $phone; # Store phone number in a hash
}
foreach $k (sort keys %phonehash){

print $k, ":", $phonehash{$k},"\n");
}
Here, we sort the list generated by keys %phonehash before printout out the values in
the hash.
What if we want to supply a different basis for comparing values? We can supply sort
with the name of a comparison function. Then comparison function we supply will be invoked
by sort. Instead of using @_ to pass parameters to the comparison function (as with normal
functions), sort will pass the values as $a and $b. The comparison function should return
1 if the first argument, $a, is smaller than the second argument, $b$, 0 if the two arguments
are equal, and 1 if the second argument is smaller than the first. So, we can sort using a
numeric comparison function as follows.
foreach $k (sort bynumber keys %somehash){

...
}
sub bynumber {
if ($a < $b) {return -1};
if ($a > $b) {return 1};
return 0;
}
In fact, this is so common that Perl supplies a builtin spaceship operator <=> that has
this effect.
sub bynumber {
return $a <=> $b; # Return -1 if $a < $b,
# +1 if $a > $b,
# 0 if $a == $b
}
The operator cmp achieves the same effect as <=> but using string comparison instead of
numeric comparison.
To sort in descending order, we simply invert the position of the arguments in the com-
parison function.
173
foreach $k (sort bynumdesc keys %somehash){
...
}
sub bynumberdesc {return $b <=> $a;}
We can also use the arguments $a and $b in more complex ways. For instance, to print
out a hash based on the numeric sorted order of its values, we can write:
foreach $k (sort bystrvalue keys %somehash){

...
}
sub bystrvalue {return $somehash{$a} cmp $somehash{$b};}
Finally, we can avoid using a separate named comparison function by just supplying the
expression that is to be evaluated in braces, as follows:
foreach $k (sort {$a <=> $b} keys %somehash){ # Same as bynumber

foreach $k (sort {$b <=> $a} keys %somehash){ # Same as bynumdesc
foreach $k (sort {$somehash{$a} cmp $somehash{$b}}
keys %somehash){ # Same as bystrvalue
174
Part IX
Appendix
175
Appendix A
Some examples of event-driven

programming in Swing
A.1 Multicasting
We want two panels, with three buttons each labelled Red, Blue and Yellow. Clicking a
button should change the colour of both panels to the appropriate colour.
Here, we make both panels listen to all six buttons. However, each panel has a ref-
erence only to the three buttons that are defined within it. Thus, we cannot directly use
getSource() to decode the identity of a button from the other panel.
An alternative solution is to use the ActionCommand associated with a button. In
the ButtonPanel, we set an ActionCommand with each button when we create it. In
ActionPerformed we examine the ActionCommand of the button that was pressed to de-
cide which colour to choose. We do not associate a listener with each button in the panel
when we construct the panel. Instead, we provide a method addListener that adds a lis-
tener to all three buttons in the panel. This is invoked by the surrounding ButtonFrame,
which defines two ButtonPanels and makes both of them listeners for each other.
ButtonPanel.java
import java.awt.*;
private JButton yellowButton;

private JButton blueButton;
177
yellowButton = new JButton("Yellow");
blueButton = new JButton("Blue");
redButton = new JButton("Red");
yellowButton.setActionCommand("YELLOW");
blueButton.setActionCommand("BLUE");
redButton.setActionCommand("RED");
add(yellowButton);
add(blueButton);
add(redButton);
}

Color color = getBackground();
String cmd = evt.getActionCommand(); // Use ActionCommand to
// determine what to do
if (cmd.equals("YELLOW")) color = Color.yellow;

else if (cmd.equals("BLUE")) color = Color.blue;
else if (cmd.equals("RED")) color = Color.red;
repaint();
}
public void addListener(ActionListener o){

yellowButton.addActionListener(o); // Add a commmon listener
blueButton.addActionListener(o); // for all buttons in
redButton.addActionListener(o); // this panel
}
}
ButtonFrame.java
import java.awt.*;
public class ButtonFrame extends JFrame implements WindowListener{

private Container contentPane;
178
private ButtonPanel b1, b2;
public ButtonFrame(){
setTitle("ButtonTest");
setSize(300, 200);
addWindowListener(this);
b1 = new ButtonPanel(); // Create two button panels

b2 = new ButtonPanel();
b1.addListener(b1); // Make each panel a listener for

b1.addListener(b2); // both sets of buttons
b2.addListener(b1);
b2.addListener(b2);
contentPane = this.getContentPane();
contentPane.setLayout(new BorderLayout()); // Set layout to

contentPane.add(b1,"North"); // ensure that
contentPane.add(b2,"South"); // panels dont
// overlap
}
public void windowClosing(WindowEvent e){

System.exit(0);
}
The main program that creates and displays a ButtonFrame is the same as before:
import java.awt.*;
public class ButtonTest

179
{ JFrame frame = new ButtonFrame();
frame.show();
}
}
A.2 Checkbox
A checkbox is a button with a one-bit statea checkbox is either selected or not selected.
In Swing, the basic checkbox class is JCheckBox. Like a button, there is only one action that
a user can perform on a checkboxto click on it. Thus, in Java the listener for JCheckBox
is the same as that for JButtonActionListener. Clicking a checkbox toggles its state.
We use the method isSelected() to determine the current state of a JCheckBox.
In CheckBoxPanel.java below, we define a JPanel with two checkboxes. The panel
CheckBoxPanel serves as an ActionListener for both its checkboxes. When a checkbox is
clicked, it examines the state of both checkboxes to determine which colour to paint the
background.
As with ButtonPanel, we have to embed CheckBoxPanel in a JFrame, called CheckBoxFrame
below, in order to display it. There is no real difference between the code for CheckBoxFrame
and ButtonFrame which we have seen earlier.
Finally, we have a main program that creates and displays a CheckBoxFrame.
CheckBoxPanel.java
import java.awt.*;
public class CheckBoxPanel extends JPanel implements ActionListener{
private JCheckBox redBox;

private JCheckBox blueBox;
public CheckBoxPanel(){
redBox = new JCheckBox("Red");

blueBox = new JCheckBox("Blue");
redBox.addActionListener(this);
blueBox.addActionListener(this);
redBox.setSelected(false);
180
blueBox.setSelected(false);
add(redBox);
add(blueBox);
Color color = getBackground();
if (blueBox.isSelected()) color = Color.blue;

if (redBox.isSelected()) color = Color.red;
if (blueBox.isSelected() && redBox.isSelected()) color = Color.green;
repaint();
}
}
CheckBoxFrame.java
import java.awt.*;
public class CheckBoxFrame extends JFrame implements WindowListener{

private Container contentPane;
public CheckBoxFrame(){
setTitle("BoxTest");
setSize(300, 200);
addWindowListener(this);
contentPane = this.getContentPane();
contentPane.add(new CheckBoxPanel());
public void windowClosing(WindowEvent e){
181
System.exit(0);
}
CheckBoxTest.java
import java.awt.*;
public class CheckBoxTest

{ JFrame frame = new CheckBoxFrame();
frame.show();
}
}
182

Programming Language Concepts: Lecture Notes

Uploaded by

Programming Language Concepts: Lecture Notes

Uploaded by

Programming Language Concepts:

3 Subclasses and inheritance 29

III Event driven programming 67

VI Functional programming 109

12 The (untyped) lambda calculus 125

13 Introducing types into the lambda calculus 141

VII Logic programming 153

VIII Scripting languages 163

Data encapsulation with classes

The starting point of object-oriented programming is to provide a more faithful implemen-

int s[100], top_of_stack = 0;

push (int i, ...){ /* push i onto stack */

int pop (...){ /* pop top of stack and return it */

bool is_empty (...){ /* is the stack empty? */

1.2 Data encapsulation (public and private)

if (s.tos == 0){ ... }

Clearly extend_stack is an implementation detail that is not intended to be revealed

private double deposit-amount;

public double threemonth-yield(){ ... } /* uses base-rate

private static int num_push = 0; /* number of pushes across all

push (int i, ...){ /* push i onto stack */

1.4 Constants (the attribute final)

area = Math.PI * radius * radius;

private int day, month, year;

public Date(int d, int m, int y){

Date d = new Date(27,1,2003);

public Date(int d, int m)

double[] darr = new double[100];

When we invoke Arrays.sort with an argument a, the type of a automatically deter-

// some methods below

no_constructor n = new no_constructor(); // Note the () after

Date d = new Date();

Date d = new Date(27,1,2003);

private int day, month, year;

public Date(int d, int m, int y){

public Date(int d, int m){

private int day, month, year;

public Date(int d, int m, int y){

2.1 Structure of a Java program

2.3 Basic Java syntax

int (4 bytes), long (8 bytes), short (2 bytes)

Some points to note.

Another point is that a typo like if (x = 7) ... will be caught as syntactically

2.3.3 Compound statements

outer_loop: // this is a loop label

String s = "Hello", t = "world";

Here, s.substring(0,3) returns the substring of length 3 starting at position 0. The

s = "A new piece of text";

is an abbreviation for something like

s = new String("A new piece of text");

Creating a new String object implicitly using an assignment statement.

This can also be done as follows (a bit more C like);

int[] a = new int[100];

public class arraycheck{

public static void main(String[] argv){

for (i = 0; i < n; i++){

for (i = 0; i < n; i++){

2.3.4 Final note

Subclasses and inheritance

3.1 Subclasses and superclasses

public Employee(String n){

public boolean setSalary(double x){ ... }

public double getSalary(){ ... }

class Manager extends Employee

public boolean setSecretary(name s){ ... }

public String getSecretary(){ ... }

double bonus(float percent){

3.2 Inheritance polymorphism

for (i = 0; i < emparray.length; i++){

if (emparray[1] instanceof Manager){

3.3 Multiple inheritance

class C3 extends C1, C2{

class C3 extends C1, C2{

3.4 The Java class hierarchy

boolean equals(Object o) // defaults to pointer equality

String toString() // converts the values of the