Notes on Data Structures and Programming

Techniques (CPSC 223, Spring 2015)

James Aspnes

Contents
0.1 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
0.1.1 License . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
0.1.2 Resources . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
0.1.3 Documentation . . . . . . . . . . . . . . . . . . . . . . . . 13
0.1.4 Questions and comments . . . . . . . . . . . . . . . . . . 13
0.2 Lecture schedule . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
0.2.1 Topics by date . . . . . . . . . . . . . . . . . . . . . . . . 13
0.2.2 Topics not covered in 2015 . . . . . . . . . . . . . . . . . 16
0.3 Syllabus . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
0.3.1 On-line course information . . . . . . . . . . . . . . . . . 16
0.3.2 Meeting times . . . . . . . . . . . . . . . . . . . . . . . . . 17
0.3.3 Synopsis of the course . . . . . . . . . . . . . . . . . . . . 17
0.3.4 Prerequisites . . . . . . . . . . . . . . . . . . . . . . . . . 17
0.3.5 Textbook . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
0.3.6 Course requirements . . . . . . . . . . . . . . . . . . . . . 17
0.3.7 Staff . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
0.3.7.1 Instructor . . . . . . . . . . . . . . . . . . . . . . 18
0.3.7.2 Teaching Fellows . . . . . . . . . . . . . . . . . . 18
0.3.7.3 Peer tutors . . . . . . . . . . . . . . . . . . . . . 18
0.3.8 Use of outside help . . . . . . . . . . . . . . . . . . . . . . 18
0.3.9 Clarifications for homework assignments . . . . . . . . . . 18
0.3.10 Late assignments . . . . . . . . . . . . . . . . . . . . . . . 18
0.4 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
0.4.1 Why should you learn to program in C? . . . . . . . . . . 19
0.4.2 Why should you learn about data structures and program-
ming techniques? . . . . . . . . . . . . . . . . . . . . . . . 20

1 The Zoo and the Zoo Annex 20

1.1 Getting an account . . . . . . . . . . . . . . . . . . . . . . . . . . 21
1.2 Getting into the room . . . . . . . . . . . . . . . . . . . . . . . . 21
1.3 Remote use . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

1
 1.3.1 Access using FastX . . . . . . . . . . . . . . . . . . . . . . 21
1.3.1.1 Getting a license key . . . . . . . . . . . . . . . 21
1.3.1.2 FastX in the Zoo Annex . . . . . . . . . . . . . . 21
1.3.1.3 Using FastX from Windows . . . . . . . . . . . . 23
1.3.1.4 Using FastX from OSX . . . . . . . . . . . . . . 24
1.3.2 Terminal access . . . . . . . . . . . . . . . . . . . . . . . . 25
1.3.3 GUI access . . . . . . . . . . . . . . . . . . . . . . . . . . 27
1.4 How to compile and run programs . . . . . . . . . . . . . . . . . 28
1.4.1 Creating the program . . . . . . . . . . . . . . . . . . . . 28
1.4.2 Compiling and running a program . . . . . . . . . . . . . 28
1.4.3 Some notes on what the program does . . . . . . . . . . . 29

2 The Linux programming environment 30

2.1 The shell . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
2.1.1 Getting a shell prompt in the Zoo . . . . . . . . . . . . . 30
2.1.2 The Unix filesystem . . . . . . . . . . . . . . . . . . . . . 31
2.1.3 Unix command-line programs . . . . . . . . . . . . . . . . 31
2.1.4 Stopping and interrupting programs . . . . . . . . . . . . 32
2.1.5 Running your own programs . . . . . . . . . . . . . . . . 33
2.1.6 Redirecting input and output . . . . . . . . . . . . . . . . 33
2.2 Text editors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
2.2.1 Writing C programs with Emacs . . . . . . . . . . . . . . 34
2.2.1.1 My favorite Emacs commands . . . . . . . . . . 35
2.2.2 Using Vi instead of Emacs . . . . . . . . . . . . . . . . . . 36
2.2.2.1 My favorite Vim commands . . . . . . . . . . . . 36
2.2.2.1.1 Normal mode . . . . . . . . . . . . . . . 36
2.2.2.1.2 Insert mode . . . . . . . . . . . . . . . 37
2.2.2.2 Settings . . . . . . . . . . . . . . . . . . . . . . . 38
2.3 Compilation tools . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
2.3.1 The GNU C compiler gcc . . . . . . . . . . . . . . . . . . 38
2.3.2 Make . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
2.3.2.1 Make gotchas . . . . . . . . . . . . . . . . . . . . 40
2.4 Debugging tools . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
2.4.1 Debugging in general . . . . . . . . . . . . . . . . . . . . . 41
2.4.2 Assertions . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
2.4.3 The GNU debugger gdb . . . . . . . . . . . . . . . . . . . 42
2.4.3.1 My favorite gdb commands . . . . . . . . . . . . 44
2.4.3.2 Debugging strategies . . . . . . . . . . . . . . . . 45
2.4.3.3 Common applications of gdb . . . . . . . . . . . 46
2.4.3.3.1 Watching your program run . . . . . . 46
2.4.3.3.2 Dealing with failed assertions . . . . . . 46
2.4.3.3.3 Dealing with segmentation faults . . . . 48
2.4.3.3.4 Dealing with infinite loops . . . . . . . 49
2.4.3.3.5 Mysterious variable changes . . . . . . 50
2.4.4 Valgrind . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
2.4.4.1 Compilation flags . . . . . . . . . . . . . . . . . 53

2
 2.4.4.2 Automated testing . . . . . . . . . . . . . . . . . 53
2.4.4.3 Examples of some common valgrind errors . . . 53
2.4.4.3.1 Uninitialized values . . . . . . . . . . . 53
2.4.4.3.2 Bytes definitely lost . . . . . . . . . . . 54
2.4.4.3.3 Invalid write or read operations . . . . 55
2.4.5 Not recommended: debugging output . . . . . . . . . . . 57
2.5 Performance tuning . . . . . . . . . . . . . . . . . . . . . . . . . . 59
2.5.1 Timing under Linux . . . . . . . . . . . . . . . . . . . . . 59
2.5.2 Profiling with gprof . . . . . . . . . . . . . . . . . . . . . 60
2.5.2.1 Effect of optimization during compilation . . . . 68
2.6 Version control . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
2.6.1 Setting up Git . . . . . . . . . . . . . . . . . . . . . . . . 69
2.6.2 Editing files . . . . . . . . . . . . . . . . . . . . . . . . . . 71
2.6.3 Renaming files . . . . . . . . . . . . . . . . . . . . . . . . 72
2.6.4 Adding and removing files . . . . . . . . . . . . . . . . . . 73
2.6.5 Recovering files from the repository . . . . . . . . . . . . 74
2.6.6 Undoing bad commits . . . . . . . . . . . . . . . . . . . . 74
2.6.7 Looking at old versions . . . . . . . . . . . . . . . . . . . 75
2.6.8 More information about Git . . . . . . . . . . . . . . . . . 76
2.7 Submitting assignments . . . . . . . . . . . . . . . . . . . . . . . 76

3 The C programming language 78

3.1 Structure of a C program . . . . . . . . . . . . . . . . . . . . . . 79
3.2 Numeric data types . . . . . . . . . . . . . . . . . . . . . . . . . . 85
3.2.1 Integer types in C . . . . . . . . . . . . . . . . . . . . . . 87
3.2.1.1 Basic integer types . . . . . . . . . . . . . . . . . 87
3.2.1.2 C99 fixed-width types . . . . . . . . . . . . . . . 89
3.2.2 size_t and ptrdiff_t . . . . . . . . . . . . . . . . . . . 91
3.2.2.1 Integer constants . . . . . . . . . . . . . . . . . . 91
3.2.2.1.1 Naming constants . . . . . . . . . . . . 92
3.2.2.2 Integer operators . . . . . . . . . . . . . . . . . . 94
3.2.2.2.1 Arithmetic operators . . . . . . . . . . 94
3.2.2.2.2 Bitwise operators . . . . . . . . . . . . 95
3.2.2.2.3 Logical operators . . . . . . . . . . . . . 97
3.2.2.2.4 Relational operators . . . . . . . . . . . 97
3.2.2.3 Converting to and from strings . . . . . . . . . . 98
3.2.3 Floating-point types . . . . . . . . . . . . . . . . . . . . . 99
3.2.3.1 Floating point basics . . . . . . . . . . . . . . . 99
3.2.3.2 Floating-point constants . . . . . . . . . . . . . 100
3.2.3.3 Operators . . . . . . . . . . . . . . . . . . . . . . 100
3.2.3.4 Conversion to and from integer types . . . . . . 100
3.2.3.5 The IEEE-754 floating-point standard . . . . . . 101
3.2.3.6 Error . . . . . . . . . . . . . . . . . . . . . . . . 102
3.2.3.7 Reading and writing floating-point numbers . . . 103
3.2.3.8 Non-finite numbers in C . . . . . . . . . . . . . . 104
3.2.3.9 The math library . . . . . . . . . . . . . . . . . 104

3
3.3 Operator precedence . . . . . . . . . . . . . . . . . . . . . . . . . 105
3.4 Programming style . . . . . . . . . . . . . . . . . . . . . . . . . . 106
3.5 Variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
3.5.1 Memory . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
3.5.2 Variables as names . . . . . . . . . . . . . . . . . . . . . . 108
3.5.2.1 Variable declarations . . . . . . . . . . . . . . . 109
3.5.2.2 Variable names . . . . . . . . . . . . . . . . . . . 110
3.5.3 Using variables . . . . . . . . . . . . . . . . . . . . . . . . 112
3.5.4 Initialization . . . . . . . . . . . . . . . . . . . . . . . . . 113
3.5.5 Storage class qualifiers . . . . . . . . . . . . . . . . . . . . 114
3.5.5.1 Scope and extent . . . . . . . . . . . . . . . . . . 114
3.5.5.1.1 Additional qualifiers for global variables 114
3.5.6 Marking variables as constant . . . . . . . . . . . . . . . . 115
3.5.6.1 Pointers to const . . . . . . . . . . . . . . . . . 115
3.6 Input and output . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
3.6.1 Character streams . . . . . . . . . . . . . . . . . . . . . . 116
3.6.2 Reading and writing single characters . . . . . . . . . . . 117
3.6.3 Formatted I/O . . . . . . . . . . . . . . . . . . . . . . . . 118
3.6.4 Rolling your own I/O routines . . . . . . . . . . . . . . . 120
3.6.5 File I/O . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
3.7 Statements and control structures . . . . . . . . . . . . . . . . . . 124
3.7.1 Simple statements . . . . . . . . . . . . . . . . . . . . . . 124
3.7.2 Compound statements . . . . . . . . . . . . . . . . . . . . 124
3.7.2.1 Conditionals . . . . . . . . . . . . . . . . . . . . 125
3.7.2.2 Loops . . . . . . . . . . . . . . . . . . . . . . . . 127
3.7.2.2.1 The while loop . . . . . . . . . . . . . . 127
3.7.2.2.2 The do..while loop . . . . . . . . . . . . 128
3.7.2.2.3 The for loop . . . . . . . . . . . . . . . 129
3.7.2.2.4 Loops with break, continue, and goto . 130
3.7.2.3 Choosing where to put a loop exit . . . . . . . . 131
3.8 Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132
3.8.1 Function definitions . . . . . . . . . . . . . . . . . . . . . 132
3.8.2 When to write a function . . . . . . . . . . . . . . . . . . 134
3.8.3 Calling a function . . . . . . . . . . . . . . . . . . . . . . 136
3.8.4 The return statement . . . . . . . . . . . . . . . . . . . . 136
3.8.5 Function declarations and modules . . . . . . . . . . . . . 137
3.8.6 Static functions . . . . . . . . . . . . . . . . . . . . . . . . 138
3.8.7 Local variables . . . . . . . . . . . . . . . . . . . . . . . . 139
3.8.8 Mechanics of function calls . . . . . . . . . . . . . . . . . 140
3.9 Pointers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141
3.9.1 Memory and addresses . . . . . . . . . . . . . . . . . . . . 141
3.9.2 Pointer variables . . . . . . . . . . . . . . . . . . . . . . . 141
3.9.2.1 Declaring a pointer variable . . . . . . . . . . . . 141
3.9.2.2 Assigning to pointer variables . . . . . . . . . . 142
3.9.2.3 Using a pointer . . . . . . . . . . . . . . . . . . . 142
3.9.2.4 Printing pointers . . . . . . . . . . . . . . . . . . 143

4
 3.9.3 The null pointer . . . . . . . . . . . . . . . . . . . . . . . 144
3.9.4 Pointers and functions . . . . . . . . . . . . . . . . . . . . 144
3.9.5 Pointer arithmetic and arrays . . . . . . . . . . . . . . . . 147
3.9.5.1 Arrays . . . . . . . . . . . . . . . . . . . . . . . 148
3.9.5.2 Arrays and functions . . . . . . . . . . . . . . . 149
3.9.5.3 Multidimensional arrays . . . . . . . . . . . . . . 150
3.9.5.4 Variable-length arrays . . . . . . . . . . . . . . . 153
3.9.6 Void pointers . . . . . . . . . . . . . . . . . . . . . . . . . 156
3.9.6.1 Alignment . . . . . . . . . . . . . . . . . . . . . 157
3.9.7 Run-time storage allocation using malloc . . . . . . . . . 158
3.9.8 Function pointers . . . . . . . . . . . . . . . . . . . . . . . 161
3.9.8.1 Function pointer declarations . . . . . . . . . . . 161
3.9.8.2 Callbacks . . . . . . . . . . . . . . . . . . . . . . 162
3.9.8.3 Dispatch tables . . . . . . . . . . . . . . . . . . . 163
3.9.9 The restrict keyword . . . . . . . . . . . . . . . . . . . . . 164
3.10 Strings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165
3.10.1 C strings . . . . . . . . . . . . . . . . . . . . . . . . . . . 166
3.10.2 String constants . . . . . . . . . . . . . . . . . . . . . . . 166
3.10.3 String buffers . . . . . . . . . . . . . . . . . . . . . . . . . 167
3.10.3.1 String buffers and the perils of gets . . . . . . . 167
3.10.4 Operations on strings . . . . . . . . . . . . . . . . . . . . 169
3.10.5 Finding the length of a string . . . . . . . . . . . . . . . . 171
3.10.5.1 The strlen tarpit . . . . . . . . . . . . . . . . . . 172
3.10.6 Comparing strings . . . . . . . . . . . . . . . . . . . . . . 173
3.10.7 Formatted output to strings . . . . . . . . . . . . . . . . . 174
3.10.8 Dynamic allocation of strings . . . . . . . . . . . . . . . . 174
3.10.9 Command-line arguments . . . . . . . . . . . . . . . . . . 175
3.11 Structured data types . . . . . . . . . . . . . . . . . . . . . . . . 176
3.11.1 Structs . . . . . . . . . . . . . . . . . . . . . . . . . . . . 176
3.11.1.1 Operations on structs . . . . . . . . . . . . . . . 179
3.11.1.2 Layout in memory . . . . . . . . . . . . . . . . . 180
3.11.1.3 Bit fields . . . . . . . . . . . . . . . . . . . . . . 180
3.11.2 Unions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 181
3.11.3 Enums . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 182
3.11.3.1 Specifying particular values . . . . . . . . . . . . 183
3.11.3.2 What most people do . . . . . . . . . . . . . . . 183
3.11.3.3 Using enum with union . . . . . . . . . . . . . . 183
3.12 Type aliases using typedef . . . . . . . . . . . . . . . . . . . . . 184
3.12.1 Opaque structs . . . . . . . . . . . . . . . . . . . . . . . . 185
3.13 Macros . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 187
3.13.1 Macros with arguments . . . . . . . . . . . . . . . . . . . 187
3.13.1.1 Multiple arguments . . . . . . . . . . . . . . . . 188
3.13.1.2 Perils of repeating arguments . . . . . . . . . . . 188
3.13.1.3 Variable-length argument lists . . . . . . . . . . 188
3.13.1.4 Macros vs. inline functions . . . . . . . . . . . . 189
3.13.2 Macros that include other macros . . . . . . . . . . . . . . 190

5
 3.13.3 More specialized macros . . . . . . . . . . . . . . . . . . . 190
3.13.3.1 Multiple expressions in a macro . . . . . . . . . 190
3.13.3.2 Non-syntactic macros . . . . . . . . . . . . . . . 191
3.13.3.3 Multiple statements in one macro . . . . . . . . 191
3.13.3.4 String expansion . . . . . . . . . . . . . . . . . . 192
3.13.3.5 Big macros . . . . . . . . . . . . . . . . . . . . . 193
3.13.4 Conditional compilation . . . . . . . . . . . . . . . . . . . 194
3.13.5 Defining macros on the command line . . . . . . . . . . . 195
3.13.6 The #if directive . . . . . . . . . . . . . . . . . . . . . . . 196
3.13.7 Debugging macro expansions . . . . . . . . . . . . . . . . 196
3.13.8 Can a macro call a preprocessor command? . . . . . . . . 197

4 Data structures and programming techniques 197

4.1 Asymptotic notation . . . . . . . . . . . . . . . . . . . . . . . . . 197
4.1.1 Two sorting algorithms . . . . . . . . . . . . . . . . . . . 197
4.1.2 Big-O to the rescue . . . . . . . . . . . . . . . . . . . . . . 199
4.1.3 Asymptotic cost of programs . . . . . . . . . . . . . . . . 200
4.1.4 Other variants of asymptotic notation . . . . . . . . . . . 202
4.2 Linked lists . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 202
4.2.1 Stacks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 203
4.2.1.1 Building a stack out of an array . . . . . . . . . 206
4.2.2 Queues . . . . . . . . . . . . . . . . . . . . . . . . . . . . 206
4.2.3 Looping over a linked list . . . . . . . . . . . . . . . . . . 210
4.2.4 Looping over a linked list backwards . . . . . . . . . . . . 211
4.2.5 Deques and doubly-linked lists . . . . . . . . . . . . . . . 212
4.2.5.1 Alternate implementation using a ring buffer . . 215
4.2.6 Circular linked lists . . . . . . . . . . . . . . . . . . . . . 220
4.2.7 What linked lists are and are not good for . . . . . . . . . 224
4.2.8 Further reading . . . . . . . . . . . . . . . . . . . . . . . . 224
4.3 Abstract data types . . . . . . . . . . . . . . . . . . . . . . . . . 224
4.3.1 A sequence type . . . . . . . . . . . . . . . . . . . . . . . 225
4.3.1.1 Interface . . . . . . . . . . . . . . . . . . . . . . 226
4.3.1.2 Implementation . . . . . . . . . . . . . . . . . . 227
4.3.1.3 Compiling and linking . . . . . . . . . . . . . . . 228
4.3.2 Designing abstract data types . . . . . . . . . . . . . . . . 230
4.3.2.1 Parnas’s Principle . . . . . . . . . . . . . . . . . 230
4.3.2.2 When to build an abstract data type . . . . . . 231
4.4 Hash tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 232
4.4.1 Dictionary data types . . . . . . . . . . . . . . . . . . . . 232
4.4.2 Basics of hashing . . . . . . . . . . . . . . . . . . . . . . . 233
4.4.3 Resolving collisions . . . . . . . . . . . . . . . . . . . . . . 233
4.4.3.1 Chaining . . . . . . . . . . . . . . . . . . . . . . 233
4.4.3.2 Open addressing . . . . . . . . . . . . . . . . . . 234
4.4.4 Choosing a hash function . . . . . . . . . . . . . . . . . . 234
4.4.4.1 Division method . . . . . . . . . . . . . . . . . . 234
4.4.4.2 Multiplication method . . . . . . . . . . . . . . . 235

6
 4.4.4.3 Universal hashing . . . . . . . . . . . . . . . . . 236
4.4.5 Maintaining a constant load factor . . . . . . . . . . . . . 238
4.4.6 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . 238
4.4.6.1 A low-overhead hash table using open addressing 238
4.4.6.2 A string to string dictionary using chaining . . . 241
4.5 Generic containers . . . . . . . . . . . . . . . . . . . . . . . . . . 246
4.5.1 Generic dictionary: interface . . . . . . . . . . . . . . . . 247
4.5.2 Generic dictionary: implementation . . . . . . . . . . . . 249
4.6 Recursion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 257
4.6.1 Example of recursion in C . . . . . . . . . . . . . . . . . . 257
4.6.2 Common problems with recursion . . . . . . . . . . . . . 261
4.6.2.1 Omitting the base case . . . . . . . . . . . . . . 261
4.6.2.2 Blowing out the stack . . . . . . . . . . . . . . . 262
4.6.2.3 Failure to make progress . . . . . . . . . . . . . 262
4.6.3 Tail-recursion and iteration . . . . . . . . . . . . . . . . . 263
4.6.3.1 Binary search: recursive and iterative versions . 264
4.6.4 Mergesort: a recursive sorting algorithm . . . . . . . . . . 266
4.6.5 Asymptotic complexity of recursive functions . . . . . . . 268
4.7 Binary trees . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 269
4.7.1 Tree basics . . . . . . . . . . . . . . . . . . . . . . . . . . 269
4.7.2 Binary tree implementations . . . . . . . . . . . . . . . . 270
4.7.3 The canonical binary tree algorithm . . . . . . . . . . . . 271
4.7.4 Nodes vs leaves . . . . . . . . . . . . . . . . . . . . . . . . 272
4.7.5 Special classes of binary trees . . . . . . . . . . . . . . . . 272
4.8 Heaps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 273
4.8.1 Priority queues . . . . . . . . . . . . . . . . . . . . . . . . 273
4.8.2 Expensive implementations of priority queues . . . . . . . 273
4.8.3 Structure of a heap . . . . . . . . . . . . . . . . . . . . . . 273
4.8.4 Packed heaps . . . . . . . . . . . . . . . . . . . . . . . . . 274
4.8.5 Bottom-up heapification . . . . . . . . . . . . . . . . . . . 274
4.8.6 Heapsort . . . . . . . . . . . . . . . . . . . . . . . . . . . 275
4.8.7 More information . . . . . . . . . . . . . . . . . . . . . . . 277
4.9 Binary search trees . . . . . . . . . . . . . . . . . . . . . . . . . . 277
4.9.1 Searching for a node . . . . . . . . . . . . . . . . . . . . . 278
4.9.2 Inserting a new node . . . . . . . . . . . . . . . . . . . . . 278
4.9.3 Deleting a node . . . . . . . . . . . . . . . . . . . . . . . . 279
4.9.4 Costs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 280
4.10 Augmented trees . . . . . . . . . . . . . . . . . . . . . . . . . . . 280
4.10.1 Applications . . . . . . . . . . . . . . . . . . . . . . . . . 280
4.11 Balanced trees . . . . . . . . . . . . . . . . . . . . . . . . . . . . 281
4.11.1 Tree rotations . . . . . . . . . . . . . . . . . . . . . . . . . 281
4.11.2 AVL trees . . . . . . . . . . . . . . . . . . . . . . . . . . . 282
4.11.2.1 Sample implementation . . . . . . . . . . . . . . 284
4.11.3 2–3 trees . . . . . . . . . . . . . . . . . . . . . . . . . . . 295
4.11.4 Red-black trees . . . . . . . . . . . . . . . . . . . . . . . . 296
4.11.5 B-trees . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 297

7
 4.11.6 Splay trees . . . . . . . . . . . . . . . . . . . . . . . . . . 299
4.11.6.1 How splaying works . . . . . . . . . . . . . . . . 299
4.11.6.2 Analysis . . . . . . . . . . . . . . . . . . . . . . 300
4.11.6.3 Other operations . . . . . . . . . . . . . . . . . . 301
4.11.6.4 Top-down splaying . . . . . . . . . . . . . . . . . 301
4.11.6.5 An implementation . . . . . . . . . . . . . . . . 304
4.11.6.6 More information . . . . . . . . . . . . . . . . . 311
4.11.7 Scapegoat trees . . . . . . . . . . . . . . . . . . . . . . . . 311
4.11.8 Skip lists . . . . . . . . . . . . . . . . . . . . . . . . . . . 311
4.11.9 Implementations . . . . . . . . . . . . . . . . . . . . . . . 311
4.12 Graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 312
4.12.1 Basic definitions . . . . . . . . . . . . . . . . . . . . . . . 312
4.12.2 Why graphs are useful . . . . . . . . . . . . . . . . . . . . 312
4.12.3 Operations on graphs . . . . . . . . . . . . . . . . . . . . 315
4.12.4 Representations of graphs . . . . . . . . . . . . . . . . . . 315
4.12.4.1 Adjacency matrices . . . . . . . . . . . . . . . . 315
4.12.4.2 Adjacency lists . . . . . . . . . . . . . . . . . . . 315
4.12.4.2.1 An implementation . . . . . . . . . . . 316
4.12.4.3 Implicit representations . . . . . . . . . . . . . . 321
4.12.5 Searching for paths in a graph . . . . . . . . . . . . . . . 321
4.12.5.1 Implementation of depth-first and breadth-first
search . . . . . . . . . . . . . . . . . . . . . . . . 323
4.12.5.2 Combined implementation of depth-first and
breadth-first search . . . . . . . . . . . . . . . . 327
4.12.5.3 Other variations on the basic algorithm . . . . . 335
4.13 Dynamic programming . . . . . . . . . . . . . . . . . . . . . . . . 335
4.13.1 Memoization . . . . . . . . . . . . . . . . . . . . . . . . . 336
4.13.2 Dynamic programming . . . . . . . . . . . . . . . . . . . . 337
4.13.2.1 More examples . . . . . . . . . . . . . . . . . . . 338
4.13.2.1.1 Longest increasing subsequence . . . . . 338
4.13.2.1.2 All-pairs shortest paths . . . . . . . . . 341
4.13.2.1.3 Longest common subsequence . . . . . 342
4.14 Randomization . . . . . . . . . . . . . . . . . . . . . . . . . . . . 345
4.14.1 Generating random values in C . . . . . . . . . . . . . . . 345
4.14.1.1 The rand function from the standard library . . 345
4.14.1.1.1 Supplying a seed with srand . . . . . . 346
4.14.1.2 Better pseudorandom number generators . . . . 347
4.14.1.3 Random numbers without the pseudo . . . . . . 347
4.14.1.4 Range issues . . . . . . . . . . . . . . . . . . . . 348
4.14.2 Randomized algorithms . . . . . . . . . . . . . . . . . . . 351
4.14.2.1 Randomized search . . . . . . . . . . . . . . . . 351
4.14.2.2 Quickselect and quicksort . . . . . . . . . . . . . 352
4.14.3 Randomized data structures . . . . . . . . . . . . . . . . . 355
4.14.3.1 Skip lists . . . . . . . . . . . . . . . . . . . . . . 356
4.14.3.2 Universal hash families . . . . . . . . . . . . . . 360
4.15 String processing . . . . . . . . . . . . . . . . . . . . . . . . . . . 361

8
 4.15.1 Radix search . . . . . . . . . . . . . . . . . . . . . . . . . 362
4.15.1.1 Tries . . . . . . . . . . . . . . . . . . . . . . . . 362
4.15.1.1.1 Searching a trie . . . . . . . . . . . . . 363
4.15.1.1.2 Inserting a new element into a trie . . . 363
4.15.1.1.3 Implementation . . . . . . . . . . . . . 363
4.15.1.2 Patricia trees . . . . . . . . . . . . . . . . . . . . 369
4.15.1.3 Ternary search trees . . . . . . . . . . . . . . . . 371
4.15.1.4 More information . . . . . . . . . . . . . . . . . 374
4.15.2 Radix sort . . . . . . . . . . . . . . . . . . . . . . . . . . . 374
4.15.2.1 Bucket sort . . . . . . . . . . . . . . . . . . . . . 375
4.15.2.2 Classic LSB radix sort . . . . . . . . . . . . . . . 375
4.15.2.3 MSB radix sort . . . . . . . . . . . . . . . . . . . 376
4.15.2.3.1 Issues with recursion depth . . . . . . . 376
4.15.2.3.2 Implementing the buckets . . . . . . . . 376
4.15.2.3.3 Further optimization . . . . . . . . . . 377
4.15.2.3.4 Sample implementation . . . . . . . . . 377

5 Other topics not covered in detail in 2015 380

5.1 More applications of function pointers . . . . . . . . . . . . . . . 380
5.1.1 Iterators . . . . . . . . . . . . . . . . . . . . . . . . . . . . 380
5.1.1.1 Option 1: Function that returns a sequence . . . 381
5.1.1.2 Option 2: Iterator with first/done/next operations382
5.1.1.3 Option 3: Iterator with function argument . . . 383
5.1.2 Closures . . . . . . . . . . . . . . . . . . . . . . . . . . . . 384
5.1.3 Objects . . . . . . . . . . . . . . . . . . . . . . . . . . . . 387
5.2 Suffix arrays . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 390
5.2.1 Why do we want to do this? . . . . . . . . . . . . . . . . . 390
5.2.2 String search algorithms . . . . . . . . . . . . . . . . . . . 390
5.2.3 Suffix trees and suffix arrays . . . . . . . . . . . . . . . . 390
5.2.3.1 Building a suffix array . . . . . . . . . . . . . . . 391
5.2.3.2 Searching a suffix array . . . . . . . . . . . . . . 392
5.3 Burrows-Wheeler transform . . . . . . . . . . . . . . . . . . . . . 392
5.3.1 Suffix arrays and the Burrows-Wheeler transform . . . . . 395
5.3.2 Sample implementation . . . . . . . . . . . . . . . . . . . 395
5.4 C++ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 400
5.4.1 Hello world . . . . . . . . . . . . . . . . . . . . . . . . . . 400
5.4.2 References . . . . . . . . . . . . . . . . . . . . . . . . . . . 401
5.4.3 Function overloading . . . . . . . . . . . . . . . . . . . . . 402
5.4.4 Classes . . . . . . . . . . . . . . . . . . . . . . . . . . . . 403
5.4.5 Operator overloading . . . . . . . . . . . . . . . . . . . . . 406
5.4.6 Templates . . . . . . . . . . . . . . . . . . . . . . . . . . . 408
5.4.7 Exceptions . . . . . . . . . . . . . . . . . . . . . . . . . . 409
5.4.8 Storage allocation . . . . . . . . . . . . . . . . . . . . . . 410
5.4.8.1 Storage allocation inside objects . . . . . . . . . 412
5.4.9 Standard library . . . . . . . . . . . . . . . . . . . . . . . 416
5.4.10 Things we haven’t talked about . . . . . . . . . . . . . . . 417

9
 5.5 Testing during development . . . . . . . . . . . . . . . . . . . . . 417
5.5.1 Unit tests . . . . . . . . . . . . . . . . . . . . . . . . . . . 417
5.5.1.1 What to put in the test code . . . . . . . . . . . 418
5.5.1.2 Example . . . . . . . . . . . . . . . . . . . . . . 418
5.5.2 Test harnesses . . . . . . . . . . . . . . . . . . . . . . . . 422
5.5.2.1 Module interface . . . . . . . . . . . . . . . . . . 422
5.5.2.1.1 stack.h . . . . . . . . . . . . . . . . . . 422
5.5.2.2 Test code . . . . . . . . . . . . . . . . . . . . . . 423
5.5.2.2.1 test-stack.c . . . . . . . . . . . . . . . . 423
5.5.2.3 Makefile . . . . . . . . . . . . . . . . . . . . . . . 425
5.5.2.3.1 Makefile . . . . . . . . . . . . . . . . . . 425
5.5.3 Stub implementation . . . . . . . . . . . . . . . . . . . . . 426
5.5.3.1 stack.c . . . . . . . . . . . . . . . . . . . . . . . 426
5.5.4 Bounded-space implementation . . . . . . . . . . . . . . . 427
5.5.4.1 stack.c . . . . . . . . . . . . . . . . . . . . . . . 427
5.5.5 First fix . . . . . . . . . . . . . . . . . . . . . . . . . . . . 429
5.5.6 Final version . . . . . . . . . . . . . . . . . . . . . . . . . 429
5.5.6.1 stack.c . . . . . . . . . . . . . . . . . . . . . . . 429
5.5.7 Moral . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 431
5.5.8 Appendix: Test macros . . . . . . . . . . . . . . . . . . . 431
5.6 Algorithm design techniques . . . . . . . . . . . . . . . . . . . . . 436
5.6.1 Basic principles of algorithm design . . . . . . . . . . . . 436
5.6.2 Specific techniques . . . . . . . . . . . . . . . . . . . . . . 437
5.6.3 Example: Finding the maximum . . . . . . . . . . . . . . 438
5.6.4 Example: Sorting . . . . . . . . . . . . . . . . . . . . . . . 439
5.7 Bit manipulation . . . . . . . . . . . . . . . . . . . . . . . . . . . 440
5.8 Persistence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 440
5.8.1 A simple solution using text files . . . . . . . . . . . . . . 441
5.8.2 Using a binary file . . . . . . . . . . . . . . . . . . . . . . 442
5.8.3 A version that updates the file in place . . . . . . . . . . 444
5.8.4 An even better version using mmap . . . . . . . . . . . . 445
5.8.5 Concurrency and fault-tolerance issues: ACIDity . . . . . 447

6 What next? 448

6.1 What’s wrong with C . . . . . . . . . . . . . . . . . . . . . . . . 449
6.2 What C++ fixes . . . . . . . . . . . . . . . . . . . . . . . . . . . 450
6.3 Other C-like languages . . . . . . . . . . . . . . . . . . . . . . . . 450
6.4 Scripting languages . . . . . . . . . . . . . . . . . . . . . . . . . . 451

7 Assignments 455
7.1 Assignment 1, due Thursday 2015-01-29, at 11:00pm . . . . . . . 455
7.1.1 Bureaucratic part . . . . . . . . . . . . . . . . . . . . . . . 455
7.1.2 A rotten cipher . . . . . . . . . . . . . . . . . . . . . . . . 455
7.1.3 Your task . . . . . . . . . . . . . . . . . . . . . . . . . . . 456
7.1.4 Hints . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 456
7.1.5 Testing your assignment . . . . . . . . . . . . . . . . . . . 456

10
 7.1.6 Submitting your assignment . . . . . . . . . . . . . . . . . 457
7.1.7 Sample solution . . . . . . . . . . . . . . . . . . . . . . . . 457
7.2 Assignment 2, due Wednesday 2015-02-04, at 11:00pm . . . . . . 458
7.2.1 Opening a safe . . . . . . . . . . . . . . . . . . . . . . . . 458
7.2.2 Submitting your assignment . . . . . . . . . . . . . . . . . 460
7.2.3 Valgrind . . . . . . . . . . . . . . . . . . . . . . . . . . . . 461
7.2.4 Sample solution . . . . . . . . . . . . . . . . . . . . . . . . 461
7.3 Assignment 3, due Wednesday 2015-02-11, at 11:00pm . . . . . . 463
7.3.1 Quadratic letter sequences . . . . . . . . . . . . . . . . . . 463
7.3.2 Your task . . . . . . . . . . . . . . . . . . . . . . . . . . . 464
7.3.3 Submitting your assignment . . . . . . . . . . . . . . . . . 465
7.3.4 Sample solution . . . . . . . . . . . . . . . . . . . . . . . . 465
7.4 Assignment 4, due Wednesday 2015-02-18, at 11:00pm . . . . . . 468
7.4.1 An ASCII art compositor . . . . . . . . . . . . . . . . . . 468
7.4.2 Submitting your assignment . . . . . . . . . . . . . . . . . 469
7.4.3 Notes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 470
7.4.3.1 Input . . . . . . . . . . . . . . . . . . . . . . . . 470
7.4.3.2 Output . . . . . . . . . . . . . . . . . . . . . . . 470
7.4.3.3 General . . . . . . . . . . . . . . . . . . . . . . . 470
7.4.4 Sample solution . . . . . . . . . . . . . . . . . . . . . . . . 471
7.5 Assignment 5, due Wednesday 2015-02-25, at 11:00pm . . . . . . 476
7.5.1 Build a Turing machine! . . . . . . . . . . . . . . . . . . . 476
7.5.2 Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . 477
7.5.3 Your task . . . . . . . . . . . . . . . . . . . . . . . . . . . 477
7.5.4 Submitting your assignment . . . . . . . . . . . . . . . . . 478
7.5.5 Sample solution . . . . . . . . . . . . . . . . . . . . . . . . 478
7.6 Assignment 6, due Wednesday 2015-03-25, at 11:00pm . . . . . . 483
7.6.1 Sinking ships . . . . . . . . . . . . . . . . . . . . . . . . . 483
7.6.2 Things to watch out for . . . . . . . . . . . . . . . . . . . 484
7.6.3 The testShips program . . . . . . . . . . . . . . . . . . . 485
7.6.4 Submitting your assignment . . . . . . . . . . . . . . . . . 487
7.6.5 Provided source files . . . . . . . . . . . . . . . . . . . . . 487
7.6.6 Sample solution . . . . . . . . . . . . . . . . . . . . . . . . 490
7.7 Assignment 7, due Wednesday 2015-04-01, at 11:00pm . . . . . . 496
7.7.1 Solitaire with big cards . . . . . . . . . . . . . . . . . . . 496
7.7.2 Explanation of the testing program . . . . . . . . . . . . . 497
7.7.3 Submitting your assignment . . . . . . . . . . . . . . . . . 498
7.7.4 Sample solution . . . . . . . . . . . . . . . . . . . . . . . . 498
7.8 Assignment 8, due Wednesday 2015-04-08, at 11:00pm . . . . . . 498
7.8.1 An ordered set . . . . . . . . . . . . . . . . . . . . . . . . 498
7.8.2 The testOrderedSet wrapper . . . . . . . . . . . . . . . 499
7.8.3 Submitting your assignment . . . . . . . . . . . . . . . . . 500
7.8.4 Sample solution . . . . . . . . . . . . . . . . . . . . . . . . 500
7.9 Assignment 9, due Wednesday 2015-04-15, at 11:00pm . . . . . . 505
7.9.1 Finding a cycle in a maze . . . . . . . . . . . . . . . . . . 505
7.9.2 Input and output format . . . . . . . . . . . . . . . . . . 505

11
 7.9.3 Submitting and testing your program . . . . . . . . . . . 507
7.9.4 Sample solution . . . . . . . . . . . . . . . . . . . . . . . . 507

8 Common C coding and debugging issues 515

% 2017-10-19T01:11:36-0400 # Course administration {#courseAdministration}

0.1 Overview

This is the course information for CPSC 223: Data Structures and Programming
Techniques for the Spring 2015 semester. This document is available in two
formats, both of which should contain the same information:
• HTML
• PDF
Code examples can be downloaded from links in the text, or can be found in the
examples directory.
The links above point to www.cs.yale.edu. In case this machine is down,
a backup copy of these files can be found at https://www.dropbox.com/sh/
omg9qcxkxeiam2o/AACRAJOTj8af6V7RC1cXBHjQa?dl=0.
This document is a work in progress, and is likely to change frequently as the
semester progresses.

0.1.1 License

Copyright © 2002–2017 by James Aspnes. Distributed under a Creative Commons

Attribution-ShareAlike 4.0 International license: https://creativecommons.org/
licenses/by-sa/4.0/.

0.1.2 Resources

• Schedule: list of lectures and events. Includes reading assignments and

pointers to lecture notes.
• Calendar: shows office hours, lectures, and assignment deadlines.
• Assignments: list of homeworks and exams.
• Notes: notes on various topics relevant to the course.
• Syllabus.
• Piazza. This is a web-based question-and-answer system for communicating
with course staff and other students.
• 2012 web pages: web pages for 2012 version of the course.

12
 • Data structure visualizations. Much better than the ASCII art (at best)
illustrations you will find in this document.1

0.1.3 Documentation

• How to use the Zoo

• Zoo help
• Coding hints
• GNU Online Documentation,
– gcc
– gdb
– ddd
– emacs
• Frequently-asked questions (FAQ) for
– C
– C (abridged)
– Unix
– Emacs
• Programming in C
• Valgrind documentation
• UNIXhelp for Users

0.1.4 Questions and comments

Please feel free to send questions or comments on the class or anything connected
to it to [email protected].
For questions about individual assignments, you may be able to get a faster
response using Piazza. Note that questions you ask there are visible to other
students if not specifically marked private, so be careful about broadcasting your
draft solutions.

0.2 Lecture schedule

0.2.1 Topics by date

2015-01-12 Introduction. What the course is about. Getting started with C:

running the compiler, the main function, integer data types, a few simple
programs. Readings: Course administration, The Zoo and the Zoo Annex,
The Linux programming environment, Structure of a C program, Basic
integer types; Kernighan and Ritchie §§1.1 and 1.2.
1 I would like to thank David Galles for making this site available and Xiao Shi for pointing

me to it.

13
2015-01-14 Arithmetic in C. Readings: Integer constants, Integer operators,
Operator precedence (we didn’t actually do the full operator precedence
table in class, but it’s a nice summary of what operators are available, and
if you don’t like typing parentheses everywhere it’s useful to have an idea
of how precedence and associativity work); K&R §§1.4, 2.2, 2.3, 2.5, 2.6,
2.8, 2.9, and 2.12.
2015-01-16 Local variables and assignment operators. The ++ and -- oper-
ators. More specialized operators: , and ?:. I/O using getchar and
putchar. Control structures: if, switch, while, for. Readings: Vari-
ables, Statements through The for loop; K&R §1.3, 1.5, 2.1, 2.4, and
3.1–3.6.
2015-01-21 Goto-like control structures: break, continue, goto, and
return. Functions. Readings: Rest of [Statements]{#statements},
Functions{#functions}; K&R §3.7, 3.8, 4.1, and 4.5. Examples from
lecture (with a bit of after-the-fact sprucing up) can be found in the
examples directory.
2015-01-26 Start of pointers and arrays: pointer types and pointer variables.
The & and * operators. Using a pointer to get a value out of a function.
Array declarations. Preventing array modification with const. Storage
allocation: malloc, free, and realloc. Readings: Pointers up through
Arrays and functions; K&R §5.1–5.4. Examples.
2015-01-28 More on pointers and arrays: Multi-dimensional arrays, C99
variable-length arrays. Function pointers. Finding storage allocation
bugs using valgrind. Readings: Rest of Pointers, Valgrind; K&R
§§5.6–5.9, 5.11. Examples from lecture: original array-of-pointers-to-arrays
implementation array2dOriginal.c, valgrind-approved version after
removing printArray2D on uninitialized data array2d.c, buggy version
demonstrating how valgrind complains about writing off the end of an
array and doing multiple frees of the same block array2dBad.c. An issue
that came up in lecture is that all of these implementations are a little bit
inefficient in that they allocate a separate block for each row, which means
extra overhead for mallocs data and the possibility that the rows may be
scattered throughout memory, causing issues with virtual memory paging.
Here is yet another implementation that gets space for both the array
of row pointers and all rows with a single call to malloc and then does
pointer arithmetic to slice the space up.2 There is also a generic version of
the simple approach in the section on multidimensional arrays.
2015-02-02 Strings in C: null-terminated strings. What goes wrong if you
forget to put on the null. Various library functions from <string.h> and
how one might implement them. The perils of gets and bounded-size
buffers, and how to use malloc and realloc to avoid them. Meaning of
argv. Readings: Strings; K&R §§5.5, 5.10, B2. Examples from lecture.
2015-02-04 Structured data types: structs, unions, and enums. Type aliases
2 Note that because each row is in the same malloc-ed block as its adjacent rows, valgrind

will not detect if you run off the end of a row in this implementation.

14
 using typedef. Opaque struct definitions and separating interface from
implementation. Readings: Structured data types, typedef; K&R Chapter
6, §2.5 (for enums). Examples from lecture, plus a working version of the
intArray implementation together with the originally stubby intArray.c
and the partial version from the end of lecture.
2015-02-09 More C stuff: Makefiles. Floating-point arithmetic and the math
library. Timing code with a profiler. Basics of file I/O. Readings: Make,
Floating point types, Performance tuning, Input and output; K&R §§2.7
and 7.5, Appendix B4. Examples from lecture.
2015-02-11 Start of data structures: efficiency of different data structures,
linked lists. Readings: Asymptotic notation, Linked lists. Examples from
lecture including improved stack.c and working queue.c.
2015-02-16 Invariants and representation functions for abstract data types.
The deque type, with two implementations: a circular doubly-linked list
and a ring buffer. Readings: Abstract data types, deques. Examples from
lecture.
2015-02-18 Hash Wednesday: set and map data types, hash tables. Readings:
Hash tables. Examples from lecture.
2015-02-23 Various C preprocessor tricks: macros with arguments, string pro-
cessing in macros, non-syntactic macros, conditional compilation. Readings:
Macros; K&R Appendix A12.3. Examples from lecture.
2015-02-25 Polymorphism in C: generic containers and object-oriented pro-
gramming (of a sort) using structs full of function pointers. Example of
using git for version control (you will not be tested on version control).
Readings: Generic containers, version control. Examples from lecture.
Unlike the actual example from lecture, this version really works, instead
of just inserting 0 over and over again.
2015-03-02 Recursion. Readings: Recursion. Examples from lecture.
2015-03-04 Exam 1. This took place at the usual class time (1:00–2:15), and
was a closed-book test potentially covering all material discussed in lecture
prior to the exam. Sample exams from previous years: 2005, 2012. Sample
solutions.
2015-03-23 Binary trees and heaps. Readings: Binary trees, Heaps. Examples
from lecture.
2015-03-25 Binary search tree implementation basics: insertion, search, dele-
tion, various kinds of traversals. Readings: Binary search trees. Examples
from lecture.
2015-03-30 Balanced binary search trees: augmented trees and AVL trees. We
also saw a little bit about red-black trees, but not enough to actually be
useful. Readings: Augmented trees, AVL trees. Example from lecture: see
AVL tree implementation.
2015-04-01 Self-adjusting binary search trees: splay trees, a little bit about
scapegoat trees. Readings: Splay trees. There was no new code in lecture
but we did spend a while looking at a pre-prepared splay tree implementa-
tion.
2015-04-06 Graphs: structure of a graph, graph representations, basic ideas of

15
 graph search. Readings: Graphs up to start of graph search.
2015-04-08 More graphs: depth-first and breadth-first search. Minimum span-
ning trees and shortest paths. Readings: Rest of graphs. BFS example
from lecture. The program I was using in lecture to make nice pictures of
graphs was dot, specifically dot -Tpng. Thanks to the good folks at ITS,
this is now installed in the Zoo along with the rest of the GraphViz tools.
2015-04-13 Dynamic programming: all-pairs shortest paths, longest increasing
subsequence, longest common subsequence. Readings: Dynamic program-
ming. Examples from lecture.
2015-04-15 Randomized data structures. Readings: Randomization. The
devurandom.c example from lecture.
2015-04-20 Data structures for strings: tries, TSTs, and variants; radix sort.
Readings: String processing.
2015-04-22 Exam 2. This took place at the usual class time (1:00–2:15),
and was a closed-book test potential covering all material discussed in
lecture during the semester. Sample exams from previous years: 2005,
2012. Sample solutions.

0.2.2 Topics not covered in 2015

Here are some topics that were not covered specifically this semester but can be
found in the notes.
• Iterators
• Suffix arrays
• Testing during development
• Algorithm design techniques
• Bit manipulation
• Persistence
• C++
• What next

0.3 Syllabus

Syllabus for Computer Science 223b, Data Structures and Programming Tech-
niques. Instructor: James Aspnes.

0.3.1 On-line course information

On-line information about the course, including the lecture schedule, lecture
notes, and information about assignments, can be found at http://www.cs.yale.
edu/homes/aspnes/classes/223/notes.html. This document will be updated
frequently during the semester, and is also available in PDF format.

16
0.3.2 Meeting times

Lectures are MW 13:00–14:15 in WLH 201 (Sudler Hall). The lecture schedule
can be found in the course notes. A calendar is also available.

0.3.3 Synopsis of the course

Topics include programming in C; data structures (arrays, stacks, queues, lists,

trees, heaps, graphs); sorting and searching; storage allocation and management;
data abstraction; programming style; testing and debugging; writing efficient
programs.

0.3.4 Prerequisites

CPSC 201, or equivalent background. See me if you aren’t sure.

0.3.5 Textbook

The textbook for this course is:

• The C Programming Language (2nd Edition), by Brian W. Kernighan and
Dennis M. Ritchie. Prentice Hall, 1988. ISBN 0131103628. The definitive
introduction to C. You should memorize this book.
If you are on the Yale campus or are using VPN to get to Yale’s network, you can
access this book at http://proquest.safaribooksonline.com/book/programming/
c/9780133086249. You do not need to buy a physical copy of this book unless
you want to.

0.3.6 Course requirements

Nine weekly homework assignments, and two in-class exams. Assignments will
be weighted equally in computing the final grade. Each exam will count as three
assignments.

0.3.7 Staff

See the calendar for open office hours.

17
0.3.7.1 Instructor
James Aspnes ([email protected], http://www.cs.yale.edu/homes/
aspnes/). Office: AKW 401. If my open office hours don’t work for you, please
send email to make an appointment.

0.3.7.2 Teaching Fellows

• Yujia Hu [email protected]
• Joshua Lockerman [email protected]
• Junaid Nomani [email protected]

0.3.7.3 Peer tutors

• Iulia Tamas [email protected]
• Dylan Visher [email protected]
• Lining Wang [email protected]
• Blake Woodworth [email protected]

0.3.8 Use of outside help

Students are free to discuss homework problems and course material with each
other, and to consult with the instructor or a TA. Solutions handed in, however,
should be the student’s own work. If a student benefits substantially from
hints or solutions received from fellow students or from outside sources, then
the student should hand in their solution but acknowledge the outside sources,
and we will apportion credit accordingly. Using outside resources in solving a
problem is acceptable but plagiarism is not.

0.3.9 Clarifications for homework assignments

From time to time, ambiguities and errors may creep into homework assignments.
Questions about the interpretation of homework assignments should be sent
to the instructor at [email protected]. Clarifications will appear in the
on-line version of the assignment.

0.3.10 Late assignments

Assignments submitted after the deadline without a Dean’s Excuse are automat-
ically assessed a 2%/hour penalty.

18
0.4 Introduction

There are two purposes to this course: to teach you to program in the C
programming language, and to teach you how to choose, implement, and use
data structures and standard programming techniques.

0.4.1 Why should you learn to program in C?

• It isthe de facto substandard of programming languages.

– C runs on everything.
– C lets you write programs that use very few resources.
– C gives you near-total control over the system, down to the level of
pushing around individual bits with your bare hands.
– C imposes very few constraints on programming style: unlike higher-
level languages, C doesn’t have much of an ideology. There are very
few programs you can’t write in C.
– Many of the programming languages people actually use (Visual Basic,
perl, python, ruby, PHP, etc.) are executed by interpreters written in
C (or C++, an extension to C).
• You will learn discipline.
– C makes it easy to shoot yourself in the foot.
– You can learn to avoid this by being careful about where you point it.
– Pain is a powerful teacher of caution.
• You will fail CS323 if you don’t learn C really well in CS223 (CS majors
only).
On the other hand, there are many reasons why you might not want to use C
later in life. It’s missing a lot of features of modern program languages, including:
• A garbage collector.
• Minimal programmer-protection features like array bounds-checking or a
strong type system.
• Non-trivial built-in data structures.
• Language support for exceptions, namespaces, object-oriented program-
ming, etc.
For most problems where minimizing programmer time and maximizing robust-
ness are more important than minimizing runtime, other languages are a better
choice. But for this class, we’ll be using C.
If you want to read a lot of flaming about what C is or is not good for, see
http://c2.com/cgi/wiki?CeeLanguage.

19
0.4.2 Why should you learn about data structures and programming
techniques?

For small programs, you don’t need much in the way of data structures. But as
soon as you are representing reasonably complicated data, you need some place
to store it. Thinking about how you want to store and organize this data can be
a good framework for organizing the rest of your program.
Many programming environments will give you a rich collection of built-in data
structures as part of their standard library. C does not: unless you use third-
party libraries, any data structure you want in C you will have to build yourself.
For most data structures this will require an understanding of pointers and
storage allocation, mechanisms often hidden in other languages. Understanding
these concepts will give you a deeper understanding of how computers actually
work, and will both let you function in minimalist environments where you don’t
have a lot of support and let you understand what more convenient environments
are doing under their abstraction barriers.
The same applies to the various programming techniques we will discuss in this
class. While some of the issues that come up are specific to C and similar low-
level languages (particular issues involving disciplined management of storage),
some techniques will apply no matter what kinds of programs you are writing
and all will help in understanding what your computer systems are doing even if
some of the details are hidden.

1 The Zoo and the Zoo Annex

The main undergraduate computing facility for Computer Science is the Zoo,
located on the third floor of AKW. The Zoo contains a large number of Linux
workstations.
You don’t need to do your work for this class in the Zoo, but that is where your
assignments will be submitted and tested, so if you do development elsewhere,
you will need to copy your files over and make sure that they work there as well.
The “Zoo Annex” is the informal name for 17HH room 111, which is reserved for
CS student use from 19:00 to 23:59 Sundays through Thursdays. The machines
in 17HH 111 run Windows, but once logged in, you can create a Linux desktop
remotely from a Zoo machine using a program called FastX. You can also
download and use FastX from your own machine to get access to the full Zoo
environment if you are running Windows or OSX.
The best place for information about the Zoo is at http://zoo.cs.yale.edu/. Below
are some points that are of particular relevance for CS223 students.

20
1.1 Getting an account

To get an account in the Zoo, follow the instructions at http://zoo.cs.yale.edu/

accounts.html. You will need your NetID and password to sign up for an account.
Even if you already have an account, you still need to use this form to register
as a CS 223 student, or you will not be able to submit assignments.

1.2 Getting into the room

The Zoo is located on the third floor of Arthur K Watson Hall, toward the
front of the building. If you are a Yale student, your ID should get you into the
building and the room. If you are not a student, you will need to get your ID
validated in AKW 008a to get in after hours.

1.3 Remote use

There are several options for remote use of the Zoo. FastX is the most straight-
forward if you want to replicate the Zoo experience elsewhere. I personally tend
to connect using ssh.

1.3.1 Access using FastX

These instructions are adapted from Stan Eisenstat’s documentation for CS323.
To use FastX, you will need an installation key. If you are downloading the
Windows or OSX version from the Yale Software Library, this will appear on
the download page. Otherwise, see the instructions in the following section.

1.3.1.1 Getting a license key

In order to use FastX, you will need a license key.
1. Go to the Yale Software Library page for FastX (Windows version).
2. Log in with your NetID and password if required.
3. Copy the installation key that appears below the “Download Now” button.

1.3.1.2 FastX in the Zoo Annex

Using FastX you can turn a window on a Windows machine in the Zoo Annex
(aka, the 17 Hillhouse, Room 111, cluster) into the same Linux desktop that you
see when you log into a Zoo node.
Do the following:

21
 1. If you are not logged in:
• Press CTRL + ALT + DELETE. A new screen with a Usage Agree-
ment will appear.
• Click “OK”. A new screen with a NetID and Password box will appear.
• Enter your NetID and Password and click “Continue”. Windows will
log you in.
2. Click on the Windows icon in the lower left-hand corner of the screen. A
new box will appear.
3. Mouse over the “Search programs and files” box in the new window and
type “fastx” (but do not hit return). A list of “Programs” will appear.
4. Click on “FastX” under “Programs” to launch FastX. (If a “Licensing”
window appears asking for an activation key, enter the installation key
from the download page, click “OK”, and click “Close” to dismiss the new
window that appears.)
5. If there is no entry named “Zoo” in the FastX window:
• Click on the green + sign in the upper left-hand corner. A new
window will appear.
• Enter “Zoo” in the Name field.
• Enter “node.zoo.cs.yale.edu” in the Host field.
• Do not change the Port or User field.
• Click “Save” to save the configuration.
6. Double click on the “Zoo” entry. A “Login” box should appear. If a
“Password” box appears instead:
• Click “Cancel”.
• Click on the green pencil in the upper left-hand corner of the FastX
window.
• Delete the contents of the User field. A grayed out “Joe” will appear.
• Click “Save” to save the configuration.
• Double click on the “Zoo” entry again. A “Login” box will appear.
7. Enter your Yale netID in the “Login” box (it should already be there) and
click “Continue”. (If a warning appears asking whether you want to accept
a new key, click on “Accept” to dismiss it.) A “Password” box will appear.
8. Enter your password in the “Password” box and click on “Continue”. A
new “Start new session” entry will appear in the right-hand side of the
FastX window.
9. Click on “Start new session” to produce a pulldown menu and click on
either “Gnome” or “KDE” to choose a window manager (Gnome is the
default in the Zoo) or “Xterm” to open a terminal window.
WARNING: If you click on “Gnome” but the desktop that appears does
not have the usual “Applications” and “Places” in the upper left-hand
corner, then:
1. Type ALT-F2 (that is, press the ALT and F2 keys simultaneously) to
open a command field.
2. Enter the command “gnome-shell –mode=classic -r” (without the
surrounding quotes) and press the “Enter” key.
10. A new window with a Zoo Linux desktop should open. You may resize it

22
 by dragging on the lower left-hand corner if you prefer a different size.
11. When you log out of the Linux desktop, close the “FastX” window by
clicking the red X in the upper right-hand corner.
12. Do not forget to log out of Windows when you are done.

1.3.1.3 Using FastX from Windows

Using FastX you can turn a window on Windows into the same Linux desktop
that you see when you log into a Zoo node.
To install the software on your own Windows machine:
1. Go to the Yale Software Library and click on “Windows”.
2. Scroll through the list and click on “FastX”.
3. Copy the installation key that appears below the “Download Now” button.
4. Click on the “Download Now” button and follow the instructions for
installing the software.
5. Launch FastX.
6. Click on the green + sign in the upper left-hand corner of the FastX
window. In the new window that appears:
• Enter “Zoo” in the Name field.
• Enter “node.zoo.cs.yale.edu” in the Host field.
• Do not change the Port field.
• Enter your Yale netID in the User field.
• Click “Save” to save the configuration.
7. Double click on the “Zoo” entry.
8. When the “Licensing” window appears asking for an activation key, enter
the number copied above and click “OK”. A new window will appear. Click
“Close” to dismiss it.
9. Quit FastX.
If you run into difficulties up to this point, seek help from a student tech or one
of the instructional staff.
Once you have installed the software, do the following to run FastX:
1. Launch FastX.
2. Double click on the “Zoo” entry. A “Password” box will appear.
3. Enter the password for your Yale netID and click “OK”. A “Start a new
session” entry will appear in the FastX window.
4. Click on “Start new session” to produce a pulldown menu and click on
either “Gnome” or “KDE” to choose a window manager (Gnome is the
default in the Zoo) or “Xterm” to open a terminal window.
WARNING: If you click on “Gnome” but the desktop that appears does
not have the usual “Applications” and “Places” in the upper left-hand
corner, then:
1. Type ALT-F2 (that is, press the ALT and F2 keys simultaneously) to
open a command field.

23
 2. Enter the command “gnome-shell –mode=classic -r” (without the
surrounding quotes) and press the “Enter” key.
5. A new window with a Zoo Linux desktop should open. You may resize it
by dragging on the lower left-hand corner if you prefer a different size.
6. When you log out of the Linux desktop, close the “FastX” window by
clicking the red X in the upper right-hand corner.

1.3.1.4 Using FastX from OSX

Using FastX you can turn a window on a Mac into the same Linux desktop that
you see when you log into a Zoo node.
To install the software on your own Mac:
1. Go to the Yale Software Library and click on “Mac”.
2. Scroll through the list and click on “FastX”.
3. Copy the installation key that appears below the “Download Now” button.
4. Click on the “Download Now” button and save the downloaded file as
“fastx_1016.dmg” (for version 1.0.16).
5. Click on the downloaded file “fastx_1016.dmg” to open the disk image. A
“FastX” disk icon will appear on your desktop, and a folder containing a
“FastX” icon will appear.
6. Drag the “FastX” icon to the folder where you would like to save it, and
launch FastX by clicking on it.
7. Click on the green + sign in the upper left-hand corner of the FastX
window. In the new window that appears:
• Enter “Zoo” in the Name field.
• Enter “node.zoo.cs.yale.edu” in the Host field.
• Do not change the Port field.
• Enter your Yale netID in the User field.
• Click “Save” to save the configuration.
8. Double click on the “Zoo” entry.
9. When the “Licensing” window appears asking for an activation key, enter
the number copied above and click “OK”. A new window will appear. Click
“Close” to dismiss it.
10. Quit FastX.
11. Drag the “FastX” disk icon and then the “fastx_1016.dmg” icon to the
Trash to clean up.
If you run into difficulties up to this point, seek help from a student tech or one
of the instructional staff.
Once you have installed the software, do the following to run FastX:
1. Launch FastX from the folder where you saved it.
2. Double click on “Zoo”. A “Password” box will appear, possibly behind the
FastX window. (If it does not, click on the “FastX” icon that is flashing in
your dock.)

24
 3. Enter the password for your Yale netID and click “OK”. A “Start a new
session” entry will appear in the FastX window.
4. Click on “Start new session” to produce a pulldown menu and click on
either “Gnome” or “KDE” to choose a window manager (Gnome is the
default in the Zoo) or “Xterm” to open a terminal window.
WARNING: If you click on “Gnome” but the desktop that appears does
not have the usual “Applications” and “Places” in the upper left-hand
corner, then:
1. Type ALT-F2 (that is, press the ALT and F2 keys simultaneously) to
open a command field.
2. Enter the command “gnome-shell –mode=classic -r” (without the
surrounding quotes) and press the “Enter” key.
5. A new window with a Zoo Linux desktop should open behind the FastX
window. You may resize it by dragging on the lower left-hand corner if
you prefer a different size.
6. When you log out of the Linux desktop, close the “FastX” window by
clicking the red dot in the upper left-hand corner.

1.3.2 Terminal access

Date: Mon, 13 Dec 2004 14:34:19 -0500 (EST)

From: Jim Faulkner <[email protected]>
Subject: Accessing the Zoo

Hello all,

I've been asked to write up a quick guide on how to access the Linux
computers in the Zoo. For those who need this information, please read
on.

There are 2 ways of accessing the Zoo nodes, by walking up to one and
logging in on the console (the computers are located on the 3rd floor of
AKW), or by connecting remotely via SSH. Telnet access is not allowed.
SSH clients for various operating systems are available here:

http://www.yale.edu/software/

Mac OSX comes with an SSH client by default. A good choice for an SSH
client if you run Microsoft Windows is PuTTY:

http://www.chiark.greenend.org.uk/~sgtatham/putty/

With the exception of a few legacy accounts, the Zoo uses your campus-wide
NetID and password for login access. However, you must sign up for a Zoo
account before access is allowed. To sign up for a Zoo account, go to

25
this web page:

http://zoo.cs.yale.edu/accounts.html

Then login with your campus-wide NetID and password. You may choose a
different shell, or set up your account to be enrolled in a class if that
is appropriate for you, but neither is necessary. Just click "Submit".
Within an hour, your Zoo account will be created, and you will receive
more information via e-mail about how to access the Zoo.

Users cannot log into zoo.cs.yale.edu (the central file server) directly,
they must log into one of the Zoo nodes. Following is the list of Zoo
nodes:

aphid.zoo.cs.yale.edu lion.zoo.cs.yale.edu
bumblebee.zoo.cs.yale.edu macaw.zoo.cs.yale.edu
cardinal.zoo.cs.yale.edu monkey.zoo.cs.yale.edu
chameleon.zoo.cs.yale.edu newt.zoo.cs.yale.edu
cicada.zoo.cs.yale.edu peacock.zoo.cs.yale.edu
cobra.zoo.cs.yale.edu perch.zoo.cs.yale.edu
cricket.zoo.cs.yale.edu python.zoo.cs.yale.edu
frog.zoo.cs.yale.edu rattlesnake.zoo.cs.yale.edu
gator.zoo.cs.yale.edu rhino.zoo.cs.yale.edu
giraffe.zoo.cs.yale.edu scorpion.zoo.cs.yale.edu
grizzly.zoo.cs.yale.edu swan.zoo.cs.yale.edu
hare.zoo.cs.yale.edu termite.zoo.cs.yale.edu
hippo.zoo.cs.yale.edu tick.zoo.cs.yale.edu
hornet.zoo.cs.yale.edu tiger.zoo.cs.yale.edu
jaguar.zoo.cs.yale.edu tucan.zoo.cs.yale.edu
koala.zoo.cs.yale.edu turtle.zoo.cs.yale.edu
ladybug.zoo.cs.yale.edu viper.zoo.cs.yale.edu
leopard.zoo.cs.yale.edu zebra.zoo.cs.yale.edu

If you have already created an account, you can SSH directly to one of
the above computers and log in with your campus-wide NetID and
password. You can also SSH to node.zoo.cs.yale.edu, which will connect
you to a random Zoo node.

Feel free to contact me if you have any questions about the Zoo.

thanks,
Jim Faulkner
Zoo Systems Administrator

26
1.3.3 GUI access

(These notes from Debayan Gupta.)

For Mac or Linux users, typing “ssh -X [email protected]” into a
terminal and then running “nautilus” will produce an X window interface.
When on Windows, I usually use XMing (I’ve included a step-by-step guide at
the end of this mail).
For transferring files, I use CoreFTP (http://www.coreftp.com). FileZilla (https:
//filezilla-project.org/) is another option.
Step-by-step guide to XMIng:
You can download Xming from here: http://sourceforge.net/projects/xming/
Download and install. Do NOT launch Xming at the end of your installation.
Once you’ve installed Xming, go to your start menu and find XLaunch (it should
be in the same folder as Xming).
1. Start XLaunch, and select “Multiple Windows”. Leave “Display Number”
as its default value. Click next.
2. Select “Start a program”. Click next.
3. Type “nautilus” (or “terminal”, if you want a terminal) into the “Start
Program” text area. Select “Using PuTTY (plink.exe)”.
4. Type in the name of the computer (use “node.zoo.cs.yale.edu”) in the
“Connect to computer” text box.
5. Type in your netID in the “Login as user” text box (you can leave the
password blank). Click next.
6. Make sure “Clipboard” is ticked. Leave everything else blank. Click next.
7. Click “Save Configuration”. When saving, make sure your filename ends
with “.xlaunch” - this will let you connect with a click (you won’t need to
do all this every time you connect).
8. Click Finish.
9. You will be prompted for your password - enter it. Ignore any security
warnings.
10. You now have a remote connection to the Zoo.
For more options and information, you can go to: http://www.straightrunning.
com/XmingNotes/

27
1.4 How to compile and run programs

See the chapter on how to use the Zoo for details of particular commands. The
basic steps are
• Creating the program with a text editor of your choosing. (I like vim for
long programs and cat for very short ones.)
• Compiling it with gcc.
• Running it.
If any of these steps fail, the next step is debugging. We’ll talk about debugging
elsewhere.

1.4.1 Creating the program

Use your favorite text editor. The program file should have a name of the form
foo.c; the .c at the end tells the C compiler the contents are C source code.
Here is a typical C program:
#include <stdio.h>

/* print the numbers from 1 to 10 */

int
main(int argc, char **argv)
{
int i;

puts("Now I will count from 1 to 10");

for(i = 1; i <= 10; i++) {
printf("%d\n", i);
}

return 0;
}
examples/count.c

1.4.2 Compiling and running a program

Here’s what happens when I compile and run it on the Zoo:

$ c99 -g3 -o count count.c
$ ./count
Now I will count from 1 to 10
1
2

28
3
4
5
6
7
8
9
10
$
The first line is the command to compile the program. The dollar sign is my
prompt, which is printed by the system to tell me it is waiting for a command.
The command calls gcc as c99 with arguments -g3 (enable maximum debugging
info), -o (specify executable file name, otherwise defaults to a.out), count (the
actual executable file name), and count.c (the source file to compile). This
tells gcc that we should compile count.c to count in C99 mode with maximum
debugging info included in the executable file.
The second line runs the output file count. Calling it ./count is necessary
because by default the shell (the program that interprets what you type) only
looks for programs in certain standard system directories. To make it run a
program in the current directory, we have to include the directory name.

1.4.3 Some notes on what the program does

Noteworthy features of this program include:

• The #include <stdio.h> in line 1. This is standard C boilerplate, and
will appear in any program you see that does input or output. The meaning
is to tell the compiler to include the text of the file /usr/include/stdio.h
in your program as if you had typed it there yourself. This particular file
contains declarations for the standard I/O library functions like puts (put
string) and printf (print formatted), as used in the program. If you don’t
put it in, your program may or may not still compile. Do it anyway.
• Line 3 is a comment; its beginning and end is marked by the /* and */
characters. Comments are ignored by the compiler but can be helpful for
other programmers looking at your code (including yourself, after you’ve
forgotten why you wrote something).
• Lines 5 and 6 declare the main function. Every C program has to have a
main function declared in exactly this way—it’s what the operating system
calls when you execute the program. The int on Line 3 says that main
returns a value of type int (we’ll describe this in more detail later in the
chapter on functions), and that it takes two arguments: argc of type int,
the number of arguments passed to the program from the command line,
and argv, of a pointer type that we will get to eventually, which is an array

29
 of the arguments (essentially all the words on the command line, including
the program name). Note that it would also work to do this as one line
(as K&R typically does); the C compiler doesn’t care about whitespace,
so you can format things however you like, subject to the constraint that
consistency will make it easier for people to read your code.
• Everything inside the curly braces is the body of the main function. This
includes
– The declaration int i;, which says that i will be a variable that
holds an int (see the chapter on Integer Types).
– Line 10, which prints an informative message using puts (discussed
in the chapter on input and output.
– The for loop on Lines 11–13, which executes its body for each value
of i from 1 to 10. We’ll explain how for loops work later. Note that
the body of the loop is enclosed in curly braces just like the body
of the main function. The only statement in the body is the call to
printf on Line 12; this includes a format string that specifies that
we want a decimal-formatted integer followed by a newline (the \n).
– The return 0; on Line 15 tells the operating system that the pro-
gram worked (the convention in Unix is that 0 means success). If
the program didn’t work for some reason, we could have returned
something else to signal an error.

2 The Linux programming environment

The Zoo runs a Unix-like operating system called Linux. Most people run Unix
with a command-line interface provided by a shell. Each line typed to the shell
tells it what program to run (the first word in the line) and what arguments
to give it (remaining words). The interpretation of the arguments is up to the
program.

2.1 The shell

When you sign up for an account in the Zoo, you are offered a choice of possible
shell programs. The examples below assume you have chosen bash, the Bourne-
again shell written by the GNU project. Other shells behave similarly for basic
commands.

2.1.1 Getting a shell prompt in the Zoo

When you log in to a Zoo node directly, you may not automatically get a shell
window. If you use the default login environment (which puts you into the KDE

30
window manager), you need to click on the picture of the display with a shell in
from of it in the toolbar at the bottom of the screen. If you run Gnome instead
(you can change your startup environment using the popup menu in the login
box), you can click on the foot in the middle of the toolbar. Either approach will
pop up a terminal emulator from which you can run emacs, gcc, and so forth.
The default login shell in the Zoo is bash, and all examples of shell command
lines given in these notes will assume bash. You can choose a different login
shell on the account sign-up page if you want to, but you are probably best off
just learning to like bash.

2.1.2 The Unix filesystem

Most of what one does with Unix programs is manipulate the filesystem.
Unix files are unstructured blobs of data whose names are given by paths
consisting of a sequence of directory names separated by slashes: for exam-
ple /home/accts/some-user/cs223/hw1.c. At any time you are in a current
working directory (type pwd to find out what it is and cd new-directory to
change it). You can specify a file below the current working directory by
giving just the last part of the pathname. The special directory names .
and .. can also be used to refer to the current directory and its parent. So
/home/accts/some-user/cs223/hw1.c is just hw1.c or ./hw1.c if your current
working directory is /home/accts/some-user/cs223, cs223/hw1.c if your cur-
rent working directory is /home/accts/some-user, and ../cs223/hw1.c if your
current working directory is /home/accts/some-user/illegal-downloads.
All Zoo machines share a common filesystem, so any files you create or change
on one Zoo machine will show up in the same place on all the others.

2.1.3 Unix command-line programs

Here are some handy Unix commands:

man man program will show you the on-line documentation (the man page) for
a program (e.g., try man man or man ls). Handy if you want to know what
a program does. On Linux machines like the ones in the Zoo you can also
get information using info program, which has an Emacs-like interface.
You can also use man function to see documentation for standard library
functions. The command man -k string will search for man pages whose
titles contain string.
Sometimes there is more than one man page with the same name. In this
case man -k will distingiush them by different manual section numbers,
e.g., printf (1) (a shell command) vs. printf (3) (a library routine).
To get a man page from a specific section, use man section name, e.g. man
3 printf.

31
ls ls lists all the files in the current directory. Some useful variants:
• ls /some/other/dir; list files in that directory instead.
• ls -l; long output format showing modification dates and owners.
mkdir mkdir dir will create a new directory in the current directory named
dir.
rmdir rmdir dir deletes a directory. It only works on directories that contain
no files.
cd cd dir changes the current working directory. With no arguments, cd
changes back to your home directory.
pwd pwd (“print working directory”) shows what your current directory is.
mv mv old-name new-name changes the name of a file. You can also use this
to move files between directories.
cp cp old-name new-name makes a copy of a file.
rm rm file deletes a file. Deleted files cannot be recovered. Use this command
carefully.
chmod chmod changes the permissions on a file or directory. See the man page
for the full details of how this works. Here are some common chmod’s:
• chmod 644 file; owner can read or write the file, others can only
read it.
• chmod 600 file; owner can read or write the file, others can’t do
anything with it.
• chmod 755 file; owner can read, write, or execute the file, others
can read or execute it. This is typically used for programs or for
directories (where the execute bit has the special meaning of letting
somebody find files in the directory).
• chmod 700 file; owner can read, write, or execute the file, others
can’t do anything with it.
emacs, gcc, make, gdb, git See corresponding sections.

2.1.4 Stopping and interrupting programs

Sometimes you may have a running program that won’t die. Aside from costing
you the use of your terminal window, this may be annoying to other Zoo users,
especially if the process won’t die even if you close the terminal window or log
out.
There are various control-key combinations you can type at a terminal window
to interrupt or stop a running program.

32
ctrl-C Interrupt the process. Many processes (including any program you write
unless you trap SIGINT using the sigaction system call) will die instantly
when you do this. Some won’t.
ctrl-Z Suspend the process. This will leave a stopped process lying around.
Type jobs to list all your stopped processes, fg to restart the last process
(or fg %1 to start process %1 etc.), bg to keep running the stopped process
in the background, kill %1 to kill process %1 politely, kill -KILL %1 to
kill process %1 whether it wants to die or not.
ctrl-D Send end-of-file to the process. Useful if you are typing test input
to a process that expects to get EOF eventually or writing programs
using cat > program.c (not really recommmended). For test input,
you are often better putting it into a file and using input redirection
(./program < test-input-file); this way you can redo the test after
you fix the bugs it reveals.
ctrl-\ Quit the process. Sends a SIGQUIT, which asks a process to quit and
dump core. Mostly useful if ctrl-C and ctrl-Z don’t work.
If you have a runaway process that you can’t get rid of otherwise, you can use
ps g to get a list of all your processes and their process ids. The kill command
can then be used on the offending process, e.g. kill -KILL 6666 if your evil
process has process id 6666. Sometimes the killall command can simplify
this procedure, e.g. killall -KILL evil kills all process with command name
evil.

2.1.5 Running your own programs

If you compile your own program, you will need to prefix it with ./ on the
command line to tell the shell that you want to run a program in the current
directory (called ‘.’) instead of one of the standard system directories. So for
example, if I’ve just built a program called count, I can run it by typing
$ ./count
Here the “$ ” is standing in for whatever your prompt looks like; you should
not type it.
Any words after the program name (separated by whitespace—spaces and/or
tabs) are passed in as arguments to the program. Sometimes you may wish to
pass more than one word as a single argument. You can do so by wrapping the
argument in single quotes, as in
$ ./count 'this is the first argument' 'this is the second argument'

2.1.6 Redirecting input and output

Some programs take input from standard input (typically the terminal). If
you are doing a lot of testing, you will quickly become tired of typing test input

33
at your program. You can tell the shell to redirect standard input from a file
by putting the file name after a < symbol, like this:
$ ./count < huge-input-file
A ‘>’ symbol is used to redirect standard output, in case you don’t want to
read it as it flies by on your screen:
$ ./count < huge-input-file > huger-output-file
A useful file for both input and output is the special file /dev/null. As input,
it looks like an empty file. As output, it eats any characters sent to it:
$ ./sensory-deprivation-experiment < /dev/null > /dev/null
You can also pipe programs together, connecting the output of one to the input
of the next. Good programs to put at the end of a pipe are head (eats all but the
first ten lines), tail (eats all but the last ten lines), more (lets you page through
the output by hitting the space bar, and tee (shows you the output but also saves
a copy to a file). A typical command might be something like ./spew | more or
./slow-but-boring | tee boring-output. Pipes can consist of a long train
of programs, each of which processes the output of the previous one and supplies
the input to the next. A typical case might be:
$ ./do-many-experiments | sort | uniq -c | sort -nr
which, if ./do-many-experiments gives the output of one experiment on each
line, produces a list of distinct experimental outputs sorted by decreasing fre-
quency. Pipes like this can often substitute for hours of real programming.

2.2 Text editors

To write your programs, you will need to use a text editor, preferably one that
knows enough about C to provide tools like automatic indentation and syntax
highlighting. There are three reasonable choices for this in the Zoo: kate, emacs,
and vim (which can also be run as vi). Kate is a GUI-style editor that comes
with the KDE window system; it plays nicely with the mouse, but Kate skills will
not translate well into other environements. Emacs and Vi have been the two
contenders for the One True Editor since the 1970s—if you learn one (or both)
you will be able to use the resulting skills everywhere. My personal preference is
to use Vi, but Emacs has the advantage of using the same editing commands as
the shell and gdb command-line interfaces.

2.2.1 Writing C programs with Emacs

To start Emacs, type emacs at the command line. If you are actually sitting at
a Zoo node it should put up a new window. If not, Emacs will take over the
current window. If you have never used Emacs before, you should immediately

34
type C-h t (this means hold down the Control key, type h, then type t without
holding down the Control key). This will pop you into the Emacs built-in
tutorial.

2.2.1.1 My favorite Emacs commands

General note: C-x means hold down Control and press x; M-x means hold down
Alt (Emacs calls it “Meta”) and press x. For M-x you can also hit Esc and then
x.
C-h Get help. Everything you could possibly want to know about Emacs is
available through this command. Some common versions: C-h t puts up
the tutorial, C-h b lists every command available in the current mode,
C-h k tells you what a particular sequence of keystrokes does, and C-h l
tells you what the last 50 or so characters you typed were (handy if Emacs
just garbled your file and you want to know what command to avoid in
the future).
C-x u Undo. Undoes the last change you made to the current buffer. Type it
again to undo more things. A lifesaver. Note that it can only undo back
to the time you first loaded the file into Emacs—if you want to be able to
back out of bigger changes, use git (described below).
C-x C-s Save. Saves changes to the current buffer out to its file on disk.
C-x C-f Edit a different file.
C-x C-c Quit out of Emacs. This will ask you if you want to save any buffers
that have been modified. You probably want to answer yes (y) for each
one, but you can answer no (n) if you changed some file inside Emacs but
want to throw the changes away.
C-f Go forward one character.
C-b Go back one character.
C-n Go to the next line.
C-p Go to the previous line.
C-a Go to the beginning of the line.
C-k Kill the rest of the line starting with the current position. Useful Emacs
idiom: C-a C-k.
C-y “Yank.” Get back what you just killed.
TAB Re-indent the current line. In C mode this will indent the line according
to Emacs’s notion of how C should be indented.
M-x compile Compile a program. This will ask you if you want to save out
any unsaved buffers and then run a compile command of your choice
(see the section on compiling programs below). The exciting thing about
M-x compile is that if your program has errors in it, you can type C-x ‘
to jump to the next error, or at least where gcc thinks the next error is.

35
2.2.2 Using Vi instead of Emacs

If you don’t find yourself liking Emacs very much, you might want to try Vim
instead. Vim is a vastly enhanced reimplementation of the classic vi editor,
which I personally find easier to use than Emacs. Type vimtutor to run the
tutorial.
One annoying feature of Vim is that it is hard to figure out how to quit. If you
don’t mind losing all of your changes, you can always get out by hitting the
Escape key a few times and then typing ~ \\\ :qa!\\\ ~
To run Vim, type vim or vim filename from the command line. Or you can use
the graphical version gvim, which pops up its own window.
Vim is a modal editor, meaning that at any time you are in one of several modes
(normal mode, insert mode, replace mode, operator-pending mode, etc.), and
the interpretation of keystrokes depends on which mode you are in. So typing
jjjj in normal mode moves the cursor down four lines, while typing jjjj in
insert mode inserts the string jjjj at the current position. Most of the time
you will be in either normal mode or insert mode. There is also a command
mode entered by hitting : that lets you type longer commands, similar to the
Unix command-line or M-x in Emacs.

2.2.2.1 My favorite Vim commands

2.2.2.1.1 Normal mode

:h Get help. (Hit Enter at the end of any command that starts with a colon.)
Escape
Get out of whatever strange mode you are in and go back to normal mode.
You will need to use this whenever you are done typing code and want to
get back to typing commands.
i Enter insert mode. You will need to do this to type anything. The command a
also enters insert mode, but puts new text after the current cursor position
instead of before it. u
Undo. Undoes the last change you made to the current buffer. Type it
again to undo more things. If you undid something by mistake, c-R (control
R) will redo the last undo
(and can also be repeated). :w
Write the current file to disk. Use :w filename to write it to filename.
Use :wa to write all files that you have modified. The command ZZ does
the
same thing without having to hit Enter at the end. :e filename

36
 Edit a different file.
:q Quit. Vi will refuse to do this if you have unwritten files. See :wa for how
to fix this, or use :q! if you want to throw away your changes and quit
anyway. The shortcuts :x and :wq do a write of the current file followed
by quitting.
h, j, k, l Move the cursor left, down, up, or right. You can also use the arrow
keys (in both normal mode and insert mode).
x Delete the current character.
D Delete to end of line.
dd Delete all of the current line. This is a special case of a more general d
command. If you precede it with a number, you can delete multiple lines:
5dd deletes the next 5 lines. If you replace the second d with a motion
command, you delete until wherever you land: d$ deletes to end of line
(D is faster), dj deletes this line and the line after it, d% deletes the next
matching group of parantheses/braces/brackets and whatever is between
them, dG deletes to end of file—there are many possibilities. All of these
save what you deleted into register "" so you can get them back with p.
yy Like dd, but only saves the line to register "" and doesn’t delete it. (Think
copy). All the variants of dd work with yy: 5yy, y$, yj, y%, etc.
p Pull whatever is in register "". (Think paste).
<< and >> Outdent or indent the current line one tab stop.
:make Run make in the current directory. You can also give it arguments, e.g.,
:make myprog, :make test. Use :cn to go to the next error if you get
errors.
:! Run a command, e.g., :! echo hello world or :! gdb myprogram. Returns
to Vim when the command exits (control-C can sometimes be helpful if
your command isn’t exiting when it should). This works best if you ran
Vim from a shell window; it doesn’t work very well if Vim is running in its
own window.

2.2.2.1.2 Insert mode

control-P and control-N These are completion commands that attempt
to expand a partial word to something it matches elsewhere in the
buffer. So if you are a good person and have named a variable
informativeVariableName instead of ivn, you can avoid having to type
the entire word by typing inf<control-P> if it’s the only word in your
buffer that starts with inf.
control-O and control-I Jump to the last cursor position before a big move /
back to the place you jumped from.

37
ESC Get out of insert mode!

2.2.2.2 Settings
Unlike Emacs, Vim’s default settings are not very good for editing C programs.
You can fix this by creating a file called .vimrc in your home directory with the
following commands:
set shiftwidth=4
set autoindent
set backup
set cindent
set hlsearch
set incsearch
set showmatch
set number
syntax on
filetype plugin on
filetype indent on
examples/sample.vimrc
(You can download this file by clicking on the link.)
In Vim, you can type e.g. :help backup to find out what each setting does.
Note that because .vimrc starts with a ., it won’t be visible to ls unless you
use ls -a or ls -A.

2.3 Compilation tools

2.3.1 The GNU C compiler gcc

A C program will typically consist of one or more files whose names end with .c.
To compile foo.c, you can type gcc foo.c. Assuming foo.c contains no errors
egregious enough to be detected by the extremely forgiving C compiler, this will
produce a file named a.out that you can then execute by typing ./a.out.
If you want to debug your program using gdb or give it a different name,
you will need to use a longer command line. Here’s one that compiles foo.c
to foo (run it using ./foo) and includes the information that gdb needs:
gcc -g3 -o foo foo.c
If you want to use C99 features, you will need to tell gcc to use C99 instead
of its own default dialect of C. You can do this either by adding the argument
-std=c99 as in gcc -std=c99 -o foo foo.c or by calling gcc as c99 as in c99
-o foo foo.c.

38
By default, gcc doesn’t check everything that might be wrong with your program.
But if you give it a few extra arguments, it will warn you about many (but not
all) potential problems: c99 -g3 -Wall -pedantic -o foo foo.c.

2.3.2 Make

For complicated programs involving multiple source files, you are probably better
off using make than calling gcc directly. Make is a “rule-based expert system”
that figures out how to compile programs given a little bit of information about
their components.
For example, if you have a file called foo.c, try typing make foo and see what
happens.
In general you will probably want to write a Makefile, which is named Makefile
or makefile and tells make how to compile programs in the same directory. Here’s
a typical Makefile:
# Any line that starts with a sharp is a comment and is ignored
# by Make.

# These lines set variables that control make's default rules.

# We STRONGLY recommend putting "-Wall -pedantic -g3" in your CFLAGS.
CC=gcc
CFLAGS=-std=c99 -Wall -pedantic -g3

# The next line is a dependency line.

# It says that if somebody types "make all"
# make must first make "hello-world".
# By default the left-hand-side of the first dependency is what you
# get if you just type "make" with no arguments.
all: hello-world

# How do we make hello-world?

# The dependency line says you need to first make hello-world.o
# and hello-library.o
hello-world: hello-world.o hello-library.o
# Subsequent lines starting with a TAB character give
# commands to execute.
# This command uses make built-in variables to avoid
# retyping (and getting things wrong):
# $@ = target hello-world
# $^ = dependencies hello-world.o and hello-library.o
$(CC) $(CFLAGS) -o $@ $^
# You can put whatever commands you want.
echo "I just built hello-world! Hooray!"

39
Figure 6: DFS tree
329
Figure 7: BFS tree
330
 * struct searchInfo *s;
* int n;
*
* s = searchInfoCreate(g);
*
* n = graph_vertices(g);
* for(i = 0; i < n; i++) {
* dfs(s, i);
* }
*
* ... use results in s ...
*
* searchInfoDestroy(s);
*
*/

/* summary information per node for dfs and bfs */

/* this is not intended to be opaque---user can read it */
/* (but should not write it!) */

#define SEARCH_INFO_NULL (-1) /* for empty slots */

struct searchInfo {
Graph graph;
int reached; /* count of reached nodes */
int *preorder; /* list of nodes in order first reached */
int *time; /* time[i] == position of node i in preorder list */
int *parent; /* parent in DFS or BFS forest */
int *depth; /* distance from root */
};

/* allocate and initialize search results structure */

/* you need to do this before passing it to dfs or bfs */
struct searchInfo *searchInfoCreate(Graph g);

/* free searchInfo data---does NOT free graph pointer */

void searchInfoDestroy(struct searchInfo *);

/* perform depth-first search starting at root, updating results */

void dfs(struct searchInfo *results, int root);

/* perform breadth-first search starting at root, updating results */

void bfs(struct searchInfo *results, int root);
examples/graphs/genericSearch.h
#include <stdlib.h>

331
#include <assert.h>

#include "graph.h"
#include "genericSearch.h"

/* create an array of n ints initialized to SEARCH_INFO_NULL */

static int *
createEmptyArray(int n)
{
int *a;
int i;

a = malloc(sizeof(*a) * n);
assert(a);

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

a[i] = SEARCH_INFO_NULL;
}

return a;
}

/* allocate and initialize search results structure */

/* you need to do this before passing it to dfs or bfs */
struct searchInfo *
searchInfoCreate(Graph g)
{
struct searchInfo *s;
int n;

s = malloc(sizeof(*s));
assert(s);

s->graph = g;
s->reached = 0;

n = graphVertexCount(g);

s->preorder = createEmptyArray(n);
s->time = createEmptyArray(n);
s->parent = createEmptyArray(n);
s->depth = createEmptyArray(n);

return s;
}

332
/* free searchInfo data---does NOT free graph pointer */
void
searchInfoDestroy(struct searchInfo *s)
{
free(s->depth);
free(s->parent);
free(s->time);
free(s->preorder);
free(s);
}

/* used inside search routines */

struct edge {
int u; /* source */
int v; /* sink */
};

/* stack/queue */
struct queue {
struct edge *e;
int bottom;
int top;
};

static void
pushEdge(Graph g, int u, int v, void *data)
{
struct queue *q;

q = data;

assert(q->top < graphEdgeCount(g) + 1);

q->e[q->top].u = u;
q->e[q->top].v = v;
q->top++;
}

/* this rather horrible function implements dfs if useQueue == 0 */

/* and bfs if useQueue == 1 */
static void
genericSearch(struct searchInfo *r, int root, int useQueue)
{
/* queue/stack */
struct queue q;

333
 /* edge we are working on */
struct edge cur;

/* start with empty q */

/* we need one space per edge */
/* plus one for the fake (root, root) edge */
q.e = malloc(sizeof(*q.e) * (graphEdgeCount(r->graph) + 1));
assert(q.e);

q.bottom = q.top = 0;

/* push the root */

pushEdge(r->graph, root, root, &q);

/* while q.e not empty */

while(q.bottom < q.top) {
if(useQueue) {
cur = q.e[q.bottom++];
} else {
cur = q.e[--q.top];
}

/* did we visit sink already? */

if(r->parent[cur.v] != SEARCH_INFO_NULL) continue;

/* no */
assert(r->reached < graphVertexCount(r->graph));
r->parent[cur.v] = cur.u;
r->time[cur.v] = r->reached;
r->preorder[r->reached++] = cur.v;
if(cur.u == cur.v) {
/* we could avoid this if we were certain SEARCH_INFO_NULL */
/* would never be anything but -1 */
r->depth[cur.v] = 0;
} else {
r->depth[cur.v] = r->depth[cur.u] + 1;
}

/* push all outgoing edges */

graphForeach(r->graph, cur.v, pushEdge, &q);
}

free(q.e);
}

void

334
dfs(struct searchInfo *results, int root)
{
genericSearch(results, root, 0);
}

void
bfs(struct searchInfo *results, int root)
{
genericSearch(results, root, 1);
}
examples/graphs/genericSearch.c
And here is some test code: genericSearchTest.c. You will need to compile
genericSearchTest.c together with both genericSearch.c and graph.c to
get it to work. This Makefile will do this for you.

4.12.5.3 Other variations on the basic algorithm

Stacks and queues are not the only options for the bucket in the generic search
algorithm. Some other choices are:
• A priority queue keyed by edge weights. If the edges have weights, the
generic tree-builder can be used to find a tree containing s with minimum
total edge weight.22 The basic idea is to always pull out the lightest edge.
The resulting algorithm runs in O(n + m log m) time (since each heap
operation takes O(log m) time), and is known as Prim’s algorithm. See
Prim’s algorithm for more details.
• A priority queue keyed by path lengths. Here we assume that edges have
lengths, and we want to build a shortest-path tree where the length of
the path is no longer just the number of edges it contains but the sum of
their weights. The basic idea is to keep track of the distance from the root
to each node in the tree, and assign each edge a key equal to the sum of
the distance to its source and its length. The resulting search algorithm,
known as Dijkstra’s algorithm, will give a shortest-path tree if all the
edge weights are non-negative. See Dijkstra’s algorithm.

4.13 Dynamic programming

Dynamic programming is a general-purpose algorithm design technique that

is most often used to solve combinatorial optimization problems, where
we are looking for the best possible input to some function chosen from an
exponentially large search space.
22 This only works if the graph is undirected, which means that for every edge uv there is a

matching edge vu with the same weight.

335
There are two parts to dynamic programming. The first part is a programming
technique: dynamic programming is essentially divide and conquer run in reverse:
we solve a big instance of a problem by breaking it up recursively into smaller
instances; but instead of carrying out the computation recursively from the top
down, we start from the bottom with the smallest instances of the problem,
solving each increasingly large instance in turn and storing the result in a
table. The second part is a design principle: in building up our table, we are
careful always to preserve alternative solutions we may need later, by delaying
commitment to particular choices to the extent that we can.
The bottom-up aspect of dynamic programming is most useful when a straight-
forward recursion would produce many duplicate subproblems. It is most efficient
when we can enumerate a class of subproblems that doesn’t include too many
extraneous cases that we don’t need for our original problem.
To take a simple example, suppose that we want to compute the n-th Fibonacci
number using the defining recurrence
• F (n) = F (n − 1) + F (n − 2)
• F (1) = F (0) = 1.
A naive approach would simply code the recurrence up directly:
int
fib(int n)
{
if(n < 2) {
return 1
} else {
return fib(n-1) + fib(n-2);
}
}
The running time of this procedure is easy to compute. The recurrence is
• T (n) = T (n − 1) + T (n − 2) + Θ(1),
whose solution is Θ(an ) where a is the golden ratio 1.6180339887498948482 . . ..
This is badly exponential.23

4.13.1 Memoization

The problem is that we keep recomputing values of fib that we’ve already
computed. We can avoid this by memoization, where we wrap our recursive
solution in a memoizer that stores previously-computed solutions in a hash
23 But it’s linear in the numerical value of the output, which means that fib(n) will actually

terminate in a reasonable amount of time on a typical modern computer when run on any n
small enough that F (n) fits in 32 bits. Running it using 64-bit (or larger) integer representations
will be slower.

336
table. Sensible programming languages will let you write a memoizer once and
apply it to arbitrary recursive functions. In less sensible programming languages
it is usually easier just to embed the memoization in the function definition itself,
like this:
int
memoFib(int n)
{
int ret;

if(hashContains(FibHash, n)) {
return hashGet(FibHash, n);
} else {
ret = memoFib(n-1) + memoFib(n-2);
hashPut(FibHash, n, ret);
return ret;
}
}
The assumption here is that FibHash is a global hash table that we have initialized
to map 0 and 1 to 1. The total cost of running this procedure is O(n), because
fib is called at most twice for each value k in 0 . . . n.
Memoization is a very useful technique in practice, but it is not popular with
algorithm designers because computing the running time of a complex memoized
procedure is often much more difficult than computing the time to fill a nice
clean table. The use of a hash table instead of an array may also add overhead
(and code complexity) that comes out in the constant factors. But it is always
the case that a memoized recursive procedure considers no more subproblems
than a table-based solution, and it may consider many fewer if we are sloppy
about what we put in our table (perhaps because we can’t easily predict what
subproblems will be useful).

4.13.2 Dynamic programming

Dynamic programming comes to the rescue. Because we know what smaller cases
we have to reduce F(n) to, instead of computing F(n) top-down, we compute it
bottom-up, hitting all possible smaller cases and storing the results in an array:
int
fib2(int n)
{
int *a;
int i;
int ret;

if(n < 2) {

337
 return 1;
} else {
a = malloc(sizeof(*a) * (n+1));
assert(a);

a[1] = a[2] = 1;

for(i = 3; i <= n; i++) {

a[i] = a[i-1] + a[i-2];
}
}

ret = a[n];
free(a);
return ret;
}
Notice the recurrence is exactly the same in this version as in our original
recursive version, except that instead of computing F(n-1) and F(n-2) recursively,
we just pull them out of the array. This is typical of dynamic-programming
solutions: often the most tedious editing step in converting a recursive algorithm
to dynamic programming is changing parentheses to square brackets. As with
memoization, the effect of this conversion is dramatic; what used to be an
exponential-time algorithm is now linear-time.

4.13.2.1 More examples

4.13.2.1.1 Longest increasing subsequence

Suppose that we want to compute the longest increasing subsequence of
an array. This is a sequence, not necessarily contiguous, of elements from the
array such that each is strictly larger than the one before it. Since there are 2n
different subsequences of an n-element array, the brute-force approach of trying
all of them might take a while.
What makes this problem suitable for dynamic programming is that any prefix
of a longest increasing subsequence is a longest increasing subsequence of the
part of the array that ends where the prefix ends; if it weren’t, we could
make the big sequence longer by choosing a longer prefix. So to find the
longest increasing subsequence of the whole array, we build up a table of longest
increasing subsequences for each initial prefix of the array. At each step, when
finding the longest increasing subsequence of elements 0 . . . i, we can just scan
through all the possible values for the second-to-last element and read the length
of the best possible subsequence ending there out of the table. When the table
is complete, we can scan for the best last element and then work backwards to
reconstruct the actual subsequence.

338
This last step requires some explanation. We don’t really want to store in
table[i] the full longest increasing subsequence ending at position i, because it
may be very big. Instead, we store the index of the second-to-last element of this
sequence. Since that second-to-last element also has a table entry that stores the
index of its predecessor, by following the indices we can generate a subsequence
of length O(n), even though we only stored O(1) pieces of information in each
table entry. This is similar to the parent pointer technique used in graph search
algorithms.
Here’s what the code looks like:
/* compute a longest strictly increasing subsequence of an array of ints */
/* input is array a with given length n */
/* returns length of LIS */
/* If the output pointer is non-null, writes LIS to output pointer. */
/* Caller should provide at least sizeof(int)*n space for output */
/* If there are multiple LIS's, which one is returned is arbitrary. */
unsigned long
longest_increasing_subsequence(const int a[], unsigned long n, int *output);
examples/dynamicProgramming/lis/lis.h
#include <stdlib.h>
#include <assert.h>

#include "lis.h"

unsigned long
longest_increasing_subsequence(const int a[], unsigned long n, int *output)
{
struct lis_data {
unsigned long length; /* length of LIS ending at this point */
unsigned long prev; /* previous entry in the LIS ending at this point
} *table;

unsigned long best; /* best entry in table */

unsigned long scan; /* used to generate output */

unsigned long i;
unsigned long j;
unsigned long best_length;

/* special case for empty table */

if(n == 0) return 0;

table = malloc(sizeof(*table) * n);

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

339
 /* default best is just this element by itself */
table[i].length = 1;
table[i].prev = n; /* default end-of-list value */

/* but try all other possibilities */

for(j = 0; j < i; j++) {
if(a[j] < a[i] && table[j].length + 1 > table[i].length) {
/* we have a winner */
table[i].length = table[j].length + 1;
table[i].prev = j;
}
}
}

/* now find the best of the lot */

best = 0;

for(i = 1; i < n; i++) {

if(table[i].length > table[best].length) {
best = i;
}
}

/* table[best].length is now our return value */

/* save it so that we don't lose it when we free table */
best_length = table[best].length;

/* do we really have to compute the output? */

if(output) {
/* yes :-( */
scan = best;
for(i = 0; i < best_length; i++) {
assert(scan >= 0);
assert(scan < n);

output[best_length - i - 1] = a[scan];

scan = table[scan].prev;
}
}

free(table);

return best_length;
}

340
examples/dynamicProgramming/lis/lis.c
A sample program that runs longest_increasing_subsequence on a list of
numbers passed in by stdin is given in test_lis.c. Here is a Makefile.
Implemented like this, the cost of finding an LIS is O(n2 ), because to compute
each entry in the array, we have to search through all the previous entries to
find the longest path that ends at a value less than the current one. This can be
improved by using a more clever data structure. If we use a binary search tree
that stores path keyed by the last value, and augment each node with a field
that represents the maximum length of any path in the subtree under that node,
then we can find the longest feasible path that we can append the current node
to in O(log n) time instead of O(n) time. This brings the total cost down to
only O(n log n).

4.13.2.1.2 All-pairs shortest paths

Suppose we want to compute the distance between any two points in a graph,
where each edge uv has a length `u v (+∞ for edges not in the graph) and
the distance between two vertices s and t$ is the minimum over all s–t paths
of the total length of the edges. There are various algorithms for doing this
for a particular s and t, but there is also a very simple dynamic programming
algorithm known as Floyd-Warshall that computes the distance between all
n2 pairs of vertices in Θ(n3 ) time.
The assumption is that the graph does not contain a negative cycle (a cycle
with total edge weight less than zero), so that for two connected nodes there
is always a shortest path that uses each intermediate vertex at most once. If a
graph does contain a negative cycle, the algorithm will detect it by reporting
the distance from i to i less than zero for some i.
Negative cycles don’t generally exist in distance graphs (unless you have the
ability to move faster than the speed of light), but they can come up in other
contexts. One example would be in currency arbitrage, where each node is some
currency, the weight of an edge uv is the logarithm of the exchange rate from
u to v, and the total weight of a path from s to t gives the logarithm of the
number of units of t you can get for one unit of s, since adding the logs along
the path corresponds to multiplying all the exchange rates. In this context a
negative cycle gives you a way to turn a dollar into less than a dollar by running
it through various other currencies, which is not useful, but a positive cycle lets
you pay for the supercomputer you bought to find it before anybody else did. If
we negate all the edge weights, we turn a positive cycle into a negative cycle,
making a fast algorithm for finding this negative cycle potentially valuable.
However, if we don’t have any negative cycles, the idea is that we can create
restricted instances of the shortest-path problem by limiting the maximum index
of any node used on the path. Let L(i, j, k) be the length of a shortest path from
i to j that uses only the vertices 0, . . . , k − 1 along the path (not counting the

341
endpoints i and j, which can be anything). When k = 0, this is just the length
of the i–j edge, or +∞ if there is no such edge. So we can start by computing
L(i, j, 0) for all i. Now given L(i, j, k) for all i and some k, we can compute
L(i, j, k + 1) by observing that any shortest i–j path that has intermediate
vertices in 0 . . . k either consists of a path with intermediate vertices in 0 . . . k − 1,
or consists of a path from i to k followed by a path from k to j, where both of
these paths have intermediate vertices in 0 . . . k − 1. So we get
• L(i, j, k + 1) = min(L(i, j, k), L(i, k, k) + L(k, j, k).
Since this takes O(1) time to compute if we have previously computed L(i, j, k)
for all i and j, we can fill in the entire table in O(n3 ) time.
Implementation details:
• If we want to reconstruct the shortest path in addition to computing its
length, we can store the first vertex for each i–j path. This will either be
(a) the first vertex in the i–j path for the previous k, or (b) the first vertex
in the i–k path.
• We don’t actually need to use a full three-dimensional array. It’s enough
to store one value for each pair i, j and let k be implicit. At each step we
let L[i][j] be min(L[i][j], L[i][k] + L[k][j]). The trick is that we don’t care
if L[i][k] or L[k][j] has already been updated, because that will only give
us paths with a few extra k vertices, which won’t be the shortest paths
anyway assuming no negative cycles.

4.13.2.1.3 Longest common subsequence

Given sequences of characters v and w, v is a subsequence of w if every character
in v appears in w in the same order. For example, aaaaa, brac, and badar
are all subsequences of abracadabra, but badcar is not. A longest common
subsequence (LCS for short) of two sequences x and y is the longest sequence that
is a subsequence of both: two longest common subsequences of abracadabra
and badcar are badar and bacar.
As with longest increasing subsequence, one can find the LCS of two sequence by
brute force, but it will take even longer. Not only are there are 2n subsequences
of a sequence of length n, but checking each subsequence of the first to see
if it is also a subsequence of the second may take some time. It is better to
solve the problem using dynamic programming. Having sequences gives an
obvious linear structure to exploit: the basic strategy will be to compute LCSs
for increasingly long prefixes of the inputs. But with two sequences we will have
to consider prefixes of both, which will give us a two-dimensional table where
rows correspond to prefixes of sequence x and columns correspond to prefixes of
sequence y.
The recursive decomposition that makes this technique work looks like this. Let
L(x, y) be the length of the longest common subsequence of x and y, where

342
x and y are strings. Let a and b be single characters. Then L(xa, yb) is the
maximum of:
• L(x, y) + 1, if a = b,
• L(xa, y), or
• L(x, yb).
The idea is that we either have a new matching character we couldn’t use before
(the first case), or we have an LCS that doesn’t use one of a or b (the remaining
cases). In each case the recursive call to LCS involves a shorter prefix of xa or
yb, with an ultimate base case L(x, y) = 0 if at least one of x or y is the empty
string. So we can fill in these values in a table, as long as we are careful to
make sure that the shorter prefixes are always filled first. If we are smart about
remembering which case applies at each step, we can even go back and extract
an actual LCS, by stitching together to places where a = b. Here’s a short C
program that does this:
#include <stdio.h>
#include <stdlib.h>
#include <assert.h>
#include <string.h>
#include <limits.h>

/* compute longest common subsequence of argv[1] and argv[2] */

/* computes longest common subsequence of x and y, writes result to lcs */

/* lcs should be pre-allocated by caller to 1 + minimum length of x or y */
void
longestCommonSubsequence(const char *x, const char *y, char *lcs)
{
int xLen;
int yLen;
int i; /* position in x */
int j; /* position in y */

xLen = strlen(x);
yLen = strlen(y);

/* best choice at each position */

/* length gives length of LCS for these prefixes */
/* prev points to previous substring */
/* newChar if non-null is new character */
struct choice {
int length;
struct choice *prev;
char newChar;
} best[xLen][yLen];

343
for(i = 0; i < xLen; i++) {
for(j = 0; j < yLen; j++) {
/* we can always do no common substring */
best[i][j].length = 0;
best[i][j].prev = 0;
best[i][j].newChar = 0;

/* if we have a match, try adding new character */

/* this is always better than the nothing we started with */
if(x[i] == y[j]) {
best[i][j].newChar = x[i];
if(i > 0 && j > 0) {
best[i][j].length = best[i-1][j-1].length + 1;
best[i][j].prev = &best[i-1][j-1];
} else {
best[i][j].length = 1;
}
}

/* maybe we can do even better by ignoring a new character */

if(i > 0 && best[i-1][j].length > best[i][j].length) {
/* throw away a character from x */
best[i][j].length = best[i-1][j].length;
best[i][j].prev = &best[i-1][j];
best[i][j].newChar = 0;
}

if(j > 0 && best[i][j-1].length > best[i][j].length) {

/* throw away a character from x */
best[i][j].length = best[i][j-1].length;
best[i][j].prev = &best[i][j-1];
best[i][j].newChar = 0;
}

}
}

/* reconstruct string working backwards from best[xLen-1][yLen-1] */

int outPos; /* position in output string */
struct choice *p; /* for chasing linked list */

outPos = best[xLen-1][yLen-1].length;
lcs[outPos--] = '\0';

for(p = &best[xLen-1][yLen-1]; p; p = p->prev) {

344
 if(p->newChar) {
assert(outPos >= 0);
lcs[outPos--] = p->newChar;
}
}
}

int
main(int argc, char **argv)
{
if(argc != 3) {
fprintf(stderr, "Usage: %s string1 string2\n", argv[0]);
return 1;
}

char output[strlen(argv[1]) + 1];

longestCommonSubsequence(argv[1], argv[2], output);

puts(output);

return 0;
}
examples/dynamicProgramming/lcs/lcs.c
The whole thing takes O(nm) time where n and m are the lengths of A and B.

4.14 Randomization

Randomization is a fundamental technique in algorithm design, that allows

programs to run quickly when the average-case behavior of an algorithm is
better than the worst-case behavior. It is also heavily used in games, both in
entertainment and gambling. The latter application gives the only example I
know of a programmer killed for writing bad code, which shows how serious
good random-number generation is.

4.14.1 Generating random values in C

If you want random values in a C program, there are three typical ways of getting
them, depending on how good (i.e. uniform, uncorrelated, and unpredictable)
you want them to be.

4.14.1.1 The rand function from the standard library

345
E.g.
#include <stdio.h>
#include <stdlib.h>

int
main(int argc, char **argv)
{
printf("%d\n", rand());
return 0;
}
examples/randomization/randOnce.c
The rand function, declared in stdlib.h, returns a random-looking integer in
the range 0 to RAND_MAX (inclusive) every time you call it. On machines using
the GNU C library RAND_MAX is equal to INT_MAX which is typically 23 1 − 1, but
RAND_MAX may be as small as 32767. There are no particularly strong guarantees
about the quality of random numbers that rand returns, but it should be good
enough for casual use, and it has the advantage that as part of the C standard
you can assume it is present almost everywhere.
Note that rand is a pseudorandom number generator: the sequence of values
it returns is predictable if you know its starting state (and is still predictable
from past values in the sequence even if you don’t know the starting state, if
you are clever enough). It is also the case that the initial seed is fixed, so that
the program above will print the same value every time you run it.
This is a feature: it permits debugging randomized programs. As John von
Neumann, who proposed pseudorandom number generators in his 1946 talk
“Various Techniques Used in Connection With Random Digits,” explained:
We see then that we could build a physical instrument to feed random
digits directly into a high-speed computing machine and could have
the control call for these numbers as needed. The real objection
to this procedure is the practical need for checking computations.
If we suspect that a calculation is wrong, almost any reasonable
check involves repeating something done before. At that point the
introduction of new random numbers would be intolerable.

4.14.1.1.1 Supplying a seed with srand

If you want to get different sequences, you need to seed the random number
generator using srand. A typical use might be:
#include <stdio.h>
#include <stdlib.h>
#include <time.h>

346
int
main(int argc, char **argv)
{
srand(time(0));
printf("%d\n", rand());
return 0;
}
examples/randomization/srandFromTime.c
Here time(0) returns the number of seconds since the epoch (00:00:00 UTC,
January 1, 1970, for POSIX systems, not counting leap seconds). Note that this
still might give repeated values if you run it twice in the same second, and it’s
extremely dangerous if you expect to distribute your code to a lot of people who
want different results, since two of your users are likely to run it twice in the
same second. See the discussion of /dev/urandom below for a better method.

4.14.1.2 Better pseudorandom number generators

There has been quite a bit of research on pseudorandom number generators over
the years, and much better pseudorandom number generators than rand are
available. The current champion for simulation work is Mersenne Twister,
which runs about 4 times faster than rand in its standard C implementation
and passes a much wider battery of statistical tests. Its English-language home
page is at http://www.math.sci.hiroshima-u.ac.jp/~m-mat/MT/emt.html. As
with rand, you still need to provide an initial seed value.
There are also cryptographically secure pseudorandom number genera-
tors, of which the most famous is Blum Blum Shub. These cannot be predicted
based on their output if seeded with a true random value (under certain crypto-
graphic assumptions: hardness of factoring for Blum Blum Shub). Unfortunately,
cryptographic PRNGs are usually too slow for day-to-day use.

4.14.1.3 Random numbers without the pseudo

If you really need actual random numbers and are on a Linux or BSD-like operat-
ing system, you can use the special device files /dev/random and /dev/urandom.
These can be opened for reading like ordinary files, but the values read from
them are a random sequence of bytes (including null characters). A typical use
might be:
#include <stdio.h>

int
main(int argc, char **argv)
{
unsigned int randval;

347
 FILE *f;

f = fopen("/dev/random", "r");
fread(&randval, sizeof(randval), 1, f);
fclose(f);

printf("%u\n", randval);

return 0;
}
examples/randomization/devRandom.c
(A similar construction can also be used to obtain a better initial seed for srand
than time(0).)
Both /dev/random and /dev/urandom derive their random bits from physically
random properties of the computer, like time between keystrokes or small
variations in hard disk rotation speeds. The difference between the two is that
/dev/urandom will always give you some random-looking bits, even if it has to
generate extra ones using a cryptographic pseudo-random number generator,
while /dev/random will only give you bits that it is confident are in fact random.
Since your computer only generates a small number of genuinely random bits per
second, this may mean that /dev/random will exhaust its pool if read too often.
In this case, a read on /dev/random will block (just like reading a terminal with
no input on it) until the pool has filled up again.
Neither /dev/random nor /dev/urandom is known to be secure against a deter-
mined attacker, but they are about the best you can do without resorting to
specialized hardware.

4.14.1.4 Range issues

The problem with rand is that getting a uniform value between 0 and 231 -1 may
not be what you want. It could be that RAND_MAX is be too small; in this case,
you may have to call rand more than once and paste together the results. But
there can be problems with RAND_MAX even if it is bigger than the values you
want.
For example, suppose you want to simulate a die roll for your video craps machine,
but you don’t want to get whacked by Johnny “The Debugger” when the Nevada
State Gaming Commission notices that 6-6 is coming up slightly less often than
it’s supposed to. A natural thing to try would be to take the output of rand
mod 6:
int d6(void) {
return rand() % 6 + 1;
}

348
The problem here is that there are 23 1 outputs from rand, and 6 doesn’t divide
23 1. So 1 and 2 are slightly more likely to come up than 3, 4, 5, or 6. This can
be particularly noticeable if we want a uniform variable from a larger range, e.g.
[0 . . . b(2/3) · 23 1c].
We can avoid this with a technique called rejection sampling, where we reject
excess parts of the output range of rand. For rolling a die, the trick is to reject
anything in the last extra bit of the range that is left over after the largest
multiple of the die size. Here’s a routine that does this, returning a uniform
value in the range 0 to n-1 for any positive n, together with a program that
demonstrates its use for rolling dice:
#include <stdio.h>
#include <stdlib.h>
#include <assert.h>
#include <time.h>

/* return a uniform random value in the range 0..n-1 inclusive */

int
randRange(int n)
{
int limit;
int r;

limit = RAND_MAX - (RAND_MAX % n);

while((r = rand()) >= limit);

return r % n;
}

int
main(int argc, char **argv)
{
int i;

srand(time(0));

for(i = 0; i < 40; i++) {

printf("%d ", randRange(6)+1);
}

putchar('\n');

return 0;
}

349
examples/randomization/randRange.c
More generally, rejection sampling can be used to get random values with
particular properties, where it’s hard to generate a value with that property
directly. Here’s a program that generates random primes:
#include <stdio.h>
#include <stdlib.h>
#include <assert.h>
#include <time.h>

/* return 1 if n is prime */
int
isprime(int n)
{
int i;

if(n % 2 == 0 || n == 1) { return 0; }

for(i = 3; i*i <= n; i += 2) {

if(n % i == 0) { return 0; }
}

return 1;
}

/* return a uniform random value in the range 0..n-1 inclusive */

int
randPrime(void)
{
int r;

/* extra parens avoid warnings */

while(!isprime((r = rand())));

return r;
}

int
main(int argc, char **argv)
{
int i;

srand(time(0));

for(i = 0; i < 10; i++) {

printf("%d\n", randPrime());

350
 }

return 0;
}
examples/randomization/randPrime.c
One temptation to avoid is to re-use your random values. If, for example, you
try to find a random prime by picking a random x and trying x, x+1, x+2, etc.,
until you hit a prime, some primes are more likely to come up than others.

4.14.2 Randomized algorithms

Randomized algorithms typically make random choices to get good average

worst-case performance in situations where a similar deterministic algorithm
would fail badly for some inputs but perform well on most inputs. The idea is
that the randomization scrambles the input space so that the adversary can’t
predict which possible input values will be bad for us. This still allows him to
make trouble if he gets lucky, but most of the time our algorithm should run
quickly.

4.14.2.1 Randomized search

This is essentially rejection sampling in disguise. Suppose that you want to find
one of many needles in a large haystack. One approach is to methodically go
through the straws/needles one at a time until you find a needle. But you may
find that your good friend the adversary has put all the needles at the end of
your list. Picking candidate at random is likely to hit a needle faster if there are
many of them.
Here is a (silly) routine that quickly finds a number whose high-order bits match
a particular pattern:
int
matchBits(int pattern)
{
int r;

while(((r = rand()) & 0x70000000) != (pattern & 0x70000000));

return r;
}
This will find a winning value in 8 tries on average. In contrast, this deterministic
version will take a lot longer for nonzero patterns:
int
matchBitsDeterministic(int pattern)

351
{
int i;

for(i = 0; (i & 0x70000000) != (pattern & 0x70000000); i++);

return i;
}
The downside of the randomized approach is that it’s hard to tell when to quit
if there are no matches; if we stop after some fixed number of trials, we get a
Monte Carlo algorithm that may give the wrong answer with small probability.
The usual solution is to either accept a small probability of failure, or interleave
a deterministic backup algorithm that always works. The latter approach gives
a Las Vegas algorithm whose running time is variable but whose correctness is
not.

4.14.2.2 Quickselect and quicksort

Quickselect, or Hoare’s FIND (Hoare, C. A. R. Algorithm 65: FIND, CACM
4(7):321–322, July 1961), is an algorithm for quickly finding the k-th largest
element in an unsorted array of n elements. It runs in O(n) time on average,
which is the best one can hope for (we have to look at every element of the array
to be sure we didn’t miss a small one that changes our answer) and better than
the O(n log n) time we get if we sort the array first using a comparison-based
sorting algorithm.
The idea is to pick a random pivot and divide the input into two piles, each
of which is likely to be roughly a constant fraction of the size of the original
input.24 It takes O(n) time to split the input up (we have to compare each
element to the pivot once), and in the recursive calls this gives a geometric series.
We can even do the splitting up in place if we are willing to reorder the elements
of our original array.
If we recurse into both piles instead of just one, we get quicksort (Hoare, C.
A. R. Algorithm 64: Quicksort. CACM 4(7):321, July 1961), a very fast and
simple comparison-based sorting algorithm. Here is an implementation of both
algorithms:
#include <stdio.h>
#include <stdlib.h>
#include <assert.h>

/* reorder an array to put elements <= pivot

* before elements > pivot.
24 The actual analysis is pretty complicated, since we are more likely to land in a bigger pile,

but it’s not hard to show that on average even the bigger pile has no more than 3/4 of the
elements.

352
 * Returns number of elements <= pivot */
static int
splitByPivot(int n, int *a, int pivot)
{
int lo;
int hi;
int temp; /* for swapping */

assert(n >= 0);

/* Dutch Flag algorithm */

/* swap everything <= pivot to bottom of array */
/* invariant is i < lo implies a[i] <= pivot */
/* and i > hi implies a[i] > pivot */
lo = 0;
hi = n-1;

while(lo <= hi) {

if(a[lo] <= pivot) {
lo++;
} else {
temp = a[hi];
a[hi--] = a[lo];
a[lo] = temp;
}
}

return lo;
}

/* find the k-th smallest element of an n-element array */

/* may reorder elements of the original array */
int
quickselectDestructive(int k, int n, int *a)
{
int pivot;
int lo;

assert(0 <= k);

assert(k < n);

if(n == 1) {
return a[0];
}

/* else */

353
 pivot = a[rand() % n]; /* we will tolerate non-uniformity */

lo = splitByPivot(n, a, pivot);

/* lo is now number of values <= pivot */

if(k < lo) {
return quickselectDestructive(k, lo, a);
} else {
return quickselectDestructive(k - lo, n - lo, a + lo);
}
}

/* sort an array in place */

void
quickSort(int n, int *a)
{
int pivot;
int lo;

if(n <= 1) {
return;
}

/* else */
pivot = a[rand() % n]; /* we will tolerate non-uniformity */

lo = splitByPivot(n, a, pivot);

quickSort(lo, a);
quickSort(n - lo, a + lo);
}

/* shuffle an array */
void
shuffle(int n, int *a)
{
int i;
int r;
int temp;

for(i = n - 1; i > 0; i--) {

r = rand() % i;
temp = a[r];
a[r] = a[i];
a[i] = temp;

354
 }
}

#define N (1024)

int
main(int argc, char **argv)
{
int a[N];
int i;

srand(0); /* use fixed value for debugging */

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

a[i] = i;
}

shuffle(N, a);

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

assert(quickselectDestructive(i, N, a) == i);
}

shuffle(N, a);

quickSort(N, a);

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

assert(a[i] == i);
}

return 0;
}
examples/randomization/quick.c

4.14.3 Randomized data structures

Suppose we insert n elements into an initially-empty binary search tree in random

order with no rebalancing. Then each element is equally likely to be the root,
and all the elements less than the root end up in the left subtree, while all
the elements greater than the root end up in the right subtree, where they are
further partitioned recursively. This is exactly what happens in quicksort, so
the structure of the tree will exactly mirror the structure of an execution of
quicksort. In particular, the average depth of a node will be O(log n), giving us

355
the same expected search cost as in a balanced binary tree.
The problem with this approach is that we don’t have any guarantees that the
input will be supplied in random order, and in the worst case we end up with a
linked list. The solution is to put the randomization into the algorithm itself,
making the structure of the tree depend on random choices made by the program
itself.

4.14.3.1 Skip lists

A skip list (Pugh, 1990) is a randomized tree-like data structure based on linked
lists. It consists of a level 0 list that is an ordinary sorted linked list, together
with higher-level lists that contain a random sampling of the elements at lower
levels. When inserted into the level i list, an element flips a coin that tells it
with probability p to insert itself in the level i+1 list as well.
Searches in a skip list are done by starting in the highest-level list and searching
forward for the last element whose key is smaller than the target; the search then
continues in the same way on the next level down. The idea is that the higher-
level lists act as express lanes to get us to our target value faster. To bound the
expected running time of a search, it helps to look at this process backwards;
the reversed search path starts at level 0 and continues going backwards until it
reaches the first element that is also in a higher level; it then jumps to the next
level up and repeats the process. On average, we hit 1 + 1/p nodes at each level
before jumping back up; for constant p (e.g. 1/2), this gives us O(log n) steps
for the search.
The space per element of a skip list also depends on p. Every element has at
least one outgoing pointer (on level 0), and on average has exactly 1/(1 − p)
expected pointers. So the space cost can also be adjusted by adjusting p. For
example, if space is at a premium, setting p = 1/10 produces 10/9 pointers per
node on average—not much more than in a linked list—but still gives O(log n)
search time.
Below is an implementation of a skip list. To avoid having to allocate a separate
array of pointers for each element, we put a length-1 array at the end of
struct skiplist and rely on C’s lack of bounds checking to make the array
longer if necessary. A dummy head element stores pointers to all the initial
elements in each level of the skip list; it is given the dummy key INT_MIN so that
searches for values less than any in the list will report this value. Aside from
these nasty tricks, the code for search and insertion is pretty straightforward.
Code for deletion is a little more involved, because we have to make sure that
we delete the leftmost copy of a key if there are duplicates (an alternative would
be to modify skiplistInsert to ignore duplicates).
#include <stdlib.h>
#include <assert.h>
#include <limits.h>

356
#include "skiplist.h"

#define MAX_HEIGHT (32)

struct skiplist {
int key;
int height; /* number of next pointers */
struct skiplist *next[1]; /* first of many */
};

/* choose a height according to a geometric distribution */

static int
chooseHeight(void)
{
int i;

for(i = 1; i < MAX_HEIGHT && rand() % 2 == 0; i++);

return i;
}

/* create a skiplist node with the given key and height */

/* does not fill in next pointers */
static Skiplist
skiplistCreateNode(int key, int height)
{
Skiplist s;

assert(height > 0);

assert(height <= MAX_HEIGHT);

s = malloc(sizeof(struct skiplist) + sizeof(struct skiplist *) * (height - 1));

assert(s);

s->key = key;
s->height = height;

return s;
}

/* create an empty skiplist */

Skiplist
skiplistCreate(void)
{

357
 Skiplist s;
int i;

/* s is a dummy head element */

s = skiplistCreateNode(INT_MIN, MAX_HEIGHT);

/* this tracks the maximum height of any node */

s->height = 1;

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

s->next[i] = 0;
}

return s;
}

/* free a skiplist */
void
skiplistDestroy(Skiplist s)
{
Skiplist next;

while(s) {
next = s->next[0];
free(s);
s = next;
}
}

/* return maximum key less than or equal to key */

/* or INT_MIN if there is none */
int
skiplistSearch(Skiplist s, int key)
{
int level;

for(level = s->height - 1; level >= 0; level--) {

while(s->next[level] && s->next[level]->key <= key) {
s = s->next[level];
}
}

return s->key;
}

/* insert a new key into s */

358
void
skiplistInsert(Skiplist s, int key)
{
int level;
Skiplist elt;

elt = skiplistCreateNode(key, chooseHeight());

assert(elt);

if(elt->height > s->height) {

s->height = elt->height;
}

/* search through levels taller than elt */

for(level = s->height - 1; level >= elt->height; level--) {
while(s->next[level] && s->next[level]->key < key) {
s = s->next[level];
}
}

/* now level is elt->height - 1, we can start inserting */

for(; level >= 0; level--) {
while(s->next[level] && s->next[level]->key < key) {
s = s->next[level];
}

/* s is last entry on this level < new element */

/* do list insert */
elt->next[level] = s->next[level];
s->next[level] = elt;
}
}

/* delete a key from s */

void
skiplistDelete(Skiplist s, int key)
{
int level;
Skiplist target;

/* first we have to find leftmost instance of key */

target = s;

for(level = s->height - 1; level >= 0; level--) {

while(target->next[level] && target->next[level]->key < key) {

359
 target = target->next[level];
}
}

/* take one extra step at bottom */

target = target->next[0];

if(target == 0 || target->key != key) {

return;
}

/* now we found target, splice it out */

for(level = s->height - 1; level >= 0; level--) {
while(s->next[level] && s->next[level]->key < key) {
s = s->next[level];
}

if(s->next[level] == target) {
s->next[level] = target->next[level];
}
}

free(target);
}
examples/trees/skiplist/skiplist.c
Here is the header file, Makefile, and test code: skiplist.h, Makefile,
test_skiplist.c.

4.14.3.2 Universal hash families

Randomization can also be useful in hash tables. Recall that in building a hash
table, we are relying on the hash function to spread out bad input distributions
over the indices of our array. But for any fixed hash function, in the worst case
we may get inputs where every key hashes to the same location. Universal
hashing (Carter and Wegman, 1979) solves this problem by choosing a hash
function at random. We may still get unlucky and have the hash function hash
all our values to the same location, but now we are relying on the random number
generator to be nice to us instead of the adversary. We can also rehash with a
new random hash function if we find out that the one we are using is bad.
The problem here is that we can’t just choose a function uniformly at random
out of the set of all possible hash functions, because there are too many of them,
meaning that we would spend more space representing our hash function than
we would on the table. The solution is to observe that we don’t need our hash
function h to be truly random; it’s enough if the probability of collision Pr[h(x)

360
= h(y)] for any fixed keys x =6 y is 1/m, where m is the size of the hash table.
The reason is that the cost of searching for x (with chaining) is linear in the
number of keys already in the table that collide with it. The expected number
of such collisions is the sum of Pr[h(x) = h(y)] over all keys y in the table, or
n/m if we have n keys. So this pairwise collision probability bound is enough to
get the desired n/m behavior out of our table. A family of hash function with
this property is called universal.
How do we get a universal hash family? For strings, we can use a table of random
values, one for each position and possible character in the string. The hash
of a string is then the exclusive or of the random values hashArray[i][s[i]]
corresponding to the actual characters in the string. If our table has size a power
of two, this has the universal property, because if two strings x and y differ in
some position i, then there is only one possible value of hashArray[i][y[i]]
(mod m) that will make the hash functions equal.
Typically to avoid having to build an arbitrarily huge table of random values,
we only has an initial prefix of the string. Here is a hash function based on this
idea, which assumes that the d data structure includes a hashArray field that
contains the random values for this particular hash table:
static unsigned long
hash_function(Dict d, const char *s)
{
unsigned const char *us;
unsigned long h;
int i;

h = 0;

us = (unsigned const char *) s;

for(i = 0; i < HASH_PREFIX_LENGTH && us[i] != '\0'; i++) {

h ^= d->hashArray[i][us[i]];
}

return h;
}
A modified version of the Dict hash table from the chapter on hash tables that
uses this hash function is given here: dict.c, dict.h, test_dict.c, Makefile.

4.15 String processing

Most of the time, when we’ve talked about the asymptotic performance of data
structures, we have implicitly assumed that the keys we are looking up are of

361
constant size. This means that computing a hash function or comparing two
keys (as in a binary search tree) takes O(1) time. What if this is not the case?
If we consider an m-character string, any reasonable hash function is going to
take O(m) time since it has to look at all of the characters. Similarly, comparing
two m-character strings may also take O(m) time. If we charge for this (as we
should!) then the cost of hash table operations goes from O(1) to O(m) and the
cost of binary search tree operations, even in a balanced tree, goes from O(log n)
to O(m log n). Even sorting becomes more expensive: a sorting algorithm that
does O(n log n) comparisons may now take O(mn log n) time. But maybe we
can exploit the structure of strings to get better performance.

4.15.1 Radix search

Radix search refers to a variety of data structures that support searching for
strings considered as sequences of digits in some large base (or radix). These
are generally faster than simple binary search trees because they usually only
require examining one byte or less of the target string at each level of the tree,
as compared to every byte in the target in a full string comparison. In many
cases, the best radix search trees are even faster than hash tables, because they
only need to look at a small part of the target string to identify it.
We’ll describe several radix search trees, starting with the simplest and working
up.

4.15.1.1 Tries
A trie is a binary tree (or more generally, a k-ary tree where k is the radix)
where the root represents the empty bit sequence and the two children of a
node representing sequence x represent the extended sequences x0 and x1 (or
generally x0, x1, . . . , x(k − 1)). So a key is not stored at a particular node but
is instead represented bit-by-bit (or digit-by-digit) along some path. Typically
a trie assumes that the set of keys is prefix-free, i.e. that no key is a prefix of
another; in this case there is a one-to-one correspondence between keys and
leaves of the trie. If this is not the case, we can mark internal nodes that
also correspond to the ends of keys, getting a slightly different data structure
known as a digital search tree. For null-terminated strings as in C, the null
terminator ensures that any set of strings is prefix-free.
Given this simple description, a trie storing a single long key would have a very
large number of nodes. A standard optimization is to chop off any path with no
branches in it, so that each leaf corresponds to the shortest unique prefix of a
key. This requires storing the key in the leaf so that we can distinguish different
keys with the same prefix.
The name trie comes from the phrase “information retrieval.” Despite the
etymology, trie is now almost always pronounced like try instead of tree to avoid

362
confusion with other tree data structures.

4.15.1.1.1 Searching a trie

Searching a trie is similar to searching a binary search tree, except that instead
of doing a comparison at each step we just look at the next bit in the target.
The time to perform a search is proportional to the number of bits in the longest
path in the tree matching a prefix of the target. This can be very fast for search
misses if the target is wildly different from all the keys in the tree.

4.15.1.1.2 Inserting a new element into a trie

Insertion is more complicated for tries than for binary search trees. The reason
is that a new element may add more than one new node. There are essentially
two cases:
1. (The simple case.) In searching for the new key, we reach a null pointer
leaving a non-leaf node. In this case we can simply add a new leaf. The
cost of this case is essentially the same as for search plus O(1) for building
the new leaf.
2. (The other case.) In searching for the new key, we reach a leaf, but the
key stored there isn’t the same as the new key. Now we have to generate a
new path for as long as the old key and the new key have the same bits,
branching out to two different leaves at the end. The cost of this operation
is within a constant factor of the cost for searching for the new leaf after
it is inserted, since that’s how long the newly-built search path will be.
In either case, the cost is bounded by the length of the new key, which is about
the best we can hope for in the worst case for any data structure.

4.15.1.1.3 Implementation
A typical trie implementation in C might look like this. It uses a GET_BIT macro
similar to the one from the chapter on bit manipulation, except that we reverse
the bits within each byte to get the right sorting order for keys.
typedef struct trie_node *Trie;

#define EMPTY_TRIE (0)

/* returns 1 if trie contains target */

int trie_contains(Trie trie, const char *target);

/* add a new key to a trie */

/* and return the new trie */
Trie trie_insert(Trie trie, const char *key);

363
/* free a trie */
void trie_destroy(Trie);

/* debugging utility: print all keys in trie */

void trie_print(Trie);
examples/trees/trie/trie.h
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <assert.h>

#include "trie.h"

#define BITS_PER_BYTE (8)

/* extract the n-th bit of x */

/* here we process bits within bytes in MSB-first order */
/* this sorts like strcmp */
#define GET_BIT(x, n) ((((x)[(n) / BITS_PER_BYTE]) & (0x1 << (BITS_PER_BYTE - 1 - (n) % BITS

#define TRIE_BASE (2)

struct trie_node {
char *key;
struct trie_node *kids[TRIE_BASE];
};

#define IsLeaf(t) ((t)->kids[0] == 0 && (t)->kids[1] == 0)

/* returns 1 if trie contains target */

int
trie_contains(Trie trie, const char *target)
{
int bit;

for(bit = 0; trie && !IsLeaf(trie); bit++) {

/* keep going */
trie = trie->kids[GET_BIT(target, bit)];
}

if(trie == 0) {
/* we lost */
return 0;
} else {
/* check that leaf really contains the target */

364
 return !strcmp(trie->key, target);
}
}

/* gcc -pedantic kills strdup! */

static char *
my_strdup(const char *s)
{
char *s2;

s2 = malloc(strlen(s) + 1);
assert(s2);

strcpy(s2, s);
return s2;
}

/* helper functions for insert */

static Trie
make_trie_node(const char *key)
{
Trie t;
int i;

t = malloc(sizeof(*t));
assert(t);

if(key) {
t->key = my_strdup(key);
assert(t->key);
} else {
t->key = 0;
}

for(i = 0; i < TRIE_BASE; i++) t->kids[i] = 0;

return t;
}

/* add a new key to a trie */

/* and return the new trie */
Trie
trie_insert(Trie trie, const char *key)
{
int bit;

365
 int bitvalue;
Trie t;
Trie kid;
const char *oldkey;

if(trie == 0) {
return make_trie_node(key);
}
/* else */
/* first we'll search for key */
for(t = trie, bit = 0; !IsLeaf(t); bit++, t = kid) {
kid = t->kids[bitvalue = GET_BIT(key, bit)];
if(kid == 0) {
/* woohoo! easy case */
t->kids[bitvalue] = make_trie_node(key);
return trie;
}
}

/* did we get lucky? */

if(!strcmp(t->key, key)) {
/* nothing to do */
return trie;
}

/* else */
/* hard case---have to extend the trie */
oldkey = t->key;
#ifdef EXCESSIVE_TIDINESS
t->key = 0; /* not required but makes data structure look tidier */
#endif

/* walk the common prefix */

while(GET_BIT(oldkey, bit) == (bitvalue = GET_BIT(key, bit))) {
kid = make_trie_node(0);
t->kids[bitvalue] = kid;
bit++;
t = kid;
}

/* then split */
t->kids[bitvalue] = make_trie_node(key);
t->kids[!bitvalue] = make_trie_node(oldkey);

return trie;
}

366
/* kill it */
void
trie_destroy(Trie trie)
{
int i;

if(trie) {
for(i = 0; i < TRIE_BASE; i++) {
trie_destroy(trie->kids[i]);
}

if(IsLeaf(trie)) {
free(trie->key);
}

free(trie);
}
}

static void
trie_print_internal(Trie t, int bit)
{
int i;
int kid;

if(t != 0) {
if(IsLeaf(t)) {
for(i = 0; i < bit; i++) putchar(' ');
puts(t->key);
} else {
for(kid = 0; kid < TRIE_BASE; kid++) {
trie_print_internal(t->kids[kid], bit+1);
}
}
}
}

void
trie_print(Trie t)
{
trie_print_internal(t, 0);
}
examples/trees/trie/trie.c
Here is a short test program that demonstrates how to use it:

367
#include <stdio.h>
#include <stdlib.h>

#include "trie.h"

/* test for trie.c */

/* reads lines from stdin and echoes lines that haven't appeared before */

/* read a line of text from stdin

* and return it (without terminating newline) as a freshly-malloc'd block.
* Caller is responsible for freeing this block.
* Returns 0 on error or EOF.
*/
char *
getline(void)
{
char *line; /* line buffer */
int n; /* characters read */
int size; /* size of line buffer */
int c;

size = 1;
line = malloc(size);
if(line == 0) return 0;

n = 0;

while((c = getchar()) != '\n' && c != EOF) {

while(n >= size - 1) {
size *= 2;
line = realloc(line, size);
if(line == 0) return 0;
}
line[n++] = c;
}

if(c == EOF && n == 0) {

/* got nothing */
free(line);
return 0;
} else {
line[n++] = '\0';
return line;
}
}

368
int
main(int argc, char **argv)
{
Trie t;
char *line;

t = EMPTY_TRIE;

while((line = getline()) != 0) {
if(!trie_contains(t, line)) {
puts(line);
}

/* try to insert it either way */

/* this tests that insert doesn't blow up on duplicates */
t = trie_insert(t, line);

free(line);
}

puts("===");

trie_print(t);

trie_destroy(t);

return 0;
}
examples/trees/trie/test_trie.c

4.15.1.2 Patricia trees

Tries perform well when all keys are short (or are distinguished by short prefixes),
but can grow very large if one inserts two keys that have a long common prefix.
The reason is that a trie has to store an internal node for every bit of the
common prefix until the two keys become distinguishable, leading to long chains
of internal nodes each of which has only one child. An optimization (described
in this paper) known as a Patricia tree eliminates these long chains by having
each node store the number of the bit to branch on, like this:
struct patricia_node {
char *key;
int bit;
struct patricia_node *kids[2];
};

369
typedef struct patricia_node *Patricia;
Now when searching for a key, instead of using the number of nodes visited so
far to figure out which bit to look at, we just read the bit out of the table. This
means in particular that we can skip over any bits that we don’t actually branch
on. We do however have to be more careful to make sure we don’t run off the
end of our target key, since it is possible that when skipping over intermediate
bits we might skip over some that distinguish our target from all keys in the
table, including longer keys. For example, a Patricia tree storing the strings
abc and abd will first test bit position 22, since that’s where abc and abd differ.
This can be trouble if we are looking for x.
Here’s the search code:
int
patricia_contains(Patricia t, const char *key)
{
int key_bits;

key_bits = BITS_PER_BYTE * (strlen(key)+1); /* +1 for the NUL */

while(t && !IsLeaf(t)) {

if(t->bit >= key_bits) {
/* can't be there */
return 0;
} else {
t = t->kids[GET_BIT(key, t->bit)];
}
}

return t && !strcmp(t->key, key);

}
The insertion code is similar in many respects to the insertion code for a trie.
The differences are that we never construct a long chain of internal nodes when
splitting a leaf (although we do have to scan through both the old and new keys
to find the first bit position where they differ), but we may sometimes have to
add a new internal node between two previously existing nodes if a new key
branches off at a bit position that was previously skipped over.
In the worst case Patricia trees are much more efficient than tries, in both space
(linear in the number of keys instead of linear in the total size of the keys) and
time complexity, often needing to examine only a very small number of bits for
misses (hits still require a full scan in strcmp to verify the correct key). The
only downside of Patricia trees is that since they work on bits, they are not quite
perfectly tuned to the byte or word-oriented structure of modern CPUs.

370
4.15.1.3 Ternary search trees
Ternary search trees were described by Jon Bentley and Bob Sedgewick in
an article in the April 1988 issue of Dr. Dobb’s Journal, available here.
The basic idea is that each node in the tree stores one character from the key
and three child pointers lt, eq, and gt. If the corresponding character in the
target is equal to the character in the node, we move to the next character in
the target and follow the eq pointer out of the node. If the target is less, follow
the lt pointer but stay at the same character. If the target is greater, follow the
gt pointer and again stay at the same character. When searching for a string,
we walk down the tree until we either reach a node that matches the terminating
NUL (a hit), or follow a null pointer (a miss).
A TST acts a bit like a 256-way trie, except that instead of storing an array
of 256 outgoing pointers, we build something similar to a small binary search
tree for the next character. Note that no explicit balancing is done within these
binary search trees. From a theoretical perspective, the worst case is that we
get a 256-node deep linked-list equivalent at each step, multiplying our search
time by 256 = O(1). In practice, only those characters that actual appear in
some key at this stage will show up, so the O(1) is likely to be a small O(1),
especially if keys are presented in random order.
TSTs are one of the fastest known data structures for implementing dictionaries
using strings as keys, beating both hash tables and tries in most cases. They
can be slower than Patricia trees if the keys have many keys with long matching
prefixes; however, a Patricia-like optimization can be applied to give a com-
pressed ternary search tree that works well even in this case. In practice,
regular TSTs are usually good enough.
Here is a simple implementation of an insert-only TST. The C code includes two
versions of the insert helper routine; the first is the original recursive version
and the second is an iterative version generated by eliminating the tail recursion
from the first.
typedef struct tst_node *TST;

#define EMPTY_TST (0)

/* returns 1 if t contains target */

int tst_contains(TST t, const char *target);

/* add a new key to a TST */

/* and return the new TST */
TST tst_insert(TST t, const char *key);

/* free a TST */
void tst_destroy(TST);

371
examples/trees/tst/tst.h
#include <stdio.h>
#include <stdlib.h>
#include <assert.h>

#include "tst.h"

struct tst_node {
char key; /* value to split on */
struct tst_node *lt; /* go here if target[index] < value */
struct tst_node *eq; /* go here if target[index] == value */
struct tst_node *gt; /* go here if target[index] > value */
};

/* returns 1 if t contains key */

int
tst_contains(TST t, const char *key)
{
assert(key);

while(t) {
if(*key < t->key) {
t = t->lt;
} else if(*key > t->key) {
t = t->gt;
} else if(*key == '\0') {
return 1;
} else {
t = t->eq;
key++;
}
}

return 0;
}

/* original recursive insert */

static void
tst_insert_recursive(TST *t, const char *key)
{
if(*t == 0) {
*t = malloc(sizeof(**t));
assert(*t);
(*t)->key = *key;
(*t)->lt = (*t)->eq = (*t)->gt = 0;

372
 }

/* now follow search */

if(*key < (*t)->key) {
tst_insert_recursive(&(*t)->lt, key);
} else if(*key > (*t)->key) {
tst_insert_recursive(&(*t)->gt, key);
} else if(*key == '\0') {
/* do nothing, we are done */
;
} else {
tst_insert_recursive(&(*t)->eq, key+1);
}
}

/* iterative version of above, since somebody asked */

/* This is pretty much standard tail-recursion elimination: */
/* The whole function gets wrapped in a loop, and recursive
* calls get replaced by assignment */
static void
tst_insert_iterative(TST *t, const char *key)
{
for(;;) {
if(*t == 0) {
*t = malloc(sizeof(**t));
assert(*t);
(*t)->key = *key;
(*t)->lt = (*t)->eq = (*t)->gt = 0;
}

/* now follow search */

if(*key < (*t)->key) {
t = &(*t)->lt;
} else if(*key > (*t)->key) {
t = &(*t)->gt;
} else if(*key == '\0') {
/* do nothing, we are done */
return;
} else {
t = &(*t)->eq;
key++;
}
}
}

373
/* add a new key to a TST */
/* and return the new TST */
TST
tst_insert(TST t, const char *key)
{
assert(key);

#ifdef USE_RECURSIVE_INSERT
tst_insert_recursive(&t, key);
#else
tst_insert_iterative(&t, key);
#endif
return t;
}

/* free a TST */
void
tst_destroy(TST t)
{
if(t) {
tst_destroy(t->lt);
tst_destroy(t->eq);
tst_destroy(t->gt);
free(t);
}
}
examples/trees/tst/tst.c
And here is some test code, almost identical to the test code for tries: test_tst.c.
The Dr. Dobb’s article contains additional code for doing deletions and partial
matches, plus some optimizations for inserts.

4.15.1.4 More information

• http://imej.wfu.edu/articles/2002/2/02/index.asp has some good Java-
based animations of radix tries, Patricia tries, and other tree-like data
structures.

4.15.2 Radix sort

The standard quicksort routine is an example of a comparison-based sorting

algorithm. This means that the only information the algorithm uses about the
items it is sorting is the return value of the compare routine. This has a rather
nice advantage of making the algorithm very general, but has the disadvantage

374
that the algorithm can extract only one bit of information from every call to
compare. Since there are n! possible ways to reorder the input sequence, this
means we need at least log(n!) = O(n log n) calls to compare to finish the sort. If
we are sorting something like strings, this can get particularly expensive, because
calls to strcmp may take time linear in the length of the strings being compared.
In the worst case, sorting n strings of length m each could take O(nm log n)
time.

4.15.2.1 Bucket sort

The core idea of radix sort is that if we want to sort values from a small range,
we can do it by making one bucket for each possible value and throw any object
with that value into the corresponding bucket. In the old days, when Solitaire
was a game played with physical pieces of cardboard, a player who suspected
that that one of these “cards” had fallen under the couch might sort the deck by
dividing it up into Spades, Hearts, Diamonds, and Club piles and then sorting
each pile recursively. The same trick works in a computer, but there the buckets
are typically implemented as an array of some convenient data structure.
If the number of possible values is too big, we may still be able to use bucket
sort digit-by-digit. The resulting algorithms are known generally as radix sort.
These are a class of algorithms designed for sorting strings in lexicographic
order—the order used by dictionaries where one string is greater than another if
the first character on which they differ is greater. One particular variant, most-
significant-byte radix sort or MSB radix sort, has the beautiful property
that its running time is not only linear in the size of the input in bytes, but
is also linear in the smallest number of characters in the input that need to
be examined to determine the correct order. This algorithm is so fast that
it takes not much more time to sort data than it does to read the data from
memory and write it back. But it’s a little trickier to explain that the original
least-significant-byte radix sort or LSB radix sort.

4.15.2.2 Classic LSB radix sort

This is the variant used for punch cards, and works well for fixed-length strings.
The idea is to sort on the least significant position first, then work backwards to
the most significant position. This works as long as each sort is stable, meaning
that it doesn’t reorder values with equal keys. For example, suppose we are
sorting the strings:
sat
bat
bad
The first pass sorts on the third column, giving:
bad

375
sat
bat
The second pass sorts on the second column, producing no change in the order
(all the characters are the same). The last pass sorts on the first column. This
moves the s after the two bs, but preserves the order of the two words starting
with b. The result is:
bad
bat
sat
There are three downsides to LSB radix sort:
1. All the strings have to be the same length (this is not necessarily a problem
if they are really fixed-width data types like ints).
2. The algorithm used to sort each position must be stable, which may require
more effort than most programmers would like to take.
3. It may be that the late positions in the strings don’t affect the
order, but we have to sort on them anyway. If we are sorting
aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa and baaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa,
we spend a lot of time matching up as against each other.

4.15.2.3 MSB radix sort

For these reasons, MSB radix sort is used more often. This is basically the radix
sort version of quicksort, where instead of partitioning our input data into two
piles based on whether each element is less than or greater than a random pivot,
we partition the input into 256 piles, one for each initial byte. We can then
sort each pile recursively using the same algorithm, taking advantage of the fact
that we know that the first byte (or later, the first k bytes) are equal and so we
only need to look at the next one. The recursion stops when we get down to 1
value, or in practice where we get down to a small enough number of values that
the cost of doing a different sorting algorithm becomes lower than the cost of
creating and tearing down the data structures for managing the piles.

4.15.2.3.1 Issues with recursion depth

The depth of recursion for MSB radix sort is equal to the length of the second-
longest string in the worst case. Since strings can be pretty long, this creates
a danger of blowing out the stack. The solution (as in quicksort) is to use
tail recursion for the largest pile. Now any pile we recurse into with an actual
procedure call is at most half the size of the original pile, so we get stack depth
at most O(log n).

4.15.2.3.2 Implementing the buckets

376
There is a trick we can do analagous to the Dutch flag algorithm where we sort
the array in place. The idea is that we first count the number of elements that
land in each bucket and assign a block of the array for each bucket, keeping
track in each block of an initial prefix of values that belong in the bucket with
the rest not yet processed. We then walk through the buckets swapping out any
elements at the top of the good prefix to the bucket they are supposed to be in.
This procedure puts at least one element in the right bucket for each swap, so
we reorder everything correctly in at most n swaps and O(n) additional work.
To keep track of each bucket, we use two pointers bucket[i] for the first element
of the bucket and top[i] for the first unused element. We could make these be
integer array indices, but this slows the code down by about 10%. This seems to
be a situation where our use of pointers is complicated enough that the compiler
can’t optimize out the array lookups.

4.15.2.3.3 Further optimization

Since we are detecting the heaviest bucket anyway, there is a straightforward
optimization that speeds the sort up noticeably on inputs with a lot of duplicates:
if everything would land in the same bucket, we can skip the bucket-sort and
just go directly to the next character.

4.15.2.3.4 Sample implementation

Here is an implementation of MSB radix sort using the ideas above:
#include <assert.h>
#include <limits.h>
#include <string.h>

#include "radixSort.h"

/* in-place MSB radix sort for null-terminated strings */

/* helper routine for swapping */

static void
swapStrings(const char **a, const char **b)
{
const char *temp;

temp = *a;
*a = *b;
*b = temp;
}

/* this is the internal routine that assumes all strings are equal for the

377
 * first k characters */
static void
radixSortInternal(int n, const char **a, int k)
{
int i;
int count[UCHAR_MAX+1]; /* number of strings with given character in position k */
int mode; /* most common position-k character */
const char **bucket[UCHAR_MAX+1]; /* position of character block in output */
const char **top[UCHAR_MAX+1]; /* first unused index in this character block */

/* loop implements tail recursion on most common character */

while(n > 1) {

/* count occurrences of each character */

memset(count, 0, sizeof(int)*(UCHAR_MAX+1));

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

count[(unsigned char) a[i][k]]++;
}

/* find the most common nonzero character */

/* we will handle this specially */
mode = 1;
for(i = 2; i < UCHAR_MAX+1; i++) {
if(count[i] > count[mode]) {
mode = i;
}
}

if(count[mode] < n) {

/* generate bucket and top fields */

bucket[0] = top[0] = a;
for(i = 1; i < UCHAR_MAX+1; i++) {
top[i] = bucket[i] = bucket[i-1] + count[i-1];
}

/* reorder elements by k-th character */

/* this is similar to dutch flag algorithm */
/* we start at bottom character and swap values out until everything is in place
/* invariant is that for all i, bucket[i] <= j < top[i] implies a[j][k] == i */
for(i = 0; i < UCHAR_MAX+1; i++) {
while(top[i] < bucket[i] + count[i]) {
if((unsigned char) top[i][0][k] == i) {
/* leave it in place, advance bucket */
top[i]++;

378
 } else {
/* swap with top of appropriate block */
swapStrings(top[i], top[(unsigned char) top[i][0][k]]++);
}
}
}

/* we have now reordered everything */

/* recurse on all but 0 and mode */
for(i = 1; i < UCHAR_MAX+1; i++) {
if(i != mode) {
radixSortInternal(count[i], bucket[i], k+1);
}
}

/* tail recurse on mode */

n = count[mode];
a = bucket[mode];
k = k+1;

} else {

/* tail recurse on whole pile */

k = k+1;
}
}
}

void
radixSort(int n, const char **a)
{
radixSortInternal(n, a, 0);
}
examples/sorting/radixSort/radixSort.c
Some additional files: radixSort.h, test_radixSort.c, Makefile, sortInput.c.
The last is a program that sorts lines on stdin and writes the result to
stdout, similar to the GNU sort utility. When compiled with -O3 and
run on my machine, this runs in about the same time as the standard sort
program when run on a 4.7 million line input file consisting of a random
shuffle of 20 copies of /usr/share/dict/words, provided sort is run with
LANG=C sort < /usr/share/dict/words to keep it from having to deal with
locale-specific collating issues. On other inputs, sort is faster. This is not
bad given how thoroughly sort has been optimized, but is a sign that further
optimization is possible.

379
5 Other topics not covered in detail in 2015
These are mostly leftovers from previous versions of the class where different
topics were emphasized.

5.1 More applications of function pointers

Here we show how to implement various mechanisms often found in more

sophisticated programming languages in C using function pointers.

5.1.1 Iterators

Suppose we have an abstract data type that represents some sort of container,
such as a list or dictionary. We’d like to be able to do something to every
element of the container; say, count them up. How can we write operations on
the abstract data type to allow this, without exposing the implementation?
To make the problem more concrete, let’s suppose we have an abstract data type
that represents the set of all non-negative numbers less than some fixed bound.
The core of its interface might look like this:
/*
* Abstract data type representing the set of numbers from 0 to
* bound-1 inclusive, where bound is passed in as an argument at creation.
*/
typedef struct nums *Nums;

/* Create a Nums object with given bound. */

Nums nums_create(int bound);

/* Destructor */
void nums_destroy(Nums);

/* Returns 1 if nums contains element, 0 otherwise */

int nums_contains(Nums nums, int element);
#include <stdlib.h>
#include "nums.h"

struct nums {
int bound;
};

Nums nums_create(int bound)

{

380
 struct nums *n;
n = malloc(sizeof(*n));
n->bound = bound;
return n;
}

void nums_destroy(Nums n) { free(n); }

int nums_contains(Nums n, int element)

{
return element >= 0 && element < n->bound;
}
From the outside, a Nums acts like the set of numbers from 0 to bound - 1;
nums_contains will insist that it contains any int that is in this set and contains
no int that is not in this set.
Let’s suppose now that we want to loop over all elements of some Nums, say to add
them together. In particular, we’d like to implement the following pseudocode,
where nums is some Nums instance:
sum = 0;
for(each i in nums) {
sum += i;
}
One way to do this would be to build the loop into some operation in nums.c,
including its body. But we’d like to be able to substitute any body for the
sum += i line. Since we can’t see the inside of a Nums, we need to have some
additional operation or operations on a Nums that lets us write the loop. How
can we do this?

5.1.1.1 Option 1: Function that returns a sequence

A data-driven approach might be to add a nums_contents function that returns
a sequence of all elements of some instance, perhaps in the form of an array or
linked list. The advantage of this approach is that once you have the sequence,
you don’t need to worry about changes to (or destruction of) the original object.
The disadvantage is that you have to deal with storage management issues, and
have to pay the costs in time and space of allocating and filling in the sequence.
This can be particularly onerous for a “virtual” container like Nums, since we
could conceivably have a Nums instance with billions of elements.
Bearing these facts in mind, let’s see what this approach might look like. We’ll
define a new function nums_contents that returns an array of ints, terminated
by a -1 sentinel:
int *

381
nums_contents(Nums n)
{
int *a;
int i;
a = malloc(sizeof(*a) * (n->bound + 1));
for(i = 0; i < n->bound; i++) a[i] = i;
a[n->bound] = -1;
return a;
}
We might use it like this:
sum = 0;
contents = nums_contents(nums);
for(p = contents; *p != -1; p++) {
sum += *p;
}
free(contents);
Despite the naturalness of the approach, returning a sequence in this case leads
to the most code complexity of the options we will examine.

5.1.1.2 Option 2: Iterator with first/done/next operations

If we don’t want to look at all the elements at once, but just want to process them
one at a time, we can build an iterator. An iterator is an object that allows you to
step through the contents of another object, by providing convenient operations
for getting the first element, testing when you are done, and getting the next
element if you are not. In C, we try to design iterators to have operations that
fit well in the top of a for loop.
For the Nums type, we’ll make each Nums its own iterator. The new operations
are given here:
int nums_first(Nums n) { return 0; }
int nums_done(Nums n, int val) { return val >= n->bound; }
int nums_next(Nums n, int val) { return val+1; }
And we use them like this:
sum = 0;
for(i = nums_first(nums); !nums_done(nums, i); i = nums_next(nums, i)) {
sum += i;
}
Not only do we completely avoid the overhead of building a sequence, we also
get much cleaner code. It helps in this case that all we need to find the next
value is the previous one; for a more complicated problem we might have to

382
create and destroy a separate iterator object that holds the state of the loop.
But for many tasks in C, the first/done/next idiom is a pretty good one.

5.1.1.3 Option 3: Iterator with function argument

Suppose we have a very complicated iteration, say one that might require several
nested loops or even a recursion to span all the elements. In this case it might
be very difficult to provide first/done/next operations, because it would be hard
to encode the state of the iteration so that we could easily pick up in the next
operation where we previously left off. What we’d really like to do is to be
able to plug arbitrary code into the innermost loop of our horrible iteration
procedure, and do it in a way that is reasonably typesafe and doesn’t violate
our abstraction barrier. This is a job for function pointers, and an example of
the functional programming style in action.
We’ll define a nums_foreach function that takes a function as an argument:
void nums_foreach(Nums n, void (*f)(int, void *), void *f_data)
{
int i;
for(i = 0; i < n->bound; i++) f(i, f_data);
}
The f_data argument is used to pass extra state into the passed-in function f;
it’s a void * because we want to let f work on any sort of extra state it likes.
Now to do our summation, we first define an extra function sum_helper, which
adds each element to an accumulator pointed to by f_data:
static void sum_helper(int i, void *f_data)
{
*((int *) f_data) += i;
}
We then feed sum_helper to the nums_foreach function:
sum = 0;
nums_foreach(nums, sum_helper, (void *) &sum);
There is a bit of a nuisance in having to define the auxiliary sum_helper function
and in all the casts to and from void, but on the whole the complexity of this
solution is not substantially greater than the first/done/next approach. Which
you should do depends on whether it’s harder to encapsulate the state of the
iterator (in which case the functional approach is preferable) or of the loop body
(in which case the first/done/next approach is preferable), and whether you
need to bail out of the loop early (which would require special support from
the foreach procedure, perhaps checking a return value from the function).
However, it’s almost always straightforward to encapsulate the state of a loop

383
body; just build a struct containing all the variables that it uses, and pass a
pointer to this struct as f_data.

5.1.2 Closures

A closure is a function plus some associated state. A simple way to implement

closures in C is to use a static local variable, but then you only get one. Better
is to allocate the state somewhere and pass it around with the function. For
example, here’s a simple functional implementation of infinite sequences:
/* a sequence is an object that returns a new value each time it is called */
struct sequence {
int (*next)(void *data);
void *data;
};

typedef struct sequence *Sequence;

Sequence
create_sequence(int (*next)(void *data), void *data)
{
Sequence s;

s = malloc(sizeof(*s));
assert(s);

s->next = next;
s->data = data;

return s;
}

int
sequence_next(Sequence s)
{
return s->next(s->data);
}
And here are some examples of sequences:
/* make a constant sequence that always returns x */
static int
constant_sequence_next(void *data)
{
return *((int *) data);
}

384
Sequence
constant_sequence(int x)
{
int *data;

data = malloc(sizeof(*data));
if(data == 0) return 0;

*data = x;

return create_sequence(constant_sequence_next, data);

}

/* make a sequence x, x+a, x+2*a, x+3*a, ... */

struct arithmetic_sequence_data {
int cur;
int step;
};

static int
arithmetic_sequence_next(void *data)
{
struct arithmetic_sequence_data *d;

d = data;
d->cur += d->step;

return d->cur;
}

Sequence
arithmetic_sequence(int x, int a)
{
struct arithmetic_sequence_data *d;

d = malloc(sizeof(*d));
if(d == 0) return 0;

d->cur = x - a; /* back up so first value returned is x */

d->step = a;

return create_sequence(arithmetic_sequence_next, d);

}

/* Return the sum of two sequences */

385
static int
add_sequences_next(void *data)
{
Sequence *s;

s = data;
return sequence_next(s[0]) + sequence_next(s[1]);
}

Sequence
add_sequences(Sequence s0, Sequence s1)
{
Sequence *s;

s = malloc(2*sizeof(*s));
if(s == 0) return 0;

s[0] = s0;
s[1] = s1;

return create_sequence(add_sequences_next, s);

}

/* Return the sequence x, f(x), f(f(x)), ... */

struct iterated_function_sequence_data {
int x;
int (*f)(int);
}

static int
iterated_function_sequence_next(void *data)
{
struct iterated_function_sequence_data *d;
int retval;

d = data;

retval = d->x;
d->x = d->f(d->x);

return retval;
}

Sequence
iterated_function_sequence(int (*f)(int), int x0)
{

386
 struct iterated_function_sequence_data *d;

d = malloc(sizeof(*d));
if(d == 0) return 0;

d->x = x0;
d->f = f;

return create_sequence(iterated_function_sequence_next, d);

}
Note that we haven’t worried about how to free the data field inside a Sequence,
and indeed it’s not obvious that we can write a generic data-freeing routine
since we don’t know what structure it has. The solution is to add more function
pointers to a Sequence, so that we can get the next value, get the sequence to
destroy itself, etc. When we do so, we have gone beyond building a closure to
building an object.

5.1.3 Objects

Here’s an example of a hierarchy of counter objects. Each counter object has (at
least) three operations: reset, next, and destroy. To call the next operation
on counter c we include c and the first argument, e.g. c->next(c) (one could
write a wrapper to enforce this).
The main trick is that we define a basic counter structure and then extend it
to include additional data, using lots of pointer conversions to make everything
work.
/* use preprocessor to avoid rewriting these */
#define COUNTER_FIELDS \
void (*reset)(struct counter *); \
int (*next)(struct counter *); \
void (*destroy)(struct counter *);

struct counter {
COUNTER_FIELDS
};

typedef struct counter *Counter;

/* minimal counter--always returns zero */

/* we don't even allocate this, just have one global one */
static void noop(Counter c) { ; }
static int return_zero(Counter c) { return 0; }
static struct counter Zero_counter = { noop, return_zero, noop };

387
Counter
make_zero_counter(void)
{
return &Zero_counter;
}

/* a fancier counter that iterates a function sequence */

/* this struct is not exported anywhere */
struct ifs_counter {

/* copied from struct counter declaration */

COUNTER_FIELDS

/* new fields */
int init;
int cur;
int (*f)(int); /* update rule */
};

static void
ifs_reset(Counter c)
{
struct ifs_counter *ic;

ic = (struct ifs_counter *) c;

ic->cur = ic->init;
}

static void
ifs_next(Counter c)
{
struct ifs_counter *ic;
int ret;

ic = (struct ifs_counter *) c;

ret = ic->cur;
ic->cur = ic->f(ic->cur);

return ret;
}

Counter
make_ifs_counter(int init, int (*f)(int))

388
{
struct ifs_counter *ic;

ic = malloc(sizeof(*ic));

ic->reset = ifs_reset;
ic->next = ifs_next;
ic->destroy = (void (*)(struct counter *)) free;

ic->init = init;
ic->cur = init;
ic->f = f;

/* it's always a Counter on the outside */

return (Counter) ic;
}
A typical use might be
static int
times2(int x)
{
return x*2;
}

void
print_powers_of_2(void)
{
int i;
Counter c;

c = make_ifs_counter(1, times2);

for(i = 0; i < 10; i++) {

printf("%d\n", c->next(c));
}

c->reset(c);

for(i = 0; i < 20; i++) {

printf("%d\n", c->next(c));
}

c->destroy(c);
}

389
5.2 Suffix arrays

These are notes on practical implementations of suffix arrays, which are a data
structure for searching quickly for substrings of a given large string.

5.2.1 Why do we want to do this?

• Answer from the old days: Fast string searching is the key to dealing with
mountains of information. Why, a modern (c. 1960) electronic computer
can search the equivalent of hundreds of pages of text in just a few hours. . .
• More recent answer:
– We still need to search enormous corpuses of text (see http://www.
google.com).
– Algorithms for finding long repeated substrings or patterns can be
useful for data compression) or detecting plagiarism.
– Finding all occurrence of a particular substring in some huge corpus
is the basis of statistical machine translation.
– We are made out of strings over a particular finite alphabet GATC.
String search is a central tool in computational biology.

5.2.2 String search algorithms

Without preprocessing, searching an n-character string for an m-character sub-

string can be done using algorithms of varying degrees of sophistication, the
worst of which run in time O(nm) (run strncmp on each position in the big
string), and best of which run in time O(n + m) (run the Boyer-Moore string
search algorithm). But we are interested in the case where we can preprocess
our big string into a data structure that will let us do lots of searches for cheap.

5.2.3 Suffix trees and suffix arrays

Suffix trees and suffix arrays are data structures for representing texts that
allow substring queries like “where does this pattern appear in the text” or “how
many times does this pattern occur in the text” to be answered quickly. Both
work by storing all suffixes of a text, where a suffix is a substring that runs to the
end of the text. Of course, storing actual copies of all suffixes of an n-character
text would take O(n2 ) space, so instead each suffix is represented by a pointer
to its first character.
A suffix array stores all the suffixes sorted in dictionary order. For example,
the suffix array of the string abracadabra is shown below. The actual contents
of the array are the indices in the left-hand column; the right-hand shows the
corresponding suffixes.

390
11 \0
10 a\0
7 abra\0
0 abracadabra\0
3 acadabra\0
5 adabra\0
8 bra\0
1 bracadabra\0
4 cadabra\0
6 dabra\0
9 ra\0
2 racadabra\0
A suffix tree is similar, but instead using an array, we use some sort of tree
data structure to hold the sorted list. A common choice given an alphabet of
some fixed size k is a trie, in which each node at depth d represents a string of
length d, and its up to k children represent all (d + 1)-character extensions of the
string. The advantage of using a suffix trie is that searching for a string of length
m takes O(m) time, since we can just walk down the trie at the rate of one
node per character in m. A further optimization is to replace any long chain of
single-child nodes with a compressed edge labeled with the concatenation all the
characters in the chain. Such compressed suffix tries can not only be searched in
linear time but can also be constructed in linear time with a sufficiently clever
algorithm (we won’t actually do this here). Of course, we could also use a simple
balanced binary tree, which might be preferable if the alphabet is large.
The disadvantage of suffix trees over suffix arrays is that they generally require
more space to store all the internal pointers in the tree. If we are indexing a
huge text (or collection of texts), this extra space may be too expensive.

5.2.3.1 Building a suffix array

A straightforward approach to building a suffix array is to run any decent
comparison-based sorting algorithm on the set of suffixes (represented by pointers
into the text). This will take O(n log n) comparisons, but in the worst case each
comparison will take O(n) time, for a total of O(n2 log n) time. This is the
approach used in the sample code below.
The original suffix array paper by Manber and Myers (“Suffix arrays: a new
method for on-line string searches,” SIAM Journal on Computing 22(5):935-948,
1993) gives an O(n log n) algorithm, somewhat resembling radix sort, for building
suffix arrays in place with only a small amount of additional space. They also
note that for random text, simple radix sorting works well, since most suffixes
become distinguishable after about logk n characters (where k is the size of the
alphabet); this gives a cost of O(n log n) to do the sort, since radix sort only
looks at the bytes it needs to once. For a comparison-based sort, the same

391
assumption gives an O(n log2 n) running time; this is a factor of log n slower,
but this may be acceptable if programmer time is more important.
The fastest approach is to build a suffix tree in O(n) time and extract the suffix
array by traversing the tree. The only complication is that we need the extra
space to build the tree, although we get it back when we throw the tree away.

5.2.3.2 Searching a suffix array

Suppose we have a suffix array corresponding to an n-character text and we
want to find all occurrences in the text of an m-character pattern. Since the
suffixes are ordered, the easiest solution is to do binary search for the first and
last occurrences of the pattern (if any) using O(log n) comparisons. (The code
below does something even lazier than this, searching for some match and then
scanning linearly for the first and last matches.) Unfortunately, each comparison
may take as much as O(m) time, since we may have to check all m characters of
the pattern. So the total cost will be O(m log n) in the worst case.
By storing additional information about the longest common prefix of consecutive
suffixes, it is possible to avoid having to re-examine every character in the
pattern for every comparison, reducing the search cost to O(m + log n). With a
sufficiently clever algorithm, this information can be computed in linear time,
and can also be used to solve quickly such problems as finding the longest
duplicate substrings, or most frequently occurring strings. For more details, see
(Gusfield, Dan. Algorithms on Strings, Trees, and Sequences: Computer Science
and Computational Biology. Cambridge University Press, 1997, §7.14.4]).
Using binary search on the suffix array, most searching tasks are now easy:
• Finding if a substring appears in the array uses binary search directly.
• Finding all occurrences requires two binary searches, one for the first
occurrence and one for the last. If we only want to count the occurrences
and not return their positions, this takes O(m + log n) time. If we want
to return their positions, it takes O(m + log n + k) time, where k is the
number of times the pattern occurs.
• Finding duplicate substrings of length m or more can be done by looking
for adjacent entries in the array with long common prefixes, which takes
O(mn) time in the worst case if done naively (and O(n) time if we have
already computed longest common prefix information.

5.3 Burrows-Wheeler transform

Closely related to suffix arrays is the Burrows-Wheeler transform (Burrows

and Wheeler, A Block-Sorting Lossless Data Compression Algorithm, DEC
Systems Research Center Technical Report number 124, 1994), which is the basis
for some of the best currently known algorithms for text compression (it’s the
technique that is used, for example, in bzip2).

392
The idea of the Burrows-Wheeler Transform is to construct an array whose rows
are all cyclic shifts of the input string in dictionary order, and return the last
column of the array. The last column will tend to have long runs of identical
characters, since whenever some substring (like the appears repeatedly in the
input, shifts that put the first character t in the last column will put the rest
of the substring he in the first columns, and the resulting rows will tend to be
sorted together. The relative regularity of the last column means that it will
compress well with even very simple compression algorithms like run-length
encoding.
Below is an example of the Burrows-Wheeler transform in action, with $ marking
end-of-text. The transformed value of abracadabra$ is $drcraaaabba, the last
column of the sorted array; note the long run of a’s (and the shorter run of b’s).
abracadabra$ abracadabra$
bracadabra$a abra$abracad
racadabra$ab acadabra$abr
acadabra$abr adabra$abrac
cadabra$abra a$abracadabr
adabra$abrac bracadabra$a
dabra$abraca --> bra$abracada
abra$abracad cadabra$abra
bra$abracada dabra$abraca
ra$abracadab racadabra$ab
a$abracadabr ra$abracadab
$abracadabra $abracadabra
The most useful property of the Burrows-Wheeler transform is that it can be
inverted; this distinguishes it from other transforms that produce long runs like
simply sorting the characters. We’ll describe two ways to do this; the first is
less efficient, but more easily grasped, and involves rebuilding the array one
column at a time, starting at the left. Observe that the leftmost column is just
all the characters in the string in sorted order; we can recover it by sorting the
rightmost column, which we have to start off with. If we paste the rightmost and
leftmost columns together, we have the list of all 2-character substrings of the
original text; sorting this list gives the first two columns of the array. (Remember
that each copy of the string wraps around from the right to the left.) We can
then paste the rightmost column at the beginning of these two columns, sort
the result, and get the first three columns. Repeating this process eventually
reconstructs the entire array, from which we can read off the original string from
any row. The initial stages of this process for abracadabra$ are shown below:
$ a $a ab $ab abr
d a da ab dab abr
r a ra ac rac aca
c a ca ad cad ada
r a ra a$ ra$ a$a
a b ab br abr bra

393
a -> b ab -> br abr -> bra
a c ac ca aca cad
a d ad da ada dab
b r br ra bra rac
b r br ra bra ra$
a $ a$ $a a$a $ab
Rebuilding the entire array in this fashion takes O(n2 ) time and O(n2 ) space.
In their paper, Burrows and Wheeler showed that one can in fact reconstruct
the original string from just the first and last columns in the array in O(n) time.
Here’s the idea: Suppose that all the characters were distinct. Then after
reconstructing the first column we would know all pairs of adjacent characters.
So we could just start with the last character $ and regenerate the string by
appending at each step the unique successor to the last character so far. If all
characters were distinct, we would never get confused about which character
comes next.
The problem is what to do with pairs with duplicate first characters, like ab
and ac in the example above. We can imagine that each a in the last column is
labeled in some unique way, so that we can talk about the first a or the third a,
but how do we know which a is the one that comes before b or d?
The trick is to look closely at how the original sort works. Look at the rows in
the original transformation. If we look at all rows that start with a, the order
they are sorted in is determined by the suffix after a. These suffixes also appear
as the prefixes of the rows that end with a, since the rows that end with a are
just the rows that start with a rotated one position. It follows that all instances
of the same letter occur in the same order in the first and last columns. So if
we use a stable sort to construct the first column, we will correctly match up
instances of letters.
This method is shown in action below. Each letter is annotated uniquely with
a count of how many identical letters equal or precede it. Sorting recovers the
first column, and combining the last and first columns gives a list of unique
pairs of adjacent annotated characters. Now start with $1 and construct the full
sequence $1 a1 b1 r1 a3 c1 a4 d1 a2 b2 r2 a5 $1. The original string is
obtained by removing the end-of-string markers and annotations: abracadabra.
$1 a1
d1 a2
r1 a3
c1 a4
r2 a5
a1 b1
a2 --> b2
a3 c1
a4 d1
b1 r1

394
b2 r2
a5 $1
Because we are only sorting single characters, we can perform the sort in linear
time using counting sort. Extracting the original string also takes linear time if
implemented reasonably.

5.3.1 Suffix arrays and the Burrows-Wheeler transform

A useful property of the Burrows-Wheeler transform is that each row of the

sorted array is essentially the same as the corresponding row in the suffix array,
except for the rotated string prefix after the $ marker. This means, among
other things, that we can compute the Burrows-Wheeler transform in linear time
using suffix trees. Ferragina and Manzini (http://people.unipmn.it/~manzini/
papers/focs00draft.pdf) have further exploited this correspondence (and some
very clever additional ideas) to design compressed suffix arrays that compress
and index a text at the same time, so that pattern searches can be done directly
on the compressed text in time close to that needed for suffix array searches.

5.3.2 Sample implementation

As mentioned above, this is a pretty lazy implementation of suffix arrays, that

doesn’t include many of the optimizations that would be necessary to deal with
huge source texts.
/* we expose this so user can iterate through it */

struct suffixArray {
size_t n; /* length of string INCLUDING final null */
const char *string; /* original string */
const char **suffix; /* suffix array of length n */
};

typedef struct suffixArray *SuffixArray;

/* construct a suffix array */

/* it is a bad idea to modify string before destroying this */
SuffixArray suffixArrayCreate(const char *string);

/* destructor */
void suffixArrayDestroy(SuffixArray);

/* return number of occurrences of substring */

/* if non-null, index of first occurrence is place in first */
size_t

395
suffixArraySearch(SuffixArray, const char *substring, size_t *first);

/* return the Burrows-Wheeler transform of the underlying string

* as malloc'd data of length sa->n */
/* note that this may have a null in the middle somewhere */
char *suffixArrayBWT(SuffixArray sa);

/* invert BWT of null-terminated string, returning a malloc'd copy of original */

char *inverseBWT(size_t len, const char *s);
examples/suffixArray/suffixArray.h
#include <stdlib.h>
#include <assert.h>
#include <string.h>
#include <limits.h>

#include "suffixArray.h"

static int
saCompare(const void *s1, const void *s2)
{
return strcmp(*((const char **) s1), *((const char **) s2));
}

SuffixArray
suffixArrayCreate(const char *s)
{
size_t i;
SuffixArray sa;

sa = malloc(sizeof(*sa));
assert(sa);

sa->n = strlen(s) + 1;
sa->string = s;

sa->suffix = malloc(sizeof(*sa->suffix) * sa->n);

assert(sa->suffix);

/* construct array of pointers to suffixes */

for(i = 0; i < sa->n; i++) {
sa->suffix[i] = s+i;
}

/* this could be a lot more efficient */

qsort(sa->suffix, sa->n, sizeof(*sa->suffix), saCompare);

396
 return sa;
}

void
suffixArrayDestroy(SuffixArray sa)
{
free(sa->suffix);
free(sa);
}

size_t
suffixArraySearch(SuffixArray sa, const char *substring, size_t *first)
{
size_t lo;
size_t hi;
size_t mid;
size_t len;
int cmp;

len = strlen(substring);

/* invariant: suffix[lo] <= substring < suffix[hi] */

lo = 0;
hi = sa->n;

while(lo + 1 < hi) {

mid = (lo+hi)/2;
cmp = strncmp(sa->suffix[mid], substring, len);

if(cmp == 0) {
/* we have a winner */
/* search backwards and forwards for first and last */
for(lo = mid; lo > 0 && strncmp(sa->suffix[lo-1], substring, len) == 0; lo--);
for(hi = mid; hi < sa->n && strncmp(sa->suffix[hi+1], substring, len) == 0; hi++

if(first) {
*first = lo;
}

return hi - lo + 1;
} else if(cmp < 0) {
lo = mid;
} else {
hi = mid;
}

397
 }

return 0;
}

char *
suffixArrayBWT(SuffixArray sa)
{
char *bwt;
size_t i;

bwt = malloc(sa->n);
assert(bwt);

for(i = 0; i < sa->n; i++) {

if(sa->suffix[i] == sa->string) {
/* wraps around to nul */
bwt[i] = '\0';
} else {
bwt[i] = sa->suffix[i][-1];
}
}

return bwt;
}

char *
inverseBWT(size_t len, const char *s)
{
/* basic trick: stable sort of s gives successor indices */
/* then we just thread through starting from the nul */

size_t *successor;
int c;
size_t count[UCHAR_MAX+1];
size_t offset[UCHAR_MAX+1];
size_t i;
char *ret;
size_t thread;

successor = malloc(sizeof(*successor) * len);

assert(successor);

/* counting sort */
for(c = 0; c <= UCHAR_MAX; c++) {
count[c] = 0;

398
 }

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

count[(unsigned char) s[i]]++;
}

offset[0] = 0;

for(c = 1; c <= UCHAR_MAX; c++) {

offset[c] = offset[c-1] + count[c-1];
}

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

successor[offset[(unsigned char) s[i]]++] = i;
}

/* find the nul */

for(thread = 0; s[thread]; thread++);

/* thread the result */

ret = malloc(len);
assert(ret);

for(i = 0, thread = successor[thread]; i < len; i++, thread = successor[thread]) {

ret[i] = s[thread];
}

return ret;
}
examples/suffixArray/suffixArray.c
Here is a Makefile and test code: Makefile, testSuffixArray.c.
The output of make test shows all occurrences of a target string, the Burrows-
Wheeler transform of the source string (second-to-last line), and its inversion
(last line, which is just the original string):
$ make test
/bin/echo -n abracadabra-abracadabra-shmabracadabra | ./testSuffixArray abra
Count: 6
abra
abra-abr
abra-shm
abracada
abracada
abracada
aaarrrdddm\x00-rrrcccaaaaaaaaaaaashbbbbbb-

399
abracadabra-abracadabra-shmabracadabra

5.4 C++

Here we will describe some basic features of C++ that are useful for implementing
abstract data types. Like all programming languages, C++ comes with an
ideology, which in this case emphasizes object-oriented features like inheritance.
We will be ignoring this ideology and treating C++ as an improved version of C.
The goal here is not to teach you all of C++, which would take a while, but
instead to give you some hints for why you might want to learn C++ on your own.
If you decide to learn C++ for real, Bjarne Stroustrup’s The C++ Programming
Language is the definitive source. A classic tutorial here aimed at C programmers
introduces C++ features one at a time (some of these features have since migrated
into C). The web site http://www.cplusplus.com has extensive tutorials and
documentation.

5.4.1 Hello world

The C++ version of “hello world” looks like this:

#include <iostream>

int
main(int argc, const char **argv)
{
std::cout << "hi\n";

return 0;
}
examples/c++/helloworld.cpp
Compile this using g++ instead of gcc. Make shows how it is done:
$ make helloworld
g++ helloworld.cpp -o helloworld
Or we could use an explicit Makefile:
CPP=g++
CPPFLAGS=-g3 -Wall

helloworld: helloworld.o
$(CPP) $(CPPFLAGS) -o $@ $^
Now the compilation looks like this:

400
$ make helloworld
g++ -g3 -Wall -c -o helloworld.o helloworld.cpp
g++ -g3 -Wall -o helloworld helloworld.o
The main difference from the C version:
1. #include <stdio.h> is replaced by #include <iostream>, which gets
the C++ version of the stdio library.
2. printf("hi\n") is replaced by std::cout << "hi\n". The stream
std::cout is the C++ wrapper for stdout; you should read this variable
name as cout in the std namespace. The << operator is overloaded for
streams so that it sends its right argument out on its left argument (see
the discussion of operator overloading below). You can also do things like
std::cout << 37, std::cout << 'q', std::cout << 4.7, etc. These
all do pretty much what you expect.
If you don’t like typing std:: before all the built-in functions and variables, you
can put using namespace std somewhere early in your program, like this:
#include <iostream>

using namespace std;

int
main(int argc, const char **argv)
{
cout << "hi\n";

return 0;
}
examples/c++/helloworld_using.cpp

5.4.2 References

Recall that in C we sometime pass objects into function by reference instead of

by value, by using a pointer:
void increment(int *x)
{
(*x)++;
}
This becomes even more useful in C++, since many of the objects we are dealing
with are quite large, and can defend themselves against dangerous modifications
by restricting access to their components. So C++ provides a special syntax
allowing function parameters to be declared as call-by-reference rather than
call-by-value. The function above could be rewritten in C++ as

401
void increment(int &x)
{
x++;
}
The int &x declaration says that x is a reference to whatever variable is passed
as the argument to increment. A reference acts exactly like a pointer that has
already had * applied to it. You can even write &x to get a pointer to the original
variable if you want to for some reason.
As with pointers, it’s polite to mark a reference with const if you don’t intend
to modify the original object:
void reportWeight(const SumoWrestler &huge)
{
cout << huge.getWeight();
}
References are also used as a return type to chain operators together; in the
expression
cout << "hi" << '\n';
the return type of the first << operator is an ostream & reference (as is cout);
this means that the '\n' gets sent to the same object. We could make the
return value be just an ostream, but then cout would be copied, which could
be expensive and would mean that the copy was no longer working on the same
internal state as the original. This same trick is used when overloading the
assignment operator.

5.4.3 Function overloading

C++ lets you define multiple functions with the same name, where the choice of
which function to call depends on the type of its arguments. Here is a program
that demonstrates this feature:
#include <iostream>

using namespace std;

const char *
typeName(int x)
{
return "int";
}

const char *
typeName(double x)

402
{
return "double";
}

const char *
typeName(char x)
{
return "char";
}

int
main(int argc, const char **argv)
{
cout << "The type of " << 3 << " is " << typeName(3) << ".\n";
cout << "The type of " << 3.1 << " is " << typeName(3.1) << ".\n";
cout << "The type of " << 'c' << " is " << typeName('c') << ".\n";

return 0;
}
examples/c++/functionOverloading.cpp
And here is what it looks like when we compile and run it:
$ make functionOverloading
g++ functionOverloading.cpp -o functionOverloading
$ ./functionOverloading
The type of 3 is int.
The type of 3.1 is double.
The type of c is char.
Internally, g++ compiles three separate functions with different (and ugly) names,
and when you use typeName on an object of a particular type, g++ picks the one
whose type matches. This is similar to what happens with built-in operators in
straight C, where + means different things depending on whether you apply it
to a pair of ints, a pair of doubles, or a pointer and an int, but C++ lets you
do it with your own functions.

5.4.4 Classes

C++ allows you to declare classes that look suspiciously like structs. The main
differences between a class and a C-style struct are that (a) classes provide
member functions or methods that operate on instances of the class and
that are called using a struct-like syntax; and (b) classes can distinguish between
private members (only accessible to methods of the class) and public members
(accessible to everybody).

403
In C, we organize abstract data types by putting the representation in a struct
and putting the operations on the data type in functions that work on this struct,
often giving the functions a prefix that hints at the type of its target (mostly to
avoid namespace collisions). Classes in C++ make this connection between a
data structure and the operations on it much more explicit.
Here is a simple example of a C++ class in action:
#include <iostream>

using namespace std;

/* counters can be incremented or read */

class Counter {
int value; /* private value */
public:
Counter(); /* constructor with default value */
Counter(int); /* constructor with specified value */
~Counter(); /* useless destructor */
int read(); /* get the value of the counter */
void increment(); /* add one to the counter */
};

Counter::Counter() { value = 0; }
Counter::Counter(int initialValue) { value = initialValue; }
Counter::~Counter() { cerr << "counter de-allocated with value " << value << '\n'; }
int Counter::read() { return value; }
void Counter::increment() { value++; }

int
main(int argc, const char **argv)
{
Counter c;
Counter c10(10);

cout << "c starts at " << c.read() << '\n';

c.increment();
cout << "c after one increment is " << c.read() << '\n';

cout << "c10 starts at " << c10.read() << '\n';

c.increment();
c.increment();
cout <<"c10 after two increments is " << c10.read() << '\n';

return 0;
}

404
examples/c++/counter.cpp
Things to notice:
1. In the class Counter declaration, the public: label introduces the public
members of the class. The member value is only accessible to member
functions of Counter. This enforces much stronger information hiding
than the default in C, although one can still use void * trickery to hunt
down and extract supposedly private data in C++ objects.
2. In addition to the member function declarations in the class declara-
tion, we also need to provide definitions. These look like ordinary func-
tion definitions, except that the class name is prepended using :: as in
Counter::read.
3. Member functions are called using struct access syntax, as in c.read().
Conceptually, each instance of a class has its own member functions, so
that c.read is the function for reading c while c10.read is the function
for reading c10. Inside a member function, names of class members refer to
members of the current instance; value inside c.read is c.value (which
otherwise is not accessible, since c.value is not public).
4. Two special member functions are Counter::Counter() and Counter::Counter(int).
These are constructors, and are identifiable as such because they are
named after the class. A constructor is called whenever a new instance
of the class is created. If you create an instance with no arguments
(as in the declaration Counter c;), you get the constructor with no
arguments. If you create an instance with arguments (as in the declaration
Counter c10(10);), you get the version with the appropriate arguments.
This is just another example of function overloading. If you don’t define
any constructors, C++ supplies a default constructor that takes no
arguments and does nothing. Note that constructors don’t have a return
type (you don’t need to preface them with void).
5. The special member function Counter::~Counter() is a destructor; it is
called when an object of type Counter is de-allocated (say, when returning
from a function with a local variable of this type). This particular destructor
is not very useful. Destructors are mostly important for objects that allocate
their own storage that needs to be de-allocated when the object is; see the
section on storage allocation below.
Compiling and running this program gives the following output. Note that the
last two lines are produced by the destructor.
c starts at 0
c after one increment is 1
c10 starts at 10
c10 after two increments is 10
counter de-allocated with value 10
counter de-allocated with value 3
One subtle difference between C and C++ is that C++ uses empty parentheses

405
() for functions with no arguments, where C would use (void). This is a bit
of a historical artifact, having to do with C allowing () for functions whose
arguments are not specified in the declaration (which was standard practice
before ANSI C).
Curiously, C++ also allows you to declare structs, with the interpretation that
a struct is exactly like a class except that all members are public by default.
So if you change class to struct in the program above, it will do exactly the
same thing. In practice, nobody who codes in C++ does this; the feature is
mostly useful to allow C code with structs to mix with C++ code.

5.4.5 Operator overloading

Sometimes when you define a new class, you also want to define new interpre-
tations of operators on that class. Here is an example of a class that defines
elements of the max-plus algebra over ints. This gives us objects that act
like ints, except that the + operator now returns the larger of its arguments and
the * operator now returns the sum.25
The mechanism in C++ for doing this is to define member functions with names
operatorsomething where something is the name of the operator we want to
define. These member functions take one less argument that the operator they
define; in effect, x + y becomes syntactic sugar for x.operator+(y) (which,
amazingly, is actually legal C++). Because these are member functions, they
are allowed to access members of other instances of the same class that would
normally be hidden.
This same mechanism is also used to define automatic type conversions out
of a type: the MaxPlus::operator int() function allows C++ to convert a
MaxPlus object to an int whenever it needs to (for example, to feed it to cout).
(Automatic type conversions into a type happen if you provide an appropriate
constructor.)
#include <iostream>
#include <algorithm> // for max

using namespace std;

/* act like ints, except + does max and * does addition */

class MaxPlus {
int value;
25 This otherwise insane-looking modification is useful for modeling scheduling problems,

where a+b is the time to do a and b in parallel, and a*b is the time to do a and b sequentially.
The reason for making the first case + and the second case * is because this makes the
distributive law a*(b+c) = (a*b)+(a*c) work. It also allows tricks like matrix multiplication
using the standard definition. See http://maxplus.org for more than you probably want to
know about this.

406
public:
MaxPlus(int);
MaxPlus operator+(const MaxPlus &);
MaxPlus operator*(const MaxPlus &);
operator int();
};

MaxPlus::MaxPlus(int x) { value = x; }

MaxPlus
MaxPlus::operator*(const MaxPlus &other)
{
return MaxPlus(value + other.value);
}

MaxPlus
MaxPlus::operator+(const MaxPlus &other)
{
/* std::max does what you expect */
return MaxPlus(max(value, other.value));
}

MaxPlus::operator int() { return value; }

int
main(int argc, const char **argv)
{
cout << "2+3 == " << (MaxPlus(2) + MaxPlus(3)) << '\n';
cout << "2*3 == " << (MaxPlus(2) * MaxPlus(3)) << '\n';

return 0;
}
examples/c++/maxPlus.cpp
Avoid the temptation to overuse operator overloading, as it can be dangerous if
used to obfuscate what an operator normally does:
MaxPlus::operator--() { godzilla.eat(tokyo); }
The general rule of thumb is that you should probably only do operator overload-
ing if you really are making things that act like numbers (yes, cout << violates
this).
Automatic type conversions can be particularly dangerous. The line
cout << (MaxPlus(2) + 3) << '\n';
is ambiguous: should the compiler convert MaxPlus(2) to an int using the

407
MaxPlus(int) constructor and use ordinary integer addition or convert 3 to
a MaxPlus using MaxPlus::operator int() and use funky MaxPlus addition?
Fortunately most C++ compilers will complain about the ambiguity and fail
rather than guessing wrong.

5.4.6 Templates

One of the things we kept running into in this class was that if we defined a
container type like a hash table, binary search tree, or priority queue, we had
to either bake in the type of the data it held or do horrible tricks with void *
pointers to work around the C type system. C++ includes a semi-principled
work-around for this problem known as templates. These are essentially macros
that take a type name as an argument, that are expanded as needed to produce
functions or classes with specific types (see Macros for an example of how to do
this if you only have C).
Typical use is to prefix a definition with template <class T> and then use T
as a type name throughout:
template <class T>
T add1(T x)
{
return x + ((T) 1);
}
Note the explicit cast to T of 1; this avoids ambiguities that might arise with
automatic type conversions.
If you put this definition in a program, you can then apply add1 to any type
that has a + operator and that you can convert 1 to. For example, the output of
this code fragment:
cout << "add1(3) == " << add1(3) << '\n';
cout << "add1(3.1) == " << add1(3.1) << '\n';
cout << "add1('c') == " << add1('c') << '\n';
cout << "add1(MaxPlus(0)) == " << add1(MaxPlus(0)) << '\n';
cout << "add1(MaxPlus(2)) == " << add1(MaxPlus(2)) << '\n';
is
add1(3) == 4
add1(3.1) == 4.1
add1('c') == d
add1(MaxPlus(0)) == 1
add1(MaxPlus(2)) == 2
By default, C++ will instantiate a template to whatever type fits in its argument.
If you want to force a particular version, you can put the type in angle brackets
after the name of whatever you defined. For example,

408
 cout << "add1<int>(3.1) == " << add1<int>(3.1) << '\n';
produces
add1<int>(3.1) == 4
because add1<int> forces its argument to be converted to an int (truncating
to 3) before adding one to it.
Because templates are really macros that get expanded as needed, it is common
to put templates in header (.h) files rather than in .cpp files. See the stack
implementation below for an example of this.

5.4.7 Exceptions

C provides no built-in mechanism for signaling that something bad happened.

So C programmers are left to come up with ad-hoc mechanisms like:
1. Calling abort to kill the program, either directly or via assert.
2. Calling exit with a nonzero exit code.
3. Returning a special error value from a function. This is often done in
library routines, because it’s rude for a library routine not to give the caller
a chance to figure out how to deal with the error. But it means coming up
with some special error value that won’t be returned normally, and these
can vary widely from one routine to another (null pointers, -1, etc.)
C++ provides a standard mechanism for signaling unusual events known as
exceptions. The actual mechanism is similar to return: the throw statement
throws an exception that may be caught by a try..catch statement anywhere
above it on the execution stack (not necessarily in the same function). Example:
#include <iostream>

using namespace std;

int fail()
{
throw "you lose";

return 5;
}

int
main(int argc, const char **argv)
{
try {
cout << fail() << '\n';
}

409
 catch(const char *s) {
cerr << "Caught error: " << s << '\n';
}

return 0;
}
examples/c++/exception.cpp
In action:
$ make exception
g++ -g3 -Wall exception.cpp -o exception
$ ./exception
Caught error: you lose
Note the use of cerr instead of cout. This sends the error message to stderr.
A try..catch statement will catch an exception only if the type matches the
type of the argument to the catch part of the statement. This can be used to
pick and choose which exceptions you want to catch. See http://www.cplusplus.
com/doc/tutorial/exceptions/ for some examples and descriptions of some C++
standard library exceptions.

5.4.8 Storage allocation

C++ programs generally don’t use malloc and free, but instead use the built-in
C++ operators new and delete. The advantage of new and delete is that they
know about types: not only does this mean that you don’t have to play games
with sizeof to figure out how much space to allocate, but if you allocate a new
object from a class with a constructor, the constructor gets called to initialize
the object, and if you delete an object, its destructor (if it has one) is called.
There are two versions of new and delete, depending on whether you want
to allocate just one object or an array of objects, plus some special syntax for
passing constructor arguments:
• To allocate a single object, use new type.
• To allocate an array of objects, use new type[size]. As with malloc, both
operations return a pointer to type.
• If you want to pass arguments to a constructor for type, use new
type(args). This only works with the single-object version, so you can’t
do new SomeClass[12] unless SomeClass has a constructor that takes no
arguments.
• To de-allocate a single object, use delete pointer-to-object.
• To de-allocate an array, use delete [] pointer-to-base-of-array. Mixing
new with delete [] or vice versa is an error that may or may not be

410
 detected by the compiler. Mixing either with malloc or free is a very bad
idea.
The program below gives examples of new and delete in action:
#include <iostream>
#include <cassert>

using namespace std;

class Noisy {
int id;
public:
Noisy(int); // create a noisy object with this id
~Noisy();
};

Noisy::Noisy(int initId) {
id = initId;
cout << "Noisy object created with id " << id << '\n';
}

Noisy::~Noisy() {
cout << "Noisy object destroyed with id " << id << '\n';
}

int
main(int argc, const char **argv)
{
int *p;
int *a;
const int n = 100;
Noisy n1(1);
Noisy *n2;

p = new int;
a = new int[n];
n2 = new Noisy(2);

*p = 5;
assert(*p == 5);

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

a[i] = i;
}

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

411
 assert(a[i] == i);
}

delete [] a;
delete p;
delete n2;

return 0;
}
examples/c++/allocation.cpp

5.4.8.1 Storage allocation inside objects

Inside objects, storage allocation gets complicated. The reason is that if the
object is copied, either by an assignment or by being passed as a call-by-value
parameter, the storage pointed to by the object will not be copied. This can
lead to two different objects that share the same internal data structures, which
is usually not something you want. Furthermore, when the object is deallocated,
it’s necessary to also deallocate any space it allocated, which can be done inside
the object’s destructor.
To avoid all these problems, any object of type T that uses new needs to have all
of:
1. A destructor T::~T().
2. A copy constructor T::T(const T &), which is a constructor that takes a
reference to another object of the same type as an argument and copies its
contents.
3. An overloaded assignment operator T::operator=(const T &) that does
the same thing, but also deallocates any internal storage of the current
object before copying new data in place of it (or possibly just copies the
contents of internal storage without doing any allocation and deallocation).
The overloaded assignment operator is particularly tricky, because you have
to make sure it doesn’t destroy the contents of the object if somebody writes
the useless self-assignment a = a, and you also need to return a reference
to *this so that you can chain assignments together as in a = b = c.
Here is an example of a Stack class that includes all of these members. Note
that it is defined using templates so we can make a stack of any type we like.
template <class T>
class Stack {
static const int initialSize = 32; /* static means this is shared across entire class
int top;
int size;
T* contents;
public:

412
 Stack(); /* create a new empty stack */

/* the unholy trinity of complex C++ objects */

~Stack(); /* destructor */
Stack(const Stack &); /* copy constructor */
Stack& operator=(const Stack &); /* overloaded assignment */

void push(T); /* push an element onto the stack */

int isEmpty(); /* return 1 if empty */
T pop(); /* pop top element from stack */
};

template <class T>

Stack<T>::Stack()
{
size = initialSize;
top = 0;
contents = new T[size];
}

template <class T>

Stack<T>::~Stack()
{
delete [] contents;
}

template <class T>

Stack<T>::Stack(const Stack<T> &other)
{
size = other.size;
top = other.top;
contents = new T[size];

for(int i = 0; i < top; i++) {

contents[i] = other.contents[i];
}
}

template <class T>

Stack<T> &
Stack<T>::operator=(const Stack<T> &other)
{
if(&other != this) {
/* this is a real assignment */

delete [] contents;

413
 size = other.size;
top = other.top;
contents = new T[size];

for(int i = 0; i < top; i++) {

contents[i] = other.contents[i];
}
}

return *this;
}

template <class T>

void
Stack<T>::push(T elt)
{
if(top >= size) {
int newSize = 2*size;
T *newContents = new T[newSize];

for(int i = 0; i < top; i++) {

newContents[i] = contents[i];
}

delete [] contents;

contents = newContents;
size = newSize;
}

contents[top++] = elt;
}

template <class T>

T
Stack<T>::pop()
{
if(top > 0) {
return contents[--top];
} else {
throw "stack empty";
}
}
examples/c++/stack/stack.h

414
Here is some code demonstrating use of the stack:
#include <iostream>

#include "stack.h"

using namespace std;

int
main(int argc, const char **argv)
{
Stack<int> s;
Stack<int> s2;

try {
s.push(1);
s.push(2);
s.push(3);

s2 = s;

cout << s.pop() << '\n';

cout << s.pop() << '\n';
cout << s.pop() << '\n';

cout << s2.pop() << '\n';

cout << s2.pop() << '\n';
cout << s2.pop() << '\n';

try {
s2.pop();
} catch(const char *err) {
cout << "Caught expected exception " << err << '\n';
}

for(int i = 0; i < 1000; i++) {

s.push(i);
}

cout << s.pop() << '\n';

} catch(const char *err) {
cerr << "Caught error " << err << '\n';
}

return 0;
}

415
examples/c++/stack/testStack.cpp

5.4.9 Standard library

C++ has a large standard library that includes implementations of many of

the data structures we’ve seen in this class. In most situations, it is easier to
use the standard library implementations than roll your own, although you
have to be careful to make sure you understand just what the standard library
implementations do. For example, here is a reimplementation of the main routine
from testStack.cpp using the stack template from #include <stack>.
#include <iostream>
#include <stack>

using namespace std;

int
main(int argc, const char **argv)
{
stack<int> s;
stack<int> s2;

s.push(1);
s.push(2);
s.push(3);

s2 = s;

cout << s.top() << '\n'; s.pop();

cout << s.top() << '\n'; s.pop();
cout << s.top() << '\n'; s.pop();

cout << s2.top() << '\n'; s2.pop();

cout << s2.top() << '\n'; s2.pop();
cout << s2.top() << '\n'; s2.pop();

for(int i = 0; i < 1000; i++) {

s.push(i);
}

cout << s.top() << '\n';

return 0;
}
examples/c++/stack/stdStack.cpp

416
One difference between the standard stack and our stack is that std::stack’s
pop member function doesn’t return anything. So we have to use top to get the
top element before popping it.
There is a chart of all the standard library data structures at http://www.
cplusplus.com/reference/stl/.

5.4.10 Things we haven’t talked about

The main thing we’ve omitted here is any discussion of object-oriented features
of C++, particularly inheritance. These are not immediately useful for the
abstract-data-type style of programming we’ve used in CS223, but can be helpful
for building more complicated systems, where we might want to have various
specialized classes of objects that can all be approached using a common interface
represented by a class that they inherit from. If you are interested in exploring
these tools further, the CS department occasionally offers a class on object-
oriented programming; Mike Fischer’s lecture notes from the last time this
course was offered can be found at http://zoo.cs.yale.edu/classes/cs427/2011a/
lectures.html.

5.5 Testing during development

It is a truth universally acknowledged that test code should be written early in

the development process. Unfortunately, most programmers (including me) tend
to assume that a program will work on the first attempt and there’s not much
point in testing it anyway, so writing and running test code often gets deferred
indefinitely. The solution is to write the test code first, and run it directly
from your Makefile every time you save and compile your program. Not only
will this guarantee that your program actually works when you are done (or at
least passes the tests you thought of), it allows you to see how the program is
improving with each positive change, and prevents you from accidentally making
new negative changes that break things that used to work.
Going one step further, we can often write our interface and test code first, build
a non-working stub implementation, and then slowly flesh out the missing pieces
until the implementation passes all the tests. This way there is always some
obvious step to do next, and we don’t find ourselves stuck staring at an empty
file.

5.5.1 Unit tests

A straightforward approach to testing is to include test code with every unit in

your program, where a unit is any part of the program that can be sensibly run

417
by itself. Typically, this will be a single function or a group of functions that
together implement some data structure.
In C, these will often make up the contents of a single source file. Though this is
probably not the best approach if you are building a production-quality testing
framework, a simple way to include unit tests in a program is to append to each
source file a test main function that can be enabled by defining a macro (I like
TEST_MAIN). You can then build this file by itself with the macro defined to get
a stand-alone test program for just this code.

5.5.1.1 What to put in the test code

Ideally, you want to use enough different inputs that every line of code in your
program is reached by some test, a goal called code coverage. For complex
programs, this may be hard to achieve, and there are programs, such as the
gcov program that comes with gcc, that will analyze how much code coverage
you get out of your tests. For simple programs, we can just try to come up with
a set of inputs that covers all our bases.
Testing can be done as black-box testing, where the test code assumes no
knowledge of the implementation, or white-box testing, where the test code
has direct access to the implementation and can observe the effects of its actions.
Black-box testing is handy if your implementation may change, and it is generally
a good idea to write black-box tests first. White-box testing can be useful if
some states of the data structure are hard to reach otherwise, or if black-box
testing is not very informative about why a particular operation is failing. The
example given below uses both.

5.5.1.2 Example
Here is an example of a simple data structure with some built-in test code
conditionally compiled by defining TEST_MAIN. The data structure implements a
counter with built-in overflow protection. The counter interface does not provide
the ability to read the counter value; instead, the user can only tell if it is zero
or not.
Because the counter is implemented internally as a uint64_t, black-box testing
of what happens with too many increments would take centuries. So we include
some white-box tests that directly access the counter value to set up this (arguably
unnecessary) test case.
The code is given below. We include both the interface file and the implemen-
tation, as well as a Makefile showing how to build and run the test program.
The Makefile includes some extra arguments to gcc to turn on the TEST_MAIN
macro and supply the extra information needed to run gcov. If you type make
test, it will make and run testCounter, and then run gcov to verify that we
did in fact hit all lines of code in the program.

418
/*
* Abstract counter type.
*
* You can increment it, decrement it, and test for zero.
*
* Increment and decrement operations return 1 if successful,
* 0 if the operation would cause underflow or overflow.
*/

typedef struct counter Counter;

/* make a new counter starting at 0 */

Counter *counterCreate(void);

/* destroy a counter */
void counterDestroy(Counter *);

/* return 1 if counter is 0, 0 otherwise */

int counterIsZero(const Counter *);

/* increment a counter, returns 1 if successful, 0 if increment would cause overflow */

int counterIncrement(Counter *);

/* decrement a counter, returns 1 if successful, 0 if decrement would cause underflow */

int counterDecrement(Counter *);
examples/unitTest/counter.h
#include <stdio.h>
#include <stdlib.h>
#include <assert.h>
#include "counter.h"

#include <stdint.h>

#define COUNTER_MAX (UINT64_MAX)

struct counter {
uint64_t value;
};

/* make a new counter starting at 0 */

Counter *
counterCreate(void)
{
Counter *c;

419
 c = malloc(sizeof(Counter));
assert(c);

c->value = 0;

return c;
}

/* destroy a counter */
void
counterDestroy(Counter *c)
{
free(c);
}

/* return 1 if counter is 0, 0 otherwise */

int
counterIsZero(const Counter *c)
{
return c->value == 0;
}

/* increment a counter, returns 1 if successful, 0 if increment would cause overflow */

int
counterIncrement(Counter *c)
{
if(c->value == COUNTER_MAX) {
return 0;
} else {
c->value++;
return 1;
}
}

/* decrement a counter, returns 1 if successful, 0 if decrement would cause underflow */

int
counterDecrement(Counter *c)
{
if(c->value == 0) {
return 0;
} else {
c->value--;
return 1;
}
}

420
#ifdef TEST_MAIN
int
main(int argc, char **argv)
{
Counter *c;

/* black box testing */

c = counterCreate(); /* 0 */

assert(counterIsZero(c));
assert(counterIncrement(c) == 1); /* 1 */
assert(!counterIsZero(c));
assert(counterIncrement(c) == 1); /* 2 */
assert(!counterIsZero(c));
assert(counterDecrement(c) == 1); /* 1 */
assert(!counterIsZero(c));
assert(counterDecrement(c) == 1); /* 0 */
assert(counterIsZero(c));
assert(counterDecrement(c) == 0); /* 0 */
assert(counterIsZero(c));
assert(counterIncrement(c) == 1); /* 1 */
assert(!counterIsZero(c));

counterDestroy(c);

/* white box testing */

c = counterCreate(); /* 0 */

assert(c->value == 0);
assert(counterIncrement(c) == 1); /* 1 */
assert(c->value == 1);
assert(counterIncrement(c) == 1); /* 2 */
assert(c->value == 2);
assert(counterDecrement(c) == 1); /* 1 */
assert(c->value == 1);
assert(counterDecrement(c) == 1); /* 0 */
assert(c->value == 0);
assert(counterDecrement(c) == 0); /* 0 */
assert(c->value == 0);
assert(counterIncrement(c) == 1); /* 1 */
assert(c->value == 1);

/* force counter value to COUNTER_MAX to test for overflow protection */

c->value = COUNTER_MAX; /* COUNTER_MAX */
assert(counterIncrement(c) == 0); /* COUNTER_MAX */
assert(c->value == COUNTER_MAX);

421
 assert(counterDecrement(c) == 1); /* COUNTER_MAX-1 */
assert(c->value == COUNTER_MAX-1);
assert(counterIncrement(c) == 1); /* COUNTER_MAX */
assert(c->value == COUNTER_MAX);

counterDestroy(c);

return 0;
}
#endif
examples/unitTest/counter.c
CC=c99
CFLAGS=-g3 -pedantic -Wall

all: seqprinter

seqprinter: main.o sequence.o

$(CC) $(CFLAGS) -o $@ $^

test: seqprinter
./seqprinter

# these rules say to rebuild main.o and sequence.o if sequence.h changes

main.o: main.c sequence.h
sequence.o: sequence.c sequence.h

clean:
$(RM) -f seqprinter *.o
examples/ADT/sequence/Makefile

5.5.2 Test harnesses

Here are some older notes on testing using a test harness that does some basic
tricks like catching segmentation faults so that a program can keep going even if
one test fails.

5.5.2.1 Module interface

The module will be a stack for storing integers.
Let’s start with the interface, which we’ll put in a file called stack.h:

5.5.2.1.1 stack.h

422
/*
* This is an "opaque struct"; it discourages people from looking at
* the inside of our structure. The actual definiton of struct stack
* is contained in stack.c.
*/
typedef struct stack *Stack;

/* constructor and destructor */

Stack stack_create(void); /* returns 0 on allocation error */
void stack_destroy(Stack);

/* push a new element onto the stack */

void stack_push(Stack , int new_element);

/* return 1 if the stack is empty, 0 otherwise */

int stack_isempty(Stack);

/* remove and return top element of stack */

/* returns STACK_EMPTY if stack is empty */
#define STACK_EMPTY (-1)
int stack_pop(Stack);
Our intent is that an Stack acts like a stack— we push things onto it using
stack_push, and then pull them off again in reverse order using stack_pop.
Ideally, we don’t ever pop the stack when it’s empty (which we can detect using
stack_isempty), but if we do, we have stack_pop return something well-defined.

5.5.2.2 Test code

Let’s write some test code to try this out. Because our initial stack implemen-
tation may be exceptionally bug-ridden, we’ll use a test harness that provides
macros for detecting and intercepting segmentation faults and similar disasters.
The various testing wrappers are defined in the files tester.h and tester.c,
from the chapter on testing; you should feel free to use it for your own purposes.
I’ve added line numbers in comments to all the TEST lines so we can find them
again later.

5.5.2.2.1 test-stack.c
#include <stdio.h>
#include <setjmp.h>
#include <signal.h>
#include <unistd.h>

#include <stdlib.h>

423
#include "stack.h"
#include "tester.h"

#define STRESS_TEST_ITERATIONS (1000000)

int
main(int argc, char **argv)
{
Stack s;
int i;

tester_init();

/* first we need to build one */

TRY { s = stack_create(); } ENDTRY;

/* 25 */ TEST_ASSERT(s != 0);

/* now we'll try pushing and popping a bit */

TRY { stack_push(s, 1); } ENDTRY;
TRY { stack_push(s, 2); } ENDTRY;
TRY { stack_push(s, 3); } ENDTRY;

/* 32 */ TEST(stack_isempty(s), 0);
/* 33 */ TEST(stack_pop(s), 3);
/* 34 */ TEST(stack_isempty(s), 0);
/* 35 */ TEST(stack_pop(s), 2);
/* 36 */ TEST(stack_isempty(s), 0);
/* 37 */ TEST(stack_pop(s), 1);
/* 38 */ TEST(stack_isempty(s), 1);
/* 39 */ TEST(stack_pop(s), STACK_EMPTY);
/* 40 */ TEST(stack_isempty(s), 1);
/* 41 */ TEST(stack_pop(s), STACK_EMPTY);

/* can we still push after popping too much? */

TRY { stack_push(s, 4); } ENDTRY;
/* 45 */ TEST(stack_isempty(s), 0);
/* 46 */ TEST(stack_pop(s), 4);
/* 47 */ TEST(stack_isempty(s), 1);
/* 48 */ TEST(stack_pop(s), STACK_EMPTY);
/* 49 */ TEST(stack_isempty(s), 1);

/* let's do some stress testing */

/* we won't use TEST for this because we might get too much output */
TRY {
for(i = 0; i < STRESS_TEST_ITERATIONS; i++) {

424
 stack_push(s, i);
}
for(i = 0; i < STRESS_TEST_ITERATIONS; i++) {
stack_push(s, 957);
if(stack_pop(s) != 957) {
/* 60 */ FAIL("wanted 957 but didn't get it");
abort();
}
}
for(i = STRESS_TEST_ITERATIONS - 1; i >= 0; i--) {
if(stack_isempty(s)) {
/* 66 */ FAIL("stack empty too early");
abort();
}
if(stack_pop(s) != i) {
/* 70 */ FAIL("got wrong value!");
abort();
}
}
} ENDTRY; /* 74 */

/* 76 */ TEST(stack_isempty(s), 1);

TRY { stack_destroy(s); } ENDTRY;

tester_report(stdout, argv[0]);
return tester_result();
}
There is a lot of test code here. In practice, we might write just a few tests to
start off with, and, to be honest, I didn’t write all of this at once. But you can
never have too many tests— if nothing else, they give an immediate sense of
gratification as the number of failed tests drops.

5.5.2.3 Makefile
• Finally, we’ll write a Makefile:

5.5.2.3.1 Makefile
CC=gcc
CFLAGS=-g3 -Wall -ansi -pedantic

all:

test: test-stack

425
 ./test-stack
@echo OK!

test-stack: test-stack.o tester.o stack.o

$(CC) $(CFLAGS) -o $@ $^

test-stack.o: stack.h tester.h

stack.o: stack.h
Note that we don’t provide a convenient shortcut for building test-stack
without running it. That’s because we want to run the test code every single
time.

5.5.3 Stub implementation

Of course, we still can’t compile anything, because we don’t have any implemen-
tation. Let’s fix that. To make it easy to write, we will try to add as little as
possible to what we already have in stack.h:

5.5.3.1 stack.c
#include <stdlib.h>
#include "stack.h"

struct stack { int dummy; };

Stack stack_create(void) { return malloc(sizeof(struct stack)); }
void stack_destroy(Stack s) { free(s); }
void stack_push(Stack s, int elem) { ; }
int stack_pop(Stack s) { return STACK_EMPTY; }
int stack_isempty(Stack s) { return 1; }
Will this work? Of course not. There’s hardly any code! But maybe it will
compile if we run make test:
$ make test
gcc -g3 -Wall -ansi -pedantic -c -o test-stack.o test-stack.c
gcc -g3 -Wall -ansi -pedantic -c -o tester.o tester.c
gcc -g3 -Wall -ansi -pedantic -c -o stack.o stack.c
gcc -g3 -Wall -ansi -pedantic -o test-stack test-stack.o tester.o stack.o
./test-stack
test-stack.c:32: TEST FAILED: stack_isempty(s) -> 1 but expected 0
test-stack.c:33: TEST FAILED: stack_pop(s) -> -1 but expected 3
test-stack.c:34: TEST FAILED: stack_isempty(s) -> 1 but expected 0
test-stack.c:35: TEST FAILED: stack_pop(s) -> -1 but expected 2
test-stack.c:36: TEST FAILED: stack_isempty(s) -> 1 but expected 0
test-stack.c:37: TEST FAILED: stack_pop(s) -> -1 but expected 1

426
test-stack.c:45: TEST FAILED: stack_isempty(s) -> 1 but expected 0
test-stack.c:46: TEST FAILED: stack_pop(s) -> -1 but expected 4
test-stack.c:60: wanted 957 but didn't get it
test-stack.c:74: Aborted (signal 6)
./test-stack: errors 8/17, signals 1, FAILs 1
make[1]: *** [test] Error 8
Hooray! It compiles on the first try! (Well, not really, but let’s pretend it did.)
Unfortunately, it only passes any tests at all by pure dumb luck. But now we
just need to get the code to pass a few more tests.

5.5.4 Bounded-space implementation

Here’s a first attempt at a stack that suffers from some artificial limits. We
retain the structure of the original broken implementation, we just put a few
more lines of code in and format it more expansively.

5.5.4.1 stack.c
#include <stdlib.h>
#include "stack.h"

#define MAX_STACK_SIZE (100)

struct stack {
int top;
int data[MAX_STACK_SIZE];
};

Stack
stack_create(void)
{
struct stack *s;

s = malloc(sizeof(*s));
s->top = 0;
return s;
}

void
stack_destroy(Stack s)
{
free(s);
}

427
void
stack_push(Stack s, int elem)
{
s->data[(s->top)++] = elem;
}

int
stack_pop(Stack s)
{
return s->data[--(s->top)];
}

int
stack_isempty(Stack s)
{
return s->top == 0;
}
Let’s see what happens now:
$ make test
gcc -g3 -Wall -ansi -pedantic -c -o test-stack.o test-stack.c
gcc -g3 -Wall -ansi -pedantic -c -o tester.o tester.c
gcc -g3 -Wall -ansi -pedantic -c -o stack.o stack.c
gcc -g3 -Wall -ansi -pedantic -o test-stack test-stack.o tester.o stack.o
./test-stack
test-stack.c:40: TEST FAILED: stack_isempty(s) -> 0 but expected 1
test-stack.c:41: TEST FAILED: stack_pop(s) -> 409 but expected -1
test-stack.c:47: TEST FAILED: stack_isempty(s) -> 0 but expected 1
test-stack.c:48: TEST FAILED: stack_pop(s) -> 0 but expected -1
test-stack.c:49: TEST FAILED: stack_isempty(s) -> 0 but expected 1
test-stack.c:74: Segmentation fault (signal 11)
test-stack.c:76: TEST FAILED: stack_isempty(s) -> 0 but expected 1
free(): invalid pointer 0x804b830!
./test-stack: errors 6/17, signals 1, FAILs 0
make[1]: *** [test] Error 6
There are still errors, but we get past several initial tests before things blow up.
Looking back at the line numbers in test-stack.c, we see that the first failed
test is the one that checks if the stack is empty after we pop from an empty stack.
The code for stack_isempty looks pretty clean, so what happened? Somewhere
s->top got set to a nonzero value, and the only place this can happen is inside
stack_pop. Aha! There’s no check in stack_pop for an empty stack, so it’s
decrementing s->top past 0. (Exercise: why didn’t the test of stack_pop fail?)

428
5.5.5 First fix

If we’re lucky, fixing this problem will make the later tests happier. Let’s try a
new version of stack_pop. We’ll leave everything else the same.
int
stack_pop(Stack s)
{
if(stack_isempty(s)) {
return STACK_EMPTY;
} else {
return s->data[--(s->top)];
}
}
And now we get:
$ make test
gcc -g3 -Wall -ansi -pedantic -c -o test-stack.o test-stack.c
gcc -g3 -Wall -ansi -pedantic -c -o tester.o tester.c
gcc -g3 -Wall -ansi -pedantic -c -o stack.o stack.c
gcc -g3 -Wall -ansi -pedantic -o test-stack test-stack.o tester.o stack.o
./test-stack
test-stack.c:74: Segmentation fault (signal 11)
test-stack.c:76: TEST FAILED: stack_isempty(s) -> 0 but expected 1
./test-stack: errors 1/17, signals 1, FAILs 0
make[1]: *** [test] Error 1
Which is much nicer. We are still failing the stress test, but that’s not terribly
surprising.

5.5.6 Final version

After some more tinkering, this is what I ended up with. This version uses a
malloc’d data field, and realloc’s it when the stack gets too big.

5.5.6.1 stack.c
#include <stdlib.h>
#include "stack.h"

struct stack {
int top; /* first unused slot in data */
int size; /* number of slots in data */
int *data; /* stack contents */
};

429
#define INITIAL_STACK_SIZE (1)
#define STACK_SIZE_MULTIPLIER (2)

Stack
stack_create(void)
{
struct stack *s;

s = malloc(sizeof(*s));
if(s == 0) return 0;

s->top = 0;
s->size = INITIAL_STACK_SIZE;
s->data = malloc(s->size * sizeof(*(s->data)));
if(s->data == 0) return 0;

/* else everything is ok */
return s;
}

void
stack_destroy(Stack s)
{
free(s->data);
free(s);
}

void
stack_push(Stack s, int elem)
{
if(s->top == s->size) {
/* need more space */
s->size *= STACK_SIZE_MULTIPLIER;
s->data = realloc(s->data, s->size * sizeof(*(s->data)));
if(s->data == 0) {
abort(); /* we have no other way to signal failure :-( */
}
}
/* now there is enough room */
s->data[s->top++] = elem;
}

int
stack_pop(Stack s)
{

430
 if(stack_isempty(s)) {
return STACK_EMPTY;
} else {
return s->data[--(s->top)];
}
}

int
stack_isempty(Stack s)
{
return s->top == 0;
}
At last we have a version that passes all tests:
$ make test
gcc -g3 -Wall -ansi -pedantic -c -o test-stack.o test-stack.c
gcc -g3 -Wall -ansi -pedantic -c -o tester.o tester.c
gcc -g3 -Wall -ansi -pedantic -c -o stack.o stack.c
gcc -g3 -Wall -ansi -pedantic -o test-stack test-stack.o tester.o stack.o
./test-stack
OK!

5.5.7 Moral

Writing a big program all at once is hard. If you can break the problem down
into little problems, it becomes easier. “Test first” is a strategy not just for
getting a well-tested program, but for giving you something easy to do at each
step— it’s usually not too hard to write one more test, and it’s usually not too
hard to get just one test working. If you can keep taking those small, easy steps,
eventually you will run out of failed tests and have a working program.

5.5.8 Appendix: Test macros

/*
* Test macros.
*
* Usage:
*
* #include <setjmp.h>
* #include <stdio.h>
* #include <signal.h>
* #include <unistd.h>
*
* testerInit(); -- Initialize internal data structures.

431
 * testerReport(FILE *, "name"); -- Print report.
* testerResult(); -- Returns # of failed tests.
*
* TRY { code } ENDTRY;
*
* Wraps code to catch seg faults, illegal instructions, etc. May not be
* nested.
* Prints a warning if a signal is caught.
* To enforce a maximum time, set alarm before entering.
*
* TEST(expr, expected_value);
*
* Evaluates expr (which should yield an integer value) inside a TRY.
* Prints a warning if evaluating expr causes a fault or returns a value
* not equal to expected_value.
*
* TEST_ASSERT(expr)
*
* Equivalent to TEST(!(expr), 0)
*
* You can also cause your own failures with FAIL:
*
* TRY {
* x = 1;
* if(x == 2) FAIL("why is x 2?");
* } ENDTRY;
*
* To limit the time taken by a test, call tester_set_time_limit with
* a new limit in seconds, e.g.
*
* tester_set_time_limit(1);
* TRY { while(1); } ENDTRY;
*
* There is an initial default limit of 10 seconds.
* If you don't want any limit, set the limit to 0.
*
*/

/* global data used by macros */

/* nothing in here should be modified directly */
extern struct tester_global_data {
jmp_buf escape_hatch; /* jump here on surprise signals */
int escape_hatch_active; /* true if escape hatch is usable */
int tests; /* number of tests performed */
int errors; /* number of tests failed */
int signals; /* number of signals caught */

432
 int expr_value; /* expression value */
int setjmp_return; /* return value from setjmp */
int try_failed; /* true if last try failed */
int user_fails; /* number of calls to FAIL */
int time_limit; /* time limit for TRY */
} TesterData;

/* set up system; call this before using macros */

void testerInit(void);

/* prints a summary report of all errors to f, prefixed with preamble */

/* If there were no errors, nothing is printed */
void testerReport(FILE *f, const char *preamble);

/* returns number of errors so far. */

int testerResult(void);

/* set a time limit t for TRY, TEST, TEST_ASSERT etc. */

/* After t seconds, an ALARM signal will interrupt the test. */
/* Set t = 0 to have no time limit. */
/* Default time limit is 10 seconds. */
void tester_set_time_limit(int t);

const char *testerStrsignal(int); /* internal hack; don't use this */

/* gruesome non-syntactic macros */

#define TRY \
TesterData.try_failed = 0; \
alarm(TesterData.time_limit); \
if(((TesterData.setjmp_return = setjmp(TesterData.escape_hatch)) == 0) \
&& (TesterData.escape_hatch_active = 1) /* one = is correct*/)
#define ENDTRY else { \
fprintf(stderr, "%s:%d: %s (signal %d)\n", \
__FILE__, __LINE__, \
testerStrsignal(TesterData.setjmp_return), \
TesterData.setjmp_return); \
TesterData.signals++; \
TesterData.try_failed = 1; \
} \
alarm(0); \
TesterData.escape_hatch_active = 0

/* another atrocity */
#define TEST(expr, expected_value) \
TesterData.tests++; \
TesterData.errors++; /* guilty until proven innocent */ \

433
 TRY { TesterData.expr_value = (expr); \
if(TesterData.expr_value != expected_value) { \
fprintf(stderr, "%s:%d: TEST FAILED: %s -> %d but expected %d\n", \
__FILE__, __LINE__, __STRING(expr), \
TesterData.expr_value, expected_value); \
} else { \
TesterData.errors--; \
} \
} \
ENDTRY; \
if(TesterData.try_failed) \
fprintf(stderr, "%s:%d: TEST FAILED: %s caught signal\n", \
__FILE__, __LINE__, __STRING(expr))

#define TEST_ASSERT(expr) TEST((expr) != 0, 1)

#define FAIL(msg) \
(fprintf(stderr, "%s:%d: %s\n", __FILE__, __LINE__, (msg)), \
TesterData.user_fails++, \
TesterData.try_failed = 1)
examples/testHarness/tester.h
#define _GNU_SOURCE /* get strsignal def */

#include <stdio.h>
#include <signal.h>
#include <string.h>
#include <setjmp.h>

#include "tester.h"

struct tester_global_data TesterData;

const char *
testerStrsignal(int sig)
{
return strsignal(sig);
}

static void
tester_sighandler(int signal)
{
if(TesterData.escape_hatch_active) {
TesterData.escape_hatch_active = 0;
longjmp(TesterData.escape_hatch, signal);
}
}

434
void
testerInit(void)
{
TesterData.escape_hatch_active = 0;
TesterData.tests = 0;
TesterData.errors = 0;
TesterData.signals = 0;
TesterData.user_fails = 0;

signal(SIGSEGV, tester_sighandler);
signal(SIGILL, tester_sighandler);
signal(SIGFPE, tester_sighandler);
signal(SIGALRM, tester_sighandler);
signal(SIGBUS, tester_sighandler);
signal(SIGABRT, tester_sighandler);
}

void
testerReport(FILE *f, const char *preamble)
{
if(TesterData.errors != 0 || TesterData.signals != 0) {
fprintf(f, "%s: errors %d/%d, signals %d, FAILs %d\n",
preamble,
TesterData.errors,
TesterData.tests,
TesterData.signals,
TesterData.user_fails);
}
}

int
testerResult(void)
{
return TesterData.errors;
}

void
tester_set_time_limit(int t)
{
TesterData.time_limit = t;
}
examples/testHarness/tester.c

435
5.6 Algorithm design techniques

5.6.1 Basic principles of algorithm design

The fundamental principle of algorithm design was best expressed by the math-
ematician George Polya: “If there is a problem you can’t solve, then there is
an easier problem you can solve: find it.” For computers, the situation is even
better: if there is any technique to make a problem easier even by a tiny bit, then
you can repeat the technique—possibly millions or even billions of times—until
the problem becomes trivial.
For example, suppose we want to find the maximum element of an array of n
ints, but we are as dumb as bricks, so it doesn’t occur to us to iterate through
the array keeping track of the largest value seen so far. We might instead be
able to solve the problem by observing that the maximum element is either (a)
the last element, or (b) the maximum of the first n − 1 elements, depending on
which is bigger. Figuring out (b) is an easier version of the original problem, so
we are pretty much done once we’ve realized we can split the problem in this
way. Here’s the code:
/* returns maximum of the n elements in a */
int
max_element(int a[], int n)
{
int prefix_max;

assert(n > 0);

if(n == 1) {
return a[0];
} else {
prefix_max = max_element(a, n-1);
if(prefix_max < a[n-1]) {
return a[n-1];
} else {
return prefix_max;
}
}
}
Note that we need a special case for a 1-element array, because the empty prefix
of such an array has no maximum element. We also assert that the array
contains at least one element, just to avoid mischief.
One problem with this algorithm (at least when coding in C) is that the recursion
may get very deep. Fortunately, there is a straightforward way to convert the
recursion to a loop. The idea is that instead of returning a value from the

436
recursive call, we put it in a variable that gets used in the next pass through the
loop. The result is
/* returns maximum of the n elements in a */
int
max_element(int a[], int n)
{
int i; /* this replaces n-1 from the recursive version */
int prefix_max;

assert(n > 0);

prefix_max = a[0]; /* this is the i == 0 case */

for(i = 1; i < n; i++) {

if(prefix_max < a[i]) {
prefix_max = a[i]; /* was return a[n-1] */
}
/* else case becomes prefix_max = prefix_max, a noop */
}

/* at the end we have to return a value for real */

return prefix_max;
}

5.6.2 Specific techniques

Algorithm design often requires both creativity and problem-specific knowledge,

but there are certain common techniques that appear over and over again. The
following classification is adapted from Anany Levitin, Introduction to the Design
& Analysis of Algorithms, Addison-Wesley, 2003.
Brute force Try all possible solutions until you find the right one.
Divide and conquer Split the problem into two or more subproblems, solve
the subproblems recursively, and then combine the solutions.
Decrease and conquer Reduce the problem to a single smaller problem, solve
that problem recursively, and then use that solution to solve the original
problem.
Transform and conquer Either (a) transform the input to a form that makes
the problem easy to solve, or (b) transform the input into the input to
another problem whose solution solves the original problem.
Use space Solve the problem using some auxiliary data structure.
Dynamic programming Construct a table of solutions for increasingly large
subproblems, where each new entry in the table is computed using previous
entries in the table.

437
Greedy method Run through your problem one step at a time, keeping track
of the single best solution at each step. Hope sincerely that this will not
lead you to make a seemingly-good choice early with bad consequences
later.
Some of these approaches work better than others—it is the role of algorithm
analysis (and experiments with real computers) to figure out which are likely to
be both correct and efficient in practice. But having all of them in your toolbox
lets you try different possibilities for a given problem.

5.6.3 Example: Finding the maximum

Though this classification is not completely well-defined, and is a bit arbitrary

for some algorithms, it does provide a useful list of things to try in solving a
problem. Here are some examples of applying the different approaches to a
simple problem, the problem of finding the maximum of an array of integers.
Brute force For index i, test if A[i] is greater than or equal to every element
in the array. When you find such an A[i], return it. For this algorithm,
T (n) = n · Θ(n) = Θ(n2 ) if implemented in the most natural way.
Divide and conquer If A has only one element, return it. Otherwise, let
m1 be the maximum of A[1] . . . A[n/2], and let m2 be the maximum of
A[n/2 + 1] . . . A[n]. Return the larger of m1 and m2 . The running time is
given by $T(n) = 2T (n/2) + Θ(1) = Θ(n).
Decrease and conquer If A has only one element, return it. Otherwise, let
m* be the maximum of A[2] . . . A[n]. Return the larger ofPA[0] and m. Now
n
the running time is given by T (n) = T (n − 1) + Θ(1) = i=1 Θ(1) = Θ(n).
Transform and conquer Sort the array, then return A[n]. Using an optimal
comparison-based sort, this takes Θ(n log n) + Θ(1) = Θ(n log n) time. The
advantage of this approach is that you probably don’t have to code up the
sorting routine yourself, since most libraries include sorting.
Use space Insert all elements into a balanced binary search tree, then return
the rightmost element. Cost is Θ(n log n) to do n insertions, plus Θ(log n)
to find the rightmost element, for a total of Θ(n log n). Sorting is equivalent
and probably easier.
Dynamic programming Create an auxiliary array B with indices 1 to n. Set
B[1] = A[1]. As i goes from 2 to n, set B[i] to the larger of B[i − 1] and
A[i], so that B[i] is always the maximum among A[1] . . . A[i]. Return B[n].
Cost: Θ(n). As is often the case, one can reduce the space to O(1) by
throwing away parts of B that we aren’t going to look at again.
Greedy method Let m = A[1]. For each element A[i] in A[2 . . . n], if A[i] > m,
set m to A[i]. Return the final value of m. Cost: Θ(n). This algorithm is
pretty much identical to the previous one.

438
5.6.4 Example: Sorting

The sorting problem asks, given as input an array A of n elements in arbitrary

order, to produce as output an array containing the same n elements in non-
decreasing order, i.e. with A[i] ≤ A[i + 1] for all i. We can apply each of the
techniques above to this problem and get a sorting algorithm (though some are
not very good).
Brute force For each of the n! permutations of the input, test if it is sorted by
checking A[i] ≤ A[i + 1] for all i. Cost if implemented naively: n! · Θ(n) =
Θ((n + 1)!). This algorithm is known as deterministic monkeysort or
deterministic bogosort. It also has a randomized variant, where the
careful generation of all n! permutations is replaced by shuffling. The
randomized variant is easier to code and runs at about the same speed
as the deterministic variant, but does not guarantee termination if the
shuffling is consistently unlucky.
Divide and conquer Sort A[1 . . . bn/2c and A[bn/2 + 1c . . . n] separately, then
merge the results (which takes Θ(n) time and Θ(n) additional space if
implemented in the most straightforward way). Cost: T (n) = 2T (n/2) +
Θ(n) = Θ(n log n) by the Master Theorem. This method gives mergesort,
one of the fastest general-purpose sorting algorithms. The merge can be
avoided by carefully splitting the array into elements less than and elements
greater than some pivot, then sorting the two resulting piles; this gives
quicksort. The performance of quicksort is often faster than mergesort in
practice, but its worst-case performance (when the pivot is chosen badly)
is just as bad as the result of insertion sort, which we will look at next.
Decrease and conquer Remove A[n], sort the remainder, then insert A[n] in
the appropriate place. This algorithm is called insertion sort. The final
insertion step requires finding the right place (which can be done fairly
quickly if one is clever) but then moving up to n − 1 elements to make
room for A[n]. Total cost is given by T (n) = T (n − 1) + Θ(n) = T (n2 ).
Transform and conquer I’m not aware of any good general transform-and-
conquer approach to sorting (there are some bad ones), but in some cases
one can transform seemingly general sorting problem (e.g. sorting strings)
into specialized sorting problems that permit faster solutions (e.g. sorting
small integers).
Use space Insert the elements into a balanced binary search tree, then read
them out from left to right. Another version: insert them into a heap.
Both take Θ(n log n) time, but are more complicated to implement than
mergesort or quicksort unless you have binary search tree or heap code
lying around already.
Dynamic programming Insertion sort may be seen as an example of this.
Greedy method Find the smallest element, mark it as used, and output it.
Repeat until no elements are left. The result is selection sort), which runs
in a respectable but suboptimal Θ(n2 ) time.

439
5.7 Bit manipulation

Sometimes it is convenient to consider a block of chars as really being a block

of bits. This requires using C’s bit operators to get at individual bits.
Here are some simple macros for extracting a particular bit from a char array,
thought of as a large vector of bits. These assume that the bytes are stored in
little-endian order, which means that the least significant bytes come first (see
Endianness). This may produce odd results if you feed them a char * that has
been converted from a larger integer type.
#define BITS_PER_BYTE (8)

/* extract the n-th bit of x */

#define GET_BIT(x, n) ((((x)[(n) / BITS_PER_BYTE]) & (0x1 << ((n) % BITS_PER_BYTE))) != 0)

/* set the n-th bit of x to 1 */

#define SET_BIT(x, n) ((x)[(n) / BITS_PER_BYTE]) |= (0x1 << ((n) % BITS_PER_BYTE))

/* set the n-th bit of x to 0 */

#define RESET_BIT(x, n) ((x)[(n) / BITS_PER_BYTE]) &= ~(0x1 << ((n) % BITS_PER_BYTE))
If you want to get multiple bits, use the right-shift operator to shift them over
to the right end of the word and then mask with bitwise AND. For example:
#define BITS_PER_BYTE (8)

/* this rather nasty expression constructs an all-ones byte */

#define BYTE_MASK ((1 << BITS_PER_BYTE) - 1)

/* extract the n-th byte from a word */

#define GET_BYTE(x, n) (((x) >> BITS_PER_BYTE * (n)) & BYTE_MASK)

/* extract n bits starting at position i from x */

#define GET_BITS(x, i, j) (((x) >> (i)) & ((1 << n) - 1))

/* another definition of GET_BIT */

#define GET_BIT2(x, n) GET_BITS(x, n, 1)
Many much more sophisticated techniques for doing bit-fiddling can be found at
http://www.jjj.de/bitwizardry/bitwizardrypage.html.

5.8 Persistence

When a C program exits, all of its global variables, local variables, and heap-
allocated blocks are lost. Its memory is reclaimed by the operating system,

440
erased, and handed out to other programs. So what happens if you want to keep
data around for later?
To make this problem concrete, let’s suppose we want to keep track of a hit
counter for web pages. From time to time, the user will run the command
count_hit number where number is an integer value in the range 0 to 99, say.
(A real application would probably be using urls, but let’s keep things as simple
as possible.) We want count_hit to print the number of times the page with the
given number has been hit, i.e. 1 the first time it is called, 2 the next time, etc.
Where can we store the counts so that they will survive to the next execution of
count_hit?

5.8.1 A simple solution using text files

The simplest solution is probably to store the data in a text file. Here’s a
program that reads a file hits, increments the appropriate value, and the writes
out a new version. To reduce the chances that data is lost (say if count_hit
blows up halfway through writing the file), the new values are written to a new
file hit~, which is then renamed to hit, taking the place of the previous version.
#include <stdio.h>
#include <stdlib.h>

#define NUM_COUNTERS (100) /* number of counters we keep track of */

#define COUNTER_FILE "/tmp/hit" /* where they are stored */
#define NEW_COUNTER_FILE COUNTER_FILE "~" /* note use of constant string concatenation */

int
main(int argc, char **argv)
{
int c;
int i;
int counts[NUM_COUNTERS];
FILE *f;

if(argc < 2) {
fprintf(stderr, "Usage: %s number\n", argv[0]);
exit(1);
}
/* else */

c = atoi(argv[1]);
if(c < 0 || c >= NUM_COUNTERS) {
fprintf(stderr, "Counter %d not in range 0..%d\n", c, NUM_COUNTERS - 1);
exit(2);
}

441
 f = fopen(COUNTER_FILE, "r");
if(f == 0) {
perror(COUNTER_FILE);
exit(3);
}

/* read them in */
for(i = 0; i < NUM_COUNTERS; i++) {
fscanf(f, "%d", &counts[i]);
}
fclose(f);

printf("%d\n", ++counts[c]);

/* write them back */

f = fopen(NEW_COUNTER_FILE, "w");
for(i = 0; i < NUM_COUNTERS; i++) {
fprintf(f, "%d\n", counts[i]);
}
fclose(f);

rename(NEW_COUNTER_FILE, COUNTER_FILE);

return 0;
}
examples/persistence/textFile.c
If you want to use this, you will need to create an initial file /tmp/hit with
NUM_COUNTERS zeroes in it.
Using a simple text file like this is the easiest way to keep data around, since
you can look at the file with a text editor or other tools if you want to do things
to it. But it means that the program has to parse the file every time it runs. We
can speed things up a little bit (and simplify the code) by storing the values in
binary.

5.8.2 Using a binary file

Here’s a version that stores the data as a binary file of exactly sizeof(int) * NUM_COUNTERS
bytes. It uses the stdio routines fread and fwrite to read and write the file.
These are much faster than the loops in the previous program, since they can
just slap the bytes directly into counts without processing them at all.
The program also supplies and extra flag b to fopen. This is ignored on Unix-like
machines but is needed on Windows machines to tell the operating system that

442
the file contains binary data (such files are stored differently from text files on
Windows).
#include <stdio.h>
#include <stdlib.h>

#define NUM_COUNTERS (100) /* number of counters we keep track of */

#define COUNTER_FILE "/tmp/hit" /* where they are stored */
#define NEW_COUNTER_FILE COUNTER_FILE "~" /* note use of constant string concatenation */

int
main(int argc, char **argv)
{
int c;
int counts[NUM_COUNTERS];
FILE *f;

if(argc < 2) {
fprintf(stderr, "Usage: %s number\n", argv[0]);
exit(1);
}
/* else */

c = atoi(argv[1]);
if(c < 0 || c >= NUM_COUNTERS) {
fprintf(stderr, "Counter %d not in range 0..%d\n", c, NUM_COUNTERS - 1);
exit(2);
}

f = fopen(COUNTER_FILE, "rb");
if(f == 0) {
perror(COUNTER_FILE);
exit(3);
}

/* read them in */
fread(counts, sizeof(*counts), NUM_COUNTERS, f);
fclose(f);

printf("%d\n", ++counts[c]);

/* write them back */

f = fopen(NEW_COUNTER_FILE, "wb");
fwrite(counts, sizeof(*counts), NUM_COUNTERS, f);
fclose(f);

443
 rename(NEW_COUNTER_FILE, COUNTER_FILE);

return 0;
}
examples/persistence/binaryFile.c
Again, you’ll have to initialize /tmp/hit to use this; in this case, you want it to
contain exactly 400 null characters. On a Linux machine you can do this with
the command dd if=/dev/zero of=/tmp/hit bs=400 count=1.
The advantage of using binary files is that reading and writing them is both
simpler and faster. The disadvantages are (a) you can’t look at or update the
binary data with your favorite text editor any more, and (b) the file may no
longer be portable from one machine to another, if the different machines have
different endianness or different values of sizeof(int). The second problem we
can deal with by converting the data to a standard word size and byte order
before storing it, but then we lose some advantages of speed.

5.8.3 A version that updates the file in place

We still may run into speed problems if NUM_COUNTERS is huge. The next program
avoids rewriting the entire file just to update one value inside it. This program
uses the fseek function to position the cursor inside the file. It opens the file
using the "r+b" flag to fopen, which means to open an existing binary file for
reading and writing.
#include <stdio.h>
#include <stdlib.h>

#define NUM_COUNTERS (100) /* number of counters we keep track of */

#define COUNTER_FILE "/tmp/hit" /* where they are stored */

int
main(int argc, char **argv)
{
int c;
int count;
FILE *f;

if(argc < 2) {
fprintf(stderr, "Usage: %s number\n", argv[0]);
exit(1);
}
/* else */

c = atoi(argv[1]);

444
 if(c < 0 || c >= NUM_COUNTERS) {
fprintf(stderr, "Counter %d not in range 0..%d\n", c, NUM_COUNTERS - 1);
exit(2);
}

f = fopen(COUNTER_FILE, "r+b");
if(f == 0) {
perror(COUNTER_FILE);
exit(3);
}

/* read counter */
fseek(f, sizeof(int) * c, SEEK_SET);
fread(&count, sizeof(int), 1, f);

printf("%d\n", ++count);

/* write it back */
fseek(f, sizeof(int) * c, SEEK_SET);
fwrite(&count, sizeof(int), 1, f);
fclose(f);

return 0;
}
examples/persistence/binaryFileFseek.c
Note that this program is not only shorter than the last one, but it also avoids
allocating the counts array. It also is less likely to run into trouble with running
out of space during writing. If we ignore issues of concurrency, this is the best
we can probably do with just stdio.

5.8.4 An even better version using mmap

We can do even better using the mmap routine, available in all POSIX-compliant
C libraries. POSIX, which is short for Portable Standard Unix, is supported by
essentially all Unix-like operating systems and NT-based versions of Microsoft
Windows. The mmap routine tells the operating system to “map” a file in the
filesystem to a region in the process’s address space. Reading bytes from this
region will read from the file; writing bytes to this region will write to the file
(although perhaps not immediately). Even better, if more than one process
calls mmap on the same file at once, they will share the memory region, so that
updates made by one process will be seen immediately by the others (with some
caveats having to do with how concurrent access to memory actually works on
real machines).

445
Here is the program using mmap:
#include <stdio.h>
#include <stdlib.h>
#include <sys/types.h>
#include <sys/stat.h>
#include <fcntl.h>
#include <sys/mman.h> /* For mmap. I think mman is short for "memory management." */

#define NUM_COUNTERS (100) /* number of counters we keep track of */

#define COUNTER_FILE "/tmp/hit" /* where they are stored */
#define NEW_COUNTER_FILE COUNTER_FILE "~" /* note use of constant string concatenation */

int
main(int argc, char **argv)
{
int c;
int *counts;
int fd;

if(argc < 2) {
fprintf(stderr, "Usage: %s number\n", argv[0]);
exit(1);
}
/* else */

c = atoi(argv[1]);
if(c < 0 || c >= NUM_COUNTERS) {
fprintf(stderr, "Counter %d not in range 0..%d\n", c, NUM_COUNTERS - 1);
exit(2);
}

/* open and map the file */

fd = open(COUNTER_FILE, O_RDWR);
if(fd < 0) {
perror(COUNTER_FILE);
exit(3);
}
counts = mmap(0, sizeof(*counts) * NUM_COUNTERS, PROT_READ|PROT_WRITE, MAP_SHARED, fd, 0

if(counts == 0) {
perror(COUNTER_FILE);
exit(4);
}

printf("%d\n", ++counts[c]);

446
 /* unmap the region and close the file just to be safe */
munmap(counts, sizeof(*counts) * NUM_COUNTERS);
close(fd);

return 0;
}
examples/persistence/binaryFileMmap.c
Now the code for actually incrementing counts[c] and writing it to the file
is trivial. Unfortunately, we have left stdio behind, and have to deal with
low-level POSIX calls like open and close to get at the file. Still, this may be
the most efficient version we can do, and becomes even better if we plan to do
many updates to the same file, since we can just keep the file open.

5.8.5 Concurrency and fault-tolerance issues: ACIDity

All of the solutions described so far can fail if you run two copies of count_hits
simultaneously. The mmap solution is probably the least vulnerable to failures,
as the worst that can happen is that some update is lost if the same locations is
updated at exactly the same time. The other solutions can fail more spectacularly;
simultaneous writes to /tmp/hit~ in the simple text file version, for example,
can produce a wide variety of forms of file corruption. For a simple web page hit
counter, this may not be a problem. If you are writing a back-end for a bank,
you probably want something less vulnerable.
Database writers aim for a property called ACIDity from the acronym ACID
= Atomicity, Consistency, Isolation, and Durability. These are defined
for a system in which the database is accessed via transactions consisting of
one or more operations. An example of a transaction might be ++counts[c],
which we can think of as consisting of two operations: reading counts[c], and
writing back counts[c]+1.
Atomicity means that either every operation in a transaction is performed or
none is. In practice, this means if the transaction fails any partial progress must
be undone.
Consistency means that at the end of a transaction the database is in a “consistent”
state. This may just mean that no data has been corrupted (e.g. in the text
data file we have exactly 100 lines and they’re all integer counts), or it may also
extend to integrity constraints enforce by the database (e.g. in a database of
airline flights, the fact that flight 2937 lands at HVN at 22:34 on 12/17 implies
that flight 2937 exists, has an assigned pilot, etc.).
Isolation says that two concurrent transactions can’t detect each other; the
partial progress of one transaction is not visible to others until the transaction
commits.

447
Durability means that the results of any committed transaction are permanent.
In practice this means there is enough information physically written to a disk
to reconstruct the transaction before the transaction finished.
How can we enforce these requirements for our hit counter? Atomicity is not
hard: if I stop a transaction after a read but before the write, no one will be the
wiser (although there is a possible problem if only half of my write succeeds).
Consistency is enforced by the fseek and mmap solutions, since they can’t change
the structure of the file. Isolation is not provided by any of our solutions,
and would require some sort of locking (e.g. using flock) to make sure that
only one program uses the file at a time. Durability is enforced by not having
count_hits return until the fclose or close operation has succeeded (although
full durability would require running fsync or msync to actually guarantee data
was written to disk).
Though it would be possible to provide full ACIDity with enough work, this is
a situation where using an existing well-debugged tool beats writing our own.
Depending on what we are allowed to do to the machine our program is running
on, we have many options for getting much better handling of concurrency. Some
standard tools we could use are:
• gdbm. This is a minimal hash-table-on-disk library that uses simplistic
locking to get isolation. The advantage of this system is that it’s probably
already installed on any Linux machine. The disadvantage is that it doesn’t
provide much functionality beyond basic transactions.
• Berkeley DB is a fancier hash-table-on-disk library that provides full
ACIDity but not much else. There is a good chance that some version of
this is also installed by default on any Linux or BSD machine you run into.
• Various toy databases like SQLite or MySQL provide tools that look very
much like serious databases with easy installation and little overhead.
These are probably the solutions most people choose, especially since
MySQL is integrated tightly with PHP and other Web-based scription
languages. Such a solution also allows other programs to access the table
without having to know a lot of details about how it is stored, because the
SQL query language hides the underlying storage format.
• Production-quality databases like PostgreSQL, SQL Server, or Oracle
provide very high levels of robustness and concurrency at the cost of
requiring non-trivial management and possibly large licensing fees. This is
what you pick if you really are running a bank.

6 What next?
Congratulations! You now know everything there is to know about programming
in C. Now what do you do?
My recommendation would be the following: learn C++, since you know 75% of

448
it already, and you will be able to escape from some (but not all) of the annoying
limitations of C. And learn a scripting language you can be comfortable with,
for writing programs quickly where performance isn’t the main requirement.

6.1 What’s wrong with C

In this class, we had to work around fundamental limitations in C on several

occasions.
C doesn’t have a garbage collector Many modern program languages will
detect and free unreachable data for you automatically. C doesn’t, so the
programmer has to spend a lot of time worrying about when and by whom
data allocated with malloc will be passed to free. Not only does this
create many possibilities for error, but it also means that certain kinds
of data structures in which a single component of the data structure is
pointed to by an unpredictable number of other components are difficult
to write in C, since it’s hard to tell when it is safe to free a component.
Garbage-collected languages avoid all of these problems at a slight cost
in performance. Though there exists a garbage collector for C/C++
http://www.hboehm.info/gc/, it isn’t 100% portable and may not work as
well as a built-in collector.
C doesn’t support any kind of polymorphism Polymorphism is when a
function can work on more than one data type. The closest C can do
is either parameterized macros (see MacrosMacros.html)), heavy use of
void * and function pointers as in qsort, or various nasty hacks where
code is automatically generated with type names filled in from a base
template (see MacrosMacros.html) again). Most modern programming
languages have some sort of support for polymorphism, allowing you to
write, for example, a generic sorting routine without resorting to void *-
like departures from the type system.
C doesn’t have exceptions Exceptions are a mechanism for doing non-
standard returns from a function when something blows up, which get
caught using an “exception handler” that is often separate from the code
handling the normal return values and which can often be used to catch
exceptions from a variety of sources. Instead, C requires function writers
to invent and document an ad-hoc protocol for indicating bad outcomes for
every function they write, and requires function users to remember to test
for bad return values. Most programmers are too lazy to do this all the
time, leading to undetected run-time errors. Most modern programming
languages fix this.
C doesn’t support object-oriented programming very well “Object-
oriented” is a buzzword with many possible meanings (but see
http://c2.com/cgi/wiki?HeInventedTheTerm). However, at minimum it
means that in addition to supporting polymorphism (described above),
your language should support strong encapsulation (controlling who

449
 can get at the internal data of an object) and inheritance (allowing one
abstract data type to be defined by extending another). You can fake
most of these things in C if you try hard enough (for example, using
function pointers), but it is always possible to muck around with internal
bits of things just because of the unlimited control C gives you over
the environment. This can quickly become dangerous in large software
projects.
C provides only limited support for avoiding namespace collisions
In a large C program, it’s impossible to guarantee that my
eat_leftovers function exported from leftovers.c doesn’t con-
flict with your eat_leftovers function in cannibalism.c. A mediocre
solution is to use longer names: leftovers_eat_leftovers vs
cannibalism_eat_leftovers, and one can also play games with
function pointers and global struct variables to allow something like
leftovers.eat_leftovers vs cannibalism.eat_leftovers. Most
modern programming languages provide an explicit package or namespace
mechanism to allow the programmer to control who sees what names
where.

6.2 What C++ fixes

On the above list, C++ fixes everything except the missing garbage collector.
If you want to learn C++, you should get a copy of The C++ Programming
Language, by Bjarne Stroustrup, which is the definitive reference manual. But
you can get a taste of it from several on-line tutorials:
• C++ tutorial for C users, by Eric Brasseur. Exactly what it says. Intro-
duces C++ features not found in C in order of increasing complexity.
• Some other on-line tutorials that assume little or no prior programming
experience:
– http://www.cplusplus.com/doc/tutorial/
– http://www.cprogramming.com/tutorial.html

6.3 Other C-like languages

C syntax has become the default for new programming languages targeted at a
general audience. Some noteworthy examples of C-like languages are Java (used
in Android), Objective-C (used in OSX and iOS), and C# (used in Windows).
Each of these fix some of the misfeatures of C (including the lack of a garbage
collector and bounds checks on arrays) while retaining much of the flavor of
C. Which to choose probably depends on what platform you are interested in
developing for.

450
6.4 Scripting languages

Much current programming is done in so-called scripting languages like

Python, Perl, PHP, JavaScript, Visual Basic, Tcl, etc. These are generally
interpreted languages similar to Lisp or Scheme under the hood, with dynamic
typing (type information is carried along with values, so type errors are detected
only at runtime but polymorphism is provided automatically), garbage collec-
tors, and support for many advanced programming features like objects and
anonymous functions. What distinguishes scripting languages from the Lisp-like
languages is that the syntax is generally more accessible to newcomers and
the language runtime usually comes with very large libraries providing built-in
tools for doing practical programming tasks like parsing odd input formats
and interfacing to databases and network services. The result is that common
programming tasks can be implemented using very few lines of code, at a cost
in performance that ranges from slight to horrendous depending on what you
are doing.
Let’s look at an example in two common scripting languages, Perl and Python.
Here are some solutions to an old assignment, which find all the palindromes on
stdin and report the first non-matching character for any non-palindrome.
The original C version looks like this:
#include <stdio.h>
#include <stdlib.h>
#include <string.h>

/* Palindrome detector.
*
* For each line of the input, prints PALINDROME if it is a palindrome
* or the index of the first non-matching character otherwise.
*
* Note: does not handle lines containing nulls.
*/

/* read a line of text from stdin

* and return it (without terminating newline) as a freshly-malloc'd block.
* Caller is responsible for freeing this block.
* Returns 0 on error or EOF.
*/
char *
getLine(void)
{
char *line; /* line buffer */
int n; /* characters read */
int size; /* size of line buffer */

451
 int c;

size = 1;
line = malloc(size);
if(line == 0) return 0;

n = 0;

while((c = getchar()) != '\n' && c != EOF) {

while(n >= size - 1) {
size *= 2;
line = realloc(line, size);
if(line == 0) return 0;
}
line[n++] = c;
}

if(c == EOF && n == 0) {

/* got nothing */
free(line);
return 0;
} else {
line[n++] = '\0';
return line;
}
}

#define IS_PALINDROME (-1)

/* returns IS_PALINDROME if s is a palindrome,

* or index of first unmatched character otherwise. */
int
testPalindrome(const char *s)
{
int n; /* length of s */
int i;

n = strlen(s);

/* we only have to check up to floor(n/2) */

for(i = 0; i < n/2; i++) {
if(s[i] != s[n-1-i]) {
return i;
}
}
/* else */

452
 return IS_PALINDROME;
}

int
main(int argc, char **argv)
{
char *line;
int mismatch;

while((line = getLine()) != 0) {
mismatch = testPalindrome(line);
if(mismatch == IS_PALINDROME) {
puts("PALINDROME");
} else {
printf("%d\n", mismatch);
}

free(line);
}

return 0;
}

examples/scripting/palindrome.c
This version is written in Perl (http://www.perl.org):
#!/usr/bin/perl

# For each line in stdin, print PALINDROME if it is a palindrome, or index of

# the first non-matching character otherwise.

while(<>) {
chomp; # remove trailing newline
if($_ eq reverse $_) {
print "PALINDROME\n";
} else {
for $i (0..length($_) - 1) {
if(substr($_, $i, 1) ne substr($_, length($_) - $i - 1, 1)) {
print $i, "\n";
last;
}
}
}
}

453
examples/scripting/palindrome.pl
The things to notice about Perl is that the syntax is deliberately very close to
C (with some idiosyncratic extensions like putting $ on the front of all variable
names), and that common tasks like reading all input lines get hidden inside
default constructions like while(<>) and the $_ variable that functions with no
arguments like chomp operate on by default. This can allow for very compact
but sometimes very incomprehensible code.
Here’s a version in Python (https://www.python.org/):
#!/usr/bin/python

"""For each line in stdin, print PALINDROME if it is a palindrome, or index of

the first non-matching character otherwise."""

import sys

for line in sys.stdin:

line = line.rstrip('\n') # remove trailing newline
if line == line[::-1]:
print("PALINDROME")
else:
mismatches = [ i for i in range(len(line)) if line[i] != line[-(i+1)] ]
print(min(mismatches))
examples/scripting/palindrome.py
Here the syntax is a little more alien if you are used to C: Python doesn’t use
curly braces for block structure, using indentation instead. The code above
uses some other odd features of the language, such as the ability to take “slices”
of sequence variables like strings (the expression line[::-1] means “take all
elements of line starting from the obvious default starting point (the empty
string before the first colon) to the obvious default ending point (the empty
string before the second colon) stepping backwards one character at a time
(the -1)), a feature the language adopted from array-processing languages like
MatLab; and the ability to do list comprehensions (the large expression assigned
to mismatches), a feature that Python adopted from Haskell and that Haskell
adopted from set theory.
What these gain in short code length they lose in speed; run times on
/usr/share/dict/words in the Zoo are

C 0.107s
Perl 0.580s
Python 2.052s

Note that for Perl and Python some of the cost is the time to start the interpreter

454
and parse the script, but factors of 10–100 are not unusual slowdowns when
moving from C to a scripting language. The selling point of these languages
is that in many applications run time is not as critical as ease and speed of
implementation.
As an even shorter example, if you just want to print all the palindromes in a
file, you can do that from the command line in one line of Perl, e.g:
$ perl -ne 'chomp; print $_, "\n" if($_ eq reverse $_)' < /usr/share/dict/words

7 Assignments

7.1 Assignment 1, due Thursday 2015-01-29, at 11:00pm

7.1.1 Bureaucratic part

Make sure that you sign up for an account on the Zoo at http://zoo.cs.yale.edu/
accounts.html. If you already have an account, you still need to check the CPSC
223 box so that you can turn in assignments. It’s best to do this as soon as
possible.
You do not need to develop your solution on the Zoo, but you will need to turn
it in there, and it will be tested using the compiler on the Zoo.

7.1.2 A rotten cipher

For this assignment, you are to implement an encoder for a polyalphabetic

substitution cipher vaguely inspired by the Enigma machines used by Germany
during World War 2. Unlike the Enigma machine, this cipher doesn’t provide
much security, so you can probably tell your non-US-national friends about it
without violating US export control laws.26
Each letter 'A' through 'Z' or 'a' through 'z' is encrypted by shifting it some
number of positions forward in the alphabet, wrapping around at the end.
The number of positions is determined by an offset that changes over time. The
initial shift is 17 positions. After encoding an uppercase letter, the shift is
increased by 5. After encoding a lowercase letter, the shift is increased by 3. To
avoid overflow on long texts, it’s probably a good idea to store the offset modulo
26.
An uppercase letter is always encoded by an uppercase letter, and a lowercase
letter is always encoded by a lowercase letter. All other characters are passed
through intact.
26 Not intended as legal advice.

455
Below is an example of encoding a famous Napoleonic palindrome using this
cipher:
Plaintext: A b l e w a s I e r e I s a w E l b a
Offset: 17 22 25 2 5 8 11 14 19 22 25 2 7 10 13 16 21 24 1
Ciphertext: R x k g b i d W x n d K z k j U g z b

7.1.3 Your task

For this assignment, you are to write a program encode.c that takes a plaintext
from stdin, encodes it using the above algorithm, and writes the result to
stdout.
For example, given the input
"Stop, thief!" cried Tom, arrestingly.

"You'll never take me alive!" replied the criminal, funereally.

Your program should output
"Jpnr, yptsw!" woihj Ccd, uorhycucygw.

"Zud'xa fztfv akxu fa znndp!" fvjiihj ctt umgnmuky, vnjdtjiwzp.

7.1.4 Hints

• You should assume that you are using the standard Latin 26-letter alphabet.
• You may assume that the characters 'A' through 'Z' and 'a' through 'z'
are represented using continuous ranges of integers, so that the expression
c - 'A' gives the position of c in the alphabet, provided c is an uppercase
character, and counting A as 0. This means that your program will not
be portable to machines that use EBCDIC or some other exotic character
representation.
• To test if a character is uppercase or lowercase, one option would be to put
#include <ctype.h> in your program and use the isupper and islower
macros. Note that these may behave oddly if you have set a locale that
uses a different alphabet. It may be safer to make your own tests.

7.1.5 Testing your assignment

Sample inputs and outputs can be found in /c/cs223/Hwk1/testFiles on the

Zoo. Note that some of these files contain non-printing characters that may
have odd effects if you send them to your screen. The safest way to test if your
program produces the same output as the sample output is probably to use cmp,
for example:

456
$ ./encode < test.in > tmp
$ cmp tmp test.out
If tmp and test.out contain the same characters, cmp will say nothing. Other-
wise it will tell you the first position where they differ.
If you want to see what characters are in a binary file, trying using od -t x1z,
as in
$ echo hi > file
$ cat file
hi
$ od -t x1z file
0000000 68 69 0a >hi.<
0000003

7.1.6 Submitting your assignment

Submit your assignment using the command:

/c/cs223/bin/submit 1 encode.c
You can test that your program compiles (and passes a few basic tests) using
the command:
/c/cs223/bin/testit 1 encode
This runs the test script in /c/cs223/Hwk1/test.encode on your submitted
assignment. You can also run this script by hand to test the version of encode.c
in your current working directory.
The unsympathetic robo-grading script used to grade this assignment may or
may not use the same tests as this command, so you should make sure your
program works on other inputs as well. You may also want to look at the style
grading checklist to see that you haven’t committed any gross atrocities against
readability, in case a human being should happen to look at your code.
You can submit your assignment more than once, but any late penalties will be
assessed based on the last submission. For more details about the submit script
and its capabilities, see here.

7.1.7 Sample solution

/*
* Encode text on stdin by alphabet rotation with shifting offset.
*
* Initially, each character 'A'..'Z' or 'a'..'z' is rotated 17 positions.
*
* After encoding an uppercase letter, the offset is increased by 5 (mod 26).

457
 *
* After encoding a lowercase letter, the offset is increased by 3 (mod 26).
*
* These parameters are set using the INITIAL_OFFSET, UPPERCASE_STEP, and LOWERCASE_STEP
* constants defined below.
*
*/
#include <stdio.h>

#define INITIAL_OFFSET (17)

#define UPPERCASE_STEP (5)

#define LOWERCASE_STEP (3)

#define MODULUS ('z' - 'a' + 1)

int
main(int argc, char **argv)
{
int offset = INITIAL_OFFSET;
int c;

while((c = getchar()) != EOF) {

if(('a' <= c) && (c <= 'z')) {
putchar(((c - 'a') + offset) % MODULUS + 'a');
offset = (offset + LOWERCASE_STEP) % MODULUS;
} else if(('A' <= c) && (c <= 'Z')) {
putchar(((c - 'A') + offset) % MODULUS + 'A');
offset = (offset + UPPERCASE_STEP) % MODULUS;
} else {
putchar(c);
}
}

return 0;
}
examples/2015/hw/1/encode.c

7.2 Assignment 2, due Wednesday 2015-02-04, at 11:00pm

7.2.1 Opening a safe

The Hollywood Hackable Safe Company has announced a new line of

electronically-controlled safes with a C API to permit maximum opening

458
speed while still protecting your valuable loot. This interface is defined
in the following file, which can also be found on the Zoo in the directory
/c/cs223/Hwk2/sourceFiles:
/*
* API for safes made by the
* Hollywood Hackable Safe Company LLC.
*/

typedef struct safe Safe; /* opaque data type for a safe */

/*
* Returns the number of tumblers on a safe.
* If this is n, the possible tumbler indices will be 0 through n-1.
* */
int numTumblers(Safe *s);

/*
* Returns the number of positions of each tumbler.
* If this is n, the possible tumbler positions will be 0 through n-1.
*/
int numPositions(Safe *s);

/* Return codes for tryCombination */

#define SAFE_BAD_COMBINATION (-1)
#define SAFE_SELF_DESTRUCTED (-2)

/*
* Try a combination.
*
* This should be an array of numTumbler(s) ints.
*
* Returns contents of safe (a non-negative int) if combination is correct
* and safe has not yet self-destructed.
*
* Returns SAFE_BAD_COMBINATION if combination is incorrect
* and safe has not yet self-destructed.
*
* Returns SAFE_SELF_DESTRUCTED if safe has self-destructed.
*
* Note: may modify combination.
*/
int tryCombination(Safe *s, int *combination);
examples/2015/hw/2/safe.h
The noteworthy function in this API is tryCombination, which takes a pointer

459
to a safe and an array of ints representing the combination, and returns either
the contents of the safe (an int), the special code SAFE_BAD_COMBINATION if the
combination is incorrect, or the special code SAFE_SELF_DESTRUCTED if the safe
blew up after seeing too many bad combinations. Note that tryCombination
does not declare its second argument to be const and may not leave it intact.
The additional functions allow you to obtain important information about the
safe, like how many tumblers it has and what values these tumblers can be set
to. The behavior of a safe given a combination with the wrong number of values
or values outside the permitted range is undefined.
Your task is to write a function openSafe that will open a safe, if possible, by
trying all possible combinations. Note that if the safe self-destructs before you
can try all the possibilities, this task may not in fact be possible. Your openSafe
function should return SAFE_SELF_DESTRUCTED in this case. Your function
should be defined in a file openSafe.c and should match the declaration in this
file:
/* Include safe.h before this file to get the definition of Safe. */

/*
* Open a safe and return the value returned by tryCombination,
* or SAFE_SELF_DESTRUCTED if the safe self-destructed.
*/
int openSafe(Safe *s);
examples/2015/hw/2/openSafe.h
It is recommended that you put the lines below in your openSafe.c file to ensure
consistency with these declarations:
#include "safe.h"
#include "openSafe.h"
You may put additional functions in openSafe.c if that would be helpful. You
should declare these static to avoid the possibility of namespace conflicts.
In addition to safe.h and openSafe.h, /c/cs223/Hwk2/sourceFiles also con-
tains a main.c file that can be compiled together with openSafe.c to generate a
program that can be called from the command line. This program generates a
safe with a pseudorandom combination based on parameters specified on the
command line, runs your openSafe routine on it, and prints the value that
openSafe returns. You should not rely on your function being tested with this
particular program.

7.2.2 Submitting your assignment

Submit your assignment as usual with

/c/cs223/bin/submit 2 openSafe.c

460
You do not need to submit any other files (and the test script will ignore them
if you do).
You can test that your program compiles and passes a few basic tests with the
command
/c/cs223/bin/testit 2 openSafe
This runs the test script in /c/cs223/Hwk2/test.openSafe on your submit-
ted assignment. You can also run this script by hand to test the version of
openSafe.c in your current working directory.

7.2.3 Valgrind

You may need to allocate storage using malloc to complete this assignment. If
you do so, you should make sure that you call free on any block you allocate
inside your openSafe function before the function returns. The test.openSafe
script attempts to detect storage leaks or other problems resulting from misuse of
these routines by running your program with valgrind. You can also use valgrind
yourself to track down the source of errors, particularly if you remember to
compile with debugging info turned on using the -g3 option to gcc. The script
/c/cs223/bin/vg gives a shortcut for running valgrind with some of the more
useful options.

7.2.4 Sample solution

#include <stdlib.h>
#include <assert.h>

#include "safe.h"
#include "openSafe.h"

/* set combination to all zeros */

static void
zeroCombination(int n, int *combination)
{
int i;

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

combination[i] = 0;
}
}

/* non-destructive version of tryCombination */

static int
nondestructiveTryCombination(Safe *s, const int *combination)

461
{
int *copy; /* duplicate of combination */
int result; /* result of tryCombination */
int n; /* number of tumblers */
int i;

n = numTumblers(s);

copy = (int *) malloc(sizeof(int) * n);

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

copy[i] = combination[i];
}

result = tryCombination(s, copy);

free(copy);

return result;
}

/* update combination to next value */

static void
nextCombination(int n, int base, int *combination)
{
int i;

/* we are essentially incrementing an n-digit number in given base */

/* this means setting any digit that overflows to 0 and continuing */
/* until we get a digit we can increment without carrying */
for(i = 0; i < n && ++(combination[i]) >= base; i++) {
combination[i] = 0;
}
}

int
openSafe(Safe *s)
{
int *combination; /* counter for combinations */
int n; /* number of tumblers */
int base; /* number of positions */
int result; /* result of tryCombination */

/* allocate space */
n = numTumblers(s);

462
 base = numPositions(s);

combination = malloc(sizeof(int) * n);

assert(combination);

for(zeroCombination(n, combination);
(result = nondestructiveTryCombination(s, combination)) == SAFE_BAD_COMBINATION;
nextCombination(n, base, combination));

free(combination);
return result;
}
examples/2015/hw/2/openSafe.c

7.3 Assignment 3, due Wednesday 2015-02-11, at 11:00pm

7.3.1 Quadratic letter sequences

Given a string s, an quadratic letter sequence for s is defined by giving

non-negative integer coefficients c0 , c1 , c2 , where at least one of c1 and c2 is not
zero, and computing the sequence of letters s[c0 + c1 · i + c2 · i2 ] for i = 0, 1, 2, . . ..
These can be used to hide secret message inside larger texts. For example, the
famous Napoleonic palindrome “Able was I ere I saw Elba” hides the word “bIb”
at positions 1, 9, and 23, which are generated by c0 = 1, c1 = 5 and c2 = 3:

i c0 + c1 i + c2 i2
0 1=1+5·0+3·0
1 9 = 1 + 5 · 1 + 3 · 12
2 23 = 1 + 5 · 2 + 3 · 22

Similarly, we can use quadratic letter sequences to reveal secret messages hidden
in the lyrics of K-pop songs:
$ ./qls hail satan < gangnam-style-excerpt.txt
470 3 5 hail
14 10 30 satan
14 56 7 satan
or even examine Act 1 of The Tempest to help resolve the Shakespeare authorship
question:27
27 Stratfordians, Oxfordians, and other conspiracy theorists might object that these results

depend critically on the precise formatting of the text. We counter this objection by observing
that we used the Project Gutenberg e-text of The Tempest, which, while not necessarily the

463
$ ./qls "Bacon" "de Vere" "Marlowe" "Stanley" "that Stratford dude" < tempest-act-one.txt
120 387 777 Bacon
120 542 906 Bacon
120 851 850 Bacon
120 1592 726 Bacon
120 1607 472 Bacon
120 2461 95 Bacon
120 2729 50 Bacon
120 3225 215 Bacon
120 3420 284 Bacon
120 4223 330 Bacon
120 4534 76 Bacon
120 5803 29 Bacon
143 46 161 Bacon
143 268 727 Bacon
143 684 1434 Bacon
[... 280 more lines of Bacon omitted ...]
19959 1178 87 Bacon
5949 239 465 Marlowe

7.3.2 Your task

Write a program qls.c that takes a text on stdin and searches for quadratic
letter sequences that start with the strings given in argv. Your program should
output all such quadratic letter sequences that it finds, using the format
printf("%d %d %d %s\n", [...]);
where [...] should be replaced by appropriate expressions to give c0 , c1 , c2 ,
and the string found.
If a string appears more than once at the start of a quadratic letter sequence,
your program should print all occurrences. The order your output lines appear
in is not important, since the test script sorts them into a canonical order. Do
whatever is convenient.
Your program should be reasonably efficient, but you do not need to get carried
away looking for a sophisticated algorithm for this problem. Simply testing all
plausible combinations of coefficients should be enough.
Because neither K-pop songs nor Elizabethan plays use null characters, you may
assume that no null characters appear in your input.
most favored by academic Shakespeare scholars, is the easiest version to obtain on-line. We
consider it further evidence of Sir Francis Bacon’s genius that not only was he able to subtly
encode his name throughout his many brilliant plays, but he was even able to anticipate the
effects of modern spelling and punctuation on this encoding.

464
You may also assume that any search strings will contain at least two characters,
in order to keep the number of outputs finite.

7.3.3 Submitting your assignment

Submit your assignment as usual with

/c/cs223/bin/submit 3 qls.c
You can run some basic tests on your submitted solution with
/c/cs223/bin/testit 3 qls
The test program is also available as /c/cs223/Hwk3/test.qls. Sample inputs
and outputs can be found in /c/cs223/Hwk3/testFiles. The title of each file
contains the test strings used, separated by - characters. Before comparing the
output of your program to the output files, you may find it helpful to run it
through sort, e.g.
./qls hail satan < hail-satan.in | sort > test.out
diff test.out hail-satan.out

7.3.4 Sample solution

/*
* Search for quadratic letter sequences starting with words from argv on stdin.
*
* A quadratic letter sequence of length n in s is a sequence of characters
*
* s[c0 + c1*i + c2*i*i]
*
* where c0, c1, c2 are all >= 0, at least one of c1 and c2 is > 0,
* and i ranges over 0, 1, 2, ..., n-1.
*
* For each QLS found, prints c0, c1, c2, and the target string to stdout.
*/

#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <assert.h>

#define NUM_COEFFICIENTS (3) /* how many coefficients to pass around */

/*
* Return true iff we get a match in s for t with coefficients c
*

465
 * Behavior is undefined if coefficients would send us off the end of s.
*/
static int
qlsMatch(const char *s, const char *t, int c[NUM_COEFFICIENTS])
{
int i;

for(i = 0; t[i] != '\0'; i++) {

if(s[c[0] + c[1] * i + c[2] * i * i] != t[i]) {
/* no match */
return 0;
}
}

return 1;
}

/*
* Search for quadratic letter sequences in s starting with t
* and print results to stdout.
*/
static void
qlsSearch(const char *s, const char *t)
{
int c[NUM_COEFFICIENTS]; /* coefficients */
int lenS; /* length of s */
int lenT; /* length of t */
int maxI; /* maximum value for i (this is lenT-1) */

lenS = strlen(s);
lenT = strlen(t);
maxI = lenT-1;

/* try all possible c[0] that will let us finish before lenS */
for(c[0] = 0; c[0] + maxI < lenS; c[0]++) {
/* if s[c[0]] isn't right, c[1] and c[2] can't fix it */
if(s[c[0]] == t[0]) {
/* try all feasible c[1] */
for(c[1] = 0; c[0] + c[1] * maxI < lenS; c[1]++) {
/* try all feasible c[2], but start at 1 if c[1] == 0 */
for(c[2] = (c[1] == 0); c[0] + c[1] * maxI + c[2] * maxI * maxI < lenS; c[2]
/* now see if we get a match */
if(qlsMatch(s, t, c)) {
printf("%d %d %d %s\n", c[0], c[1], c[2], t);
}
}

466
 }
}
}
}

/* used internally by getContents; initial size of buffer */

#define INITIAL_BUFFER_SIZE (16)

/*
* Return a single string holding all characters from stdin.
*
* This is malloc'd data that the caller should eventually free.
*/
static char *
getContents(void)
{
size_t size;
size_t len;
char *text;
int c;

size = INITIAL_BUFFER_SIZE;
len = 0;

text = malloc(size);
assert(text);

while((c = getchar()) != EOF) {

/* grow the buffer if full */
if(len >= size) {
size *= 2;
text = realloc(text, size);
assert(text);
}

text[len++] = c;
}

/* cleanup */
text = realloc(text, len+1);
assert(text);

text[len] = '\0';

return text;
}

467
int
main(int argc, char **argv)
{
int i;
char *s;

s = getContents();

for(i = 1; i < argc; i++) {

qlsSearch(s, argv[i]);
}

free(s);

return 0;
}
examples/2015/hw/3/qls.c

7.4 Assignment 4, due Wednesday 2015-02-18, at 11:00pm

7.4.1 An ASCII art compositor

For this assignment you are to write a program that takes from stdin a sequence
of instructions for pasting ASCII art pictures together, reads those pictures from
files, and writes the combined picture to stdout.
Each instruction is of the form row column filename, suitable for reading
with scanf("%d %d %s", &row, &col, filename);, where row and col are
declared as ints and filename is a suitably large buffer of chars. Such an
instruction means to paste the contents of file filename into the picture with
each character shifted row rows down and column columns to the right of its
position in file filename. When pasting an image, all characters other than space
(' ', or ASCII code 32) overwrite any characters from earlier files at the same
position. Spaces should be treated as transparent, having no effect on the final
image.
For example, suppose that the current directory contains these files:
# # #
\==========/
\......../
examples/2015/hw/4/ship
/\

468
 /vv\
/vvvv\
||
examples/2015/hw/4/tree
* * *
____|_|_|_____
|_____________|
|___HAPPY_____|
|__BIRTHDAY___|
|_____________|
examples/2015/hw/4/cake
Then this is what we should get from executing the command:
$ echo "1 1 ship 3 5 ship 3 19 tree 7 2 ship 13 4 ship 4 22 tree 5 6 cake" | ./compositor

# # #
\==========/
\......#.# # /\
\==========/ /vv\/\
\....*.*.* /vvv/vv\
____|_|_|_____|/vvvv\
|_____________| ||
\===|___HAPPY_____|
\..|__BIRTHDAY___|
|_____________|

# # #
\==========/
\......../
examples/2015/hw/4/example.out

7.4.2 Submitting your assignment

For this assignment, you may submit whatever source files you like, along with
a file Makefile that will generate the program compositor when make is called
with no arguments (see the instructions for using make.)
You can test your submitted assignment using the public test script with
/c/cs223/bin/testit 4 public
You may also test your unsubmitted assignment in the current working directory
with

469
/c/cs223/Hwk4/test.public
The test script is intended mostly to guard against trivial errors in output format
and is not necessarily exhaustive.

7.4.3 Notes

7.4.3.1 Input
For parsing the commands on stdin, we recommend using scanf. You can test
for end of file by checking if scanf correctly parsed all three arguments, as in
int row;
int col;
char filename[BUFFER_SIZE];

while(scanf("%d %d %s", &row, &col, filename) == 3) {

/* do something with this file */
}
You may assume that row and col are always non-negative.
Your program should exit with a non-zero error code if it cannot open a file
for reading. Because scanf’s %s conversion specifier only reads up to the next
whitespace character, you may assume that filenames do not contain whitespace.
You may also assume that no filename appearing in the input will require more
than 2048 bytes to store (including the terminal null character).28
You may assume that the input contains no null characters.

7.4.3.2 Output
Your output should include newline and space characters to put the composited
characters in the appropriate rows and columns. It should not include any more
of such characters than are absolutely necessary.
For example, there should never be a space at the end of a line (even if there
is a space at the end of a line in one of the input files). Similarly, there should
not be any blank lines at the end of your output. You may, however, find it
necessary to add a newline to the end of the last line to avoid having the output
end in the middle of a line.

7.4.3.3 General
You may assume that the final picture is not so big that you can’t store a row
or column number for one of its characters in an int.
28 Normally this is a dangerous thing to assume, but this assignment is complicated enough

already.

470
7.4.4 Sample solution

I wrote two versions of this. The first used a jagged array to represent an image,
but I decided I didn’t like it and did another version using a sorted linked list of
points. This second version is shown below.
/*
* Alternate version of ASCII art thing using a queue.
*/

#include <stdio.h>
#include <stdlib.h>
#include <assert.h>

/*
* Idea of this data structure is that we have a sorted array
* of pixels, where each pixel specifies a row, column, and character
* to put in that position. The sort order is row then column.
*
* This is organized as a queue in the sense that we can push
* new pixels on to the end of it, although as it happens we
* never actually dequeue anything.
*/
struct pixel {
int row;
int col;
char value;
};

struct queue {
size_t top; /* number of elements */
size_t size; /* number of allocated slots */
struct pixel *pixels; /* pixel values, sorted by row then column */
};

#define QUEUE_INITIAL_SIZE (16)

/* create new empty queue */

struct queue *
queueCreate(void)
{
struct queue *q;

q = malloc(sizeof(struct queue));
assert(q);

471
 q->top = 0;
q->size = QUEUE_INITIAL_SIZE;

q->pixels = malloc(sizeof(struct pixel) * q->size);

assert(q->pixels);

return q;
}

/* clean up queue */
void
queueDestroy(struct queue *q)
{
free(q->pixels);
free(q);
}

/* add a new pixel to queue */

void
queuePush(struct queue *q, struct pixel p)
{
while(q->top >= q->size) {
q->size *= 2;
q->pixels = realloc(q->pixels, sizeof(struct pixel) * q->size);
assert(q->pixels);
}

q->pixels[q->top++] = p;
}

/* returns malloc'd data, free with queueDestroy */

struct queue *
queueRead(const char *filename)
{
FILE *f;
struct queue *q;
struct pixel p;
int c;

q = queueCreate();

f = fopen(filename, "r");
if(f == 0) {
perror(filename);
exit(1);
}

472
 p.row = p.col = 0;

while((c = getc(f)) != EOF) {

switch(c) {
case '\n':
p.row++;
p.col = 0;
break;
case ' ':
p.col++;
break;
default:
p.value = c;
queuePush(q, p);
p.col++;
break;
}
}

fclose(f);

return q;
}

/* write pixels in queue to stdout */

void
queueWrite(const struct queue *q)
{
int outputRow = 0;
int outputCol = 0;
int i;

for(i = 0; i < q->top; i++) {

while(outputRow < q->pixels[i].row) {
putchar('\n');
outputRow++;
outputCol = 0;
}
while(outputCol < q->pixels[i].col) {
putchar(' ');
outputCol++;
}
putchar(q->pixels[i].value);
outputCol++;
}

473
 /* end last row */
putchar('\n');
}

/*
* Merge two queues, creating a new, freshly-allocated queue.
* New queue is sorted. If there are pixels in both left
* and right with the same row and column, the one from right
* overwrites the one from left.
*/
struct queue *
queueMerge(const struct queue *left, const struct queue *right)
{
int l = 0;
int r = 0;
struct queue *q;

q = queueCreate();

while(l < left->top && r < right->top) {

if(left->pixels[l].row < right->pixels[r].row) {
queuePush(q, left->pixels[l++]);
} else if(left->pixels[l].row == right->pixels[r].row) {
if(left->pixels[l].col < right->pixels[r].col) {
queuePush(q, left->pixels[l++]);
} else if(left->pixels[l].col == right->pixels[r].col) {
/* right wins but both increment */
queuePush(q, right->pixels[r++]);
l++;
} else {
/* right is earlier */
queuePush(q, right->pixels[r++]);
}
} else {
/* right is earlier */
queuePush(q, right->pixels[r++]);
}
}

/* clean out whichever tail is still nonempty */

while(l < left->top) {
queuePush(q, left->pixels[l++]);
}

while(r < right->top) {

474
 queuePush(q, right->pixels[r++]);
}

return q;
}

/* in-place offset by r rows and c columns */

void
queueOffset(struct queue *q, int r, int c)
{
int i;

for(i = 0; i < q->top; i++) {

q->pixels[i].row += r;
q->pixels[i].col += c;
}
}

/* max filename size as promised in assignment text */

#define BUFFER_SIZE (2048)

int
main(int argc, char **argv)
{
struct queue *merged; /* holding place for result of merge */
struct queue *left; /* accumulated picture */
struct queue *right; /* new picture */
int row; /* row offset for new picture */
int col; /* column offset for new picture */
char filename[BUFFER_SIZE]; /* filename for new picture */

if(argc != 1) {
fprintf(stderr, "Usage: %s\n", argv[0]);
return 1;
}

for(left = queueCreate(); scanf("%d %d %s", &row, &col, filename) == 3; left = merged) {

right = queueRead(filename);
queueOffset(right, row, col);

merged = queueMerge(left, right);

queueDestroy(left);
queueDestroy(right);
}

475
 queueWrite(left);

queueDestroy(left);

return 0;
}
examples/2015/hw/4/compositor.c
Here is a Makefile.

7.5 Assignment 5, due Wednesday 2015-02-25, at 11:00pm

7.5.1 Build a Turing machine!

A Turing machine is a hypothetical computational device that consists of a

little trolley, called the controller, that rolls around on an infinite paper tape
on which it can write letters. The controller itself has a small number of states
that affect its behavior, and a program that tells it what to do when it is in a
particular state and sees a particular letter.
In C terms, we can think of a Turing machine as consisting of an infinite array of
chars (representing the tape), an integer index into the array (representing the
current position of the controller), and an int (representing the current state of
the controller). At each step, the machine
1. Looks at the symbol s on the current tape cell.
2. Looks up its state q and symbol s in the table representing its program.
This table will specify an action, which consists of three parts:
a. A new symbol to write to the current tape cell.
b. A direction - or + to move.
c. A new state to switch to.
If the new state is 0, the machine halts and takes no more steps.
For this assignment, you are to write a Turing machine simulator. The program
for machine will be supplied in argv; argv[i] will give the behavior of the
machine when in state i, where the first three characters specify what to do
if the machine sees an 'a', the second three characters specify what to do if
the machine sees a 'b', and so on. Each of these three-letter groups will look
like (new-symbol,direction,new-state), where new-symbol is the new symbol to
write (a lower-case letter), direction is the direction to move ('+', meaning one
position to the right, or '-', meaning one position to the left) and new-state is
the new state to go to (a digit). Your program should run this program starting
in state 1 on some location in the middle of a tape initially made up entirely of
'a' characters, and continue until the program enters the special halting state 0.
It should then print the number of steps until this occurred.

476
You may assume that the program is argv is complete in the sense that it
includes rules for any combination of state and symbol you will encounter while
executing it. You are not required to detect if this assumption is violated.

7.5.2 Example

The program
b+2a-0 a-1a-1
gives instructions for what to do in state 1 (b+2a-0) and state 2 (a-1a-1). In
state 1, if the controller reads an a, the triple b+2 means that it should write
b, move right (+), and switch to state 2. If instead it reads a b, the triple a-0
means that it should write a, move left (-), and halt (0). In state 2, the machine
always writes a, moves left, and switches to state 1.
Below is a depiction of this machine’s execution. It passes through 4 states
(including both the initial state and the final halting state) using a total of 3
steps. The controller and its current state is shown above its current position on
the tape at each point in time. To avoid having to put in infinitely long lines,
only the middle three tape cells are shown.
1
aaa

2
aba

1
aba

0
aaa

7.5.3 Your task

You should submit a Makefile and whatever source files are needed to generate a
program ./turing when make is called with no arguments. The turing program
should simulate a Turing machine as described above and print the number of
steps that it takes until it halts in decimal format, followed by a newline. It
should not produce any other output. For example, using the program above,
your program should print 3:
$ ./turing b+2a-0 a-1a-1
3

477
For more complex programs you may get different results. Here is a 3 state, 3
symbol program that runs for a bit longer:
$ ./turing b+2a-0c-3 b-3c+2b-2 b-1a+2c-1
92649163
You may assume that tape symbols can always be represented by lowercase
letters, that states can always be represented by single digits, and that argv is
in the correct format (although it may be worth including a few sanity checks in
your program just in case).
Not all Turing machine programs will halt. Your program is not required to
detect if the Turing machine it is simulating will halt eventually or not (although
it should notice if it does halt).

7.5.4 Submitting your assignment

Submit all files needed to build your program as usual using /c/cs223/bin/submit
5 filename.
There is a public test script in /c/cs223/Hwk5/test.public. You can run this
on your submitted files with /c/cs223/bin/testit 5 public.

7.5.5 Sample solution

/*
* Simple Turing machine simulator.
*
* Tape holds symbols 0 (default) through 2.
*
* Controller programming is specified in argv:
*
* argv[i] gives transitions for state i as six characters.
*
* Each triple of characters is <action><direction><new-state>
*
* where <action> is one of:
*
* a,b,c: write this value to tape
*
* <direction> is one of:
*
* -: go left
* +: go right
* .: stay put
*

478
 * The three pairs give the transition for reading 0, 1, 2 from tape.
*
* State 0 is the halting state.
*
* On halting, prints number of transitions followed by contents
* of all tape cells that have ever been visited by the
* finite-state controller.
*/

#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <assert.h>
#include <sys/types.h>

struct configuration {
unsigned int state;/* state of head */
size_t leftmost; /* leftmost cell visited */
size_t rightmost; /* rightmost cell visited */
size_t current; /* current cell */
size_t tapeLength; /* current allocated space for tape */
char *tape; /* contents of cells */
};

/* increase the size of the tape and recenter contents in middle */

void
configurationExpand(struct configuration *c)
{
size_t newTapeLength;
char *oldTape;
char *newTape;
size_t i;
ssize_t offset;

newTapeLength = 4*c->tapeLength;
newTape = malloc(newTapeLength);
assert(newTape);

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

newTape[i] = 0;
}

/* copy old tape */

offset = newTapeLength / 2 - c->current;

for(i = c->leftmost; i <= c->rightmost; i++) {

479
 newTape[i + offset] = c->tape[i];
}

oldTape = c->tape;
c->tape = newTape;
c->tapeLength = newTapeLength;
c->current += offset;
c->leftmost += offset;
c->rightmost += offset;

free(oldTape);
}

#define INITIAL_TAPE_LENGTH (16)

struct configuration *
configurationCreate(void)
{
struct configuration *c;
size_t i;

c = malloc(sizeof(struct configuration));
assert(c);

c->state = 1;
c->tapeLength = INITIAL_TAPE_LENGTH;
c->leftmost = c->rightmost = c->current = c->tapeLength / 2;
c->tape = malloc(c->tapeLength);
assert(c->tape);

for(i = 0; i < c->tapeLength; i++) {

c->tape[i] = 0;
}

return c;
}

void
configurationDestroy(struct configuration *c)
{
free(c->tape);
free(c);
}

#define SYMBOL_BASE ('a')

#define STATE_BASE ('0')

480
/* used for debugging mostly */
void
configurationPrint(const struct configuration *c)
{
size_t i;

for(i = c->leftmost; i < c->current; i++) {

putchar(' ');
}
putchar(STATE_BASE + c->state);
putchar('\n');

for(i = c->leftmost; i <= c->rightmost; i++) {

putchar(SYMBOL_BASE + c->tape[i]);
}
putchar('\n');
}

int
main(int argc, char **argv)
{
struct configuration *c;
char cellValue;
const char *transition;
size_t steps;

if(argc == 1) {
fprintf(stderr, "Usage: %s transitions\n", argv[0]);
return 1;
}

c = configurationCreate();
steps = 0;

while(c->state != 0) {
steps++;

/* execute the next transition */

assert(c->state < argc);

cellValue = c->tape[c->current];
assert(0 <= cellValue);
assert(3*(cellValue+1) <= strlen(argv[c->state]));

transition = argv[c->state] + 3*c->tape[c->current];

481
 c->tape[c->current] = transition[0] - SYMBOL_BASE;

switch(transition[1]) {
case '-':
if(c->current == 0) {
configurationExpand(c);
}
c->current--;
if(c->current < c->leftmost) {
c->leftmost = c->current;
}
break;
case '+':
if(c->current == c->tapeLength - 1) {
configurationExpand(c);
}
c->current++;
if(c->current > c->rightmost) {
c->rightmost = c->current;
}
break;
case '.':
/* do nothing */
break;
default:
fprintf(stderr, "Bad direction '%c'\n", transition[2]);
exit(2);
break;
}

c->state = transition[2] - STATE_BASE;

#ifdef PRINT_CONFIGURATION
configurationPrint(c);
#endif
}

/* print number of steps */

printf("%zu\n", steps);

configurationDestroy(c);

return 0;
}

482
examples/2015/hw/5/turing.c
CC=gcc
CFLAGS=-std=c99 -Wall -pedantic -g3

all: turing

compositor: turing.o
$(CC) $(CFLAGS) -o $@ $^

clean:
$(RM) turing *.o
examples/2015/hw/5/Makefile

7.6 Assignment 6, due Wednesday 2015-03-25, at 11:00pm

7.6.1 Sinking ships

For this assignment, you are to implement a data structure for playing a game
involving ships placed in a large square grid. Each ship occupies one more more
squares in either a vertical or horizontal line, and has a name that consists of a
single char other than a period (which will be used to report the absence of a
ship). Ships have a bounded maximum length; attempts to place ships longer
than this length have no effect.
All type and constant definitions for the data type, and all function declarations,
are given in the file ships.h, which is shown below, and which you can also find
in /c/cs223/Hwk6/sourceFiles/ships.h. The playing field is represented by
a struct field (which you get to define). A new struct field is created by
fieldCreate, and when no longer needed should be destroyed by fieldDestroy.
These data types from ships.h control ship naming and placement. Note that
uint32_t is defined in stdint.h (which is also included by inttypes.h. You
will need to include one of these files before ships.h to get this definition.
typedef uint32_t coord;

struct position {
coord x;
coord y;
};

struct ship {
struct position topLeft; /* coordinates of top left corner */
int direction; /* HORIZONTAL or VERTICAL */
unsigned int length; /* length of ship */

483
 char name; /* name of ship */
};
Actual placement is done using the fieldPlaceShip function, declared as follows:
void fieldPlaceShip(struct field *f, struct ship s);
A ship of length m placed horizontally with its top left corner at position (x, y)
will occupy positions (x, y) through (x+m−1, y). If instead it is placed vertically,
it will occupy positions (x, y) through (x, y + m − 1). If any of these coordinates
exceed the maximum coordinate COORD_MAX (defined in ships.h), the ship will
not be placed. The ship will also not be placed if its name field is equal to
NO_SHIP_NAME or if the length exceeds MAX_SHIP_LENGTH.
If the new ship will occupy any position as a ship previously placed in the field,
the previous ship will be removed. It is possible for many ships to be removed
at once in this way.
The fieldAttack function can be used to remove a ship at a particular location
without placing a new ship. It returns the name of the removed ship, if any, or
NO_SHIP_NAME if there is no ship at that location.
Finally, the fieldCountShips returns the number of ships still present in the
field.
Your job is to write an implementation of these functions, which you should
probably put in a file ships.c. You must also supply a Makefile, which,
when make is called with no arguments, generates a test program testShips
from your implementation and the file testShips.c that we will provide. You
should not count on precisely this version of testShips.c being supplied; your
implementation should work with any main program that respects the interface
in ships.h.

7.6.2 Things to watch out for

You should write your implementation so that it will continue to work if the
typedef for coord, or the definitions of the constants COORD_MAX, NO_SHIP_NAME,
SHIP_MAX_LENGTH, HORIZONTAL, or VERTICAL change. You may, however, assume
that coord is an unsigned integer type and the COORD_MAX is the largest value
that can be represented by this type.
If it helps in crafting your implementation, you may assume that
MAX_SHIP_LENGTH will alway be a reasonably small constant. You do
not need to worry about implementing a data structure that will handle huge
ships efficiently. On the other hand, COORD_MAX as defined in the default
ships.h is 23 2 − 1, so you will need to be able to deal with a field with at
least 26 4 possible locations, a consideration you should take into account when
choosing a data structure to represent a field.

484
7.6.3 The testShips program

The supplied testShips program creates a field, processes commands from

stdin, then destroys the field when it reaches EOF or a command that it can’t
parse. Each command is either of the form
+ x y vertical length name
which means to call fieldPlaceShip for a new ship at coordinates (x,y), with
direction HORIZONTAL if isVertical is false and VERTICAL otherwise, and length
length and name name.
The command
- x y
calls fieldAttack on coordinates (x,y).
Here is a small input file:
+ 0 0 0 3 a
+ 1 2 0 4 b
+ 3 0 1 3 c
- 1 0
After processing the first line, the contents of the field should look like this:
aaa..
.....
.....
.....
.....
After processing the second line:
aaa..
.....
.bbbb
.....
.....
The third line places a ship vertically that intersects one of the previous ships,
sinking it:
aaac.
...c.
...c.
.....
.....
Had ship c been shifted one position to the left, it would have sunken a as well.
Finally, the last line drops a bomb at (1, 0), sinking a:

485
...c.
...c.
...c.
.....
.....
The input files used by test.public can be found in /c/cs223/Hwk6/testFiles.
Some of these were generated randomly using the script /c/cs223/Hwk6/makeRandom,
which you should feel free to use for your own nefarious purposes.
Because the interface in ships.h gives no way to find out what ships are currently
in the field, the test program will not actually produce pictures like the above.
Instead, it prints after each command a line giving the name of the ship sunken by
fieldAttack (or NO_SHIP_NAME if no ship is sunk or fieldPlaceShip is called)
and the number of ships left in the field following the attack. So the user must
imagine the carnage as the 100000 ships in randomSparseBig.in somehow leave
only 25336 survivors in randomSparseBig.out, demonstrating the importance
of strict navigational rules in real life.

Figure 8: Why you should respect people’s space

486
7.6.4 Submitting your assignment

Submit your assignment as usual with

/c/cs223/bin/submit 6
You can run the public test script in /c/cs223/Hwk6/test.public on your
submitted files with
/c/cs223/bin/testit 6 public

7.6.5 Provided source files

#define HORIZONTAL (0) /* place ship horizontally */

#define VERTICAL (1) /* place ship vertically */

#define MAX_SHIP_LENGTH (17) /* length of longest ship (width is always 1) */

#define NO_SHIP_NAME ('.') /* what to return when hitting no ship */

/*
* Type for coordinates, and their maximum possible value.
*
* Include <stdint.h> before this header file
* to get the definition of uint32_t
* and its maximum value UINT32_MAX.
*/
typedef uint32_t coord;
#define COORD_MAX (UINT32_MAX)

/*
* Non-opaque structs for passing around positions and ship placements.
*/
struct position {
coord x;
coord y;
};

struct ship {
struct position topLeft; /* coordinates of top left corner */
int direction; /* HORIZONTAL or VERTICAL */
unsigned int length; /* length of ship */
char name; /* name of ship */
};

/*

487
 * Create a playing field for holding ships.
*/
struct field *fieldCreate(void);

/*
* Free all space associated with a field.
*/
void fieldDestroy(struct field *);

/*
* Place a ship in a field with given placement and name.
*
* If placement.length is less than one or greater than MAX_SHIP_LENGTH,
* or if some part of the ship would have a coordinate greater than COORD_MAX,
* or if the ship's name is NO_SHIP_NAME,
* the function returns without placing a ship.
*
* Placing a new ship that intersects any previously-placed ships
* sinks the previous ships, removing them from the field.
*/
void fieldPlaceShip(struct field *f, struct ship s);

/*
* Attack!
*
* Drop a shell at given position.
*
* Returns NO_SHIP_NAME if attack misses (does not intersect any ship).
*
* Otherwise returns name of ship hit.
*
* Hitting a ship sinks it, removing it from the field.
*/
char fieldAttack(struct field *f, struct position p);

/*
* Return number of ships in the field.
*/
size_t fieldCountShips(const struct field *f);
examples/2015/hw/6/ships.h
#include <stdio.h>
#include <stdlib.h>
#include <assert.h>
#include <stdint.h>
#include <inttypes.h>

488
#include "ships.h"

#define PLACE_SHIP ('+') /* command to place a new ship */

#define ATTACK ('-') /* command to attack a location */

int
main(int argc, char **argv)
{
struct field *f; /* where we keep our ships */
int command; /* command char */
struct ship s; /* ship we are placing */
struct position p; /* location to attack */
int sank; /* ship we sank */

if(argc != 1) {
fprintf(stderr, "Usage: %s\n", argv[0]);
return 1;
}

f = fieldCreate();

while((command = getchar()) != EOF) {

switch(command) {
case PLACE_SHIP:
if(scanf("%" SCNu32 " %" SCNu32 "%d %u %c ", &s.topLeft.x, &s.topLeft.y, &s.
/* not enough args */
fprintf(stderr, "Not enough enough args to %c\n", PLACE_SHIP);
return 1;
}
/* else */

/* fix the direction to match actual definitions */

s.direction = s.direction ? VERTICAL : HORIZONTAL;

fieldPlaceShip(f, s);
sank = NO_SHIP_NAME;

break;

case ATTACK:
if(scanf("%" SCNu32 " %" SCNu32 " ", &p.x, &p.y) != 2) {
fprintf(stderr, "Not enough enough args to %c\n", ATTACK);
return 1;
}
/* else */

489
 sank = fieldAttack(f, p);

break;

default:
/* bad command */
fprintf(stderr, "Bad command %c\n", command);
return 1;
break;
}

printf("%c %zu\n", sank, fieldCountShips(f));

}

fieldDestroy(f);

return 0;
}
examples/2015/hw/6/testShips.c

7.6.6 Sample solution

#include <stdlib.h>
#include <assert.h>
#include <string.h>
#include <stdint.h>

#include "ships.h"

/* basic hash table */

struct field {
size_t size; /* number of slots in table */
size_t occupancy; /* number of elements in table */
struct elt **table; /* hash table, malloc'd */
};

struct elt {
struct elt *next; /* pointer to next element in linked list */
struct ship ship; /* ship in this element */
};

/* picked more or less at whim from http://planetmath.org/goodhashtableprimes */

#define X_HASH_FACTOR (201326611)
#define Y_HASH_FACTOR (3145739)

490
static size_t
hash(struct position p)
{
return X_HASH_FACTOR * p.x + Y_HASH_FACTOR * p.y;
}

#define DEFAULT_INITIAL_SIZE (8)

/* like fieldCreate, but argument gives initial size */

static struct field *
fieldCreateInternal(size_t initialSize)
{
struct field *f;
size_t i;

f = malloc(sizeof(struct field));
assert(f);

f->size = initialSize;
f->occupancy = 0;

f->table = malloc(sizeof(struct elt *) * f->size);

for(i = 0; i < f->size; i++) {

f->table[i] = 0;
}

return f;
}

struct field *
fieldCreate(void)
{
return fieldCreateInternal(DEFAULT_INITIAL_SIZE);
}

/* destroy contents of f but don't free f itself */

static void
fieldDestroyContents(struct field *f)
{
size_t i;
struct elt *e;
struct elt *next;

for(i = 0; i < f->size; i++) {

491
 for(e = f->table[i]; e != 0; e = next) {
next = e->next;
free(e);
}
}

free(f->table);
}

void
fieldDestroy(struct field *f)
{
fieldDestroyContents(f);
free(f);
}

/* when to grow field */

#define MAX_ALPHA (1)

/*
* Helper for fieldPlaceShip.
*
* This skips all the sanity-checking in fieldPlaceShip,
* and just performs the hash table insertion.
*/
static void
fieldInsertShip(struct field *f, struct ship s)
{
size_t h; /* hashed coordinates */
struct elt *e; /* new element to insert */

h = hash(s.topLeft) % f->size;

e = malloc(sizeof(struct elt));
assert(e);

e->ship = s;
e->next = f->table[h];
f->table[h] = e;
f->occupancy++;
}

void
fieldPlaceShip(struct field *f, struct ship s)
{

492
struct field *f2;
struct elt *e;
struct position pos;
size_t i;

/* test if we can just throw this away */

if(s.name == NO_SHIP_NAME
|| s.length == 0
|| s.length > MAX_SHIP_LENGTH
|| (s.direction == HORIZONTAL && s.topLeft.x > COORD_MAX - (s.length - 1))
|| (s.direction == VERTICAL && s.topLeft.y > COORD_MAX - (s.length - 1))
)
{
return;
}
/* else */

if(f->occupancy >= f->size * MAX_ALPHA) {

/* grow the field */
f2 = fieldCreateInternal(f->size * 2);

/* copy to new field */

for(i = 0; i < f->size; i++) {
for(e = f->table[i]; e != 0; e = e->next) {
/* skip testing for occupancy or intersections */
fieldInsertShip(f2, e->ship);
}
}

/* transplant new field into old field */

fieldDestroyContents(f);
*f = *f2;

free(f2);
}

/* check for intersections */

pos = s.topLeft;
for(i = 0; i < s.length; i++) {
if(s.direction == HORIZONTAL) {
pos.x = s.topLeft.x + i;
} else {
pos.y = s.topLeft.y + i;
}

fieldAttack(f, pos);

493
 }

/* call helper to do the actual hash table insertion */

fieldInsertShip(f, s);
}

/*
* Helper for fieldAttack.
*
* If there is a ship with topLeft at given position, return pointer
* to location in hash table that points to it (either table entry
* or next component).
*
* If not, return null.
*/
static struct elt **
fieldShipAt(struct field *f, struct position p)
{
struct elt **prev; /* previous pointer */

for(prev = &f->table[hash(p) % f->size]; *prev != 0; prev = &((*prev)->next)) {

if((*prev)->ship.topLeft.x == p.x && (*prev)->ship.topLeft.y == p.y) {
return prev;
}
}

/* didn't find anything */

return 0;
}

/*
* Attack!
*
* Drop a shell at given position.
*
* Returns 0 if attack misses (does not intersect any ship).
*
* Otherwise returns name of ship hit,
* which should be freed by caller when no longer needed.
*
* Hitting a ship sinks it, removing it from the field.
*/
char
fieldAttack(struct field *f, struct position p)
{
struct position p2;

494
 int i;
int direction;
struct elt **prev;
struct elt *freeMe;
char name;

for(direction = 0; direction <= 1; direction++) {

for(i = 0; i < MAX_SHIP_LENGTH && i <= (direction == HORIZONTAL ? p.x : p.y); i++) {
if(direction == HORIZONTAL) {
p2.x = p.x - i;
p2.y = p.y;
} else {
p2.x = p.x;
p2.y = p.y - i;
}

prev = fieldShipAt(f, p2);

if(prev) {
/* if we sink anybody, it will be this ship */
/* but maybe it doesn't reach */
/* or points in the wrong direction */
if((*prev)->ship.length > i && (*prev)->ship.direction == direction) {
/* got it */
freeMe = *prev;
*prev = freeMe->next;

name = freeMe->ship.name;
free(freeMe);

f->occupancy--;

return name;
} else {
/* didn't get it */
/* maybe try again in other direction */
break;
}
}
}
}

/* didn't get anything */

return NO_SHIP_NAME;
}

495
/*
* Return number of ships in the field.
*/
size_t
fieldCountShips(const struct field *f)
{
return f->occupancy;
}
examples/2015/hw/6/ships.c

7.7 Assignment 7, due Wednesday 2015-04-01, at 11:00pm

7.7.1 Solitaire with big cards

For this assignment you are to implement a strategy for playing a card game
involving moving cards (represented by uint64_ts) down through a sequence of
n piles. The interface to your strategy is given in the file strategy.h, shown
below:
/*
* Interface for card-playing strategy.
*
* The deal function supplies a new card to the strategy. Each possible card will only be d
*
* The play function should return a card that has been dealt previously but not yet played.
* If asked for a card when the hand is empty, its behavior is undefined.
*/

#include <stdint.h>

typedef uint64_t Card; /* representation of a card */

/* opaque type for strategy data */

typedef struct strategy Strategy;

/* set up a new strategy for numPiles many piles */

Strategy *strategyCreate(int numPiles);

/* clean up all space used by a strategy */

void strategyDestroy(Strategy *);

/* add a card to the current hand */

void strategyDeal(Strategy *, Card);

496
/* play a card from pile k */
Card strategyPlay(Strategy *, int k);
examples/2015/hw/7/strategy.h
Initially, the player has n piles, numbered 1 through n. The strategyDeal
function is called to indicate that a new card has been dealt to pile n. The
strategyPlay function is called to indicate that a card should be moved from
pile k to pile k-1; this function should return the card to move. Cards moved to
pile 0 leave the game and are not used again. Each card is unique: once a card is
dealt, the same card will never be dealt again during the same play of the game.
The choice of when to deal and when to play from pile is controlled by some
external entity, which at some point will stop and compute the smallest card in
each pile. The goal of the strategy is to make these smallest cards be as large as
possible, giving priority to the highest-numbered piles: given two runs of the
game, the better-scoring one is the one that has the larger smallest card in pile
n, or, if both have the same smallest card in pile n, the one that has the larger
smallest card in pile n − 1, and so forth. A tie would require that both runs end
with the same smallest card in every pile. An empty pile counts as UINT64_MAX
for this purpose (although note that a strategy has no control over which piles
are empty).
Your job is to implement a strategy that produces the best possible result
for any sequence of calls to strategyDeal and strategyPlay. Your strategy
implementation will most likely need to keep track of which cards are available in
each pile, as this information is not provided by the caller. Your strategyPlay
function should only make legal moves: that is, it should only play cards that are
actually present in the appropriate pile. You may assume that strategyPlay is
never called on an empty pile.
Your implementation should consist of a file strategy.c and any support-
ing source and header files that you need other than strategy.h, which we
have provided for you. You should also supply a file Makefile that gener-
ates a program testStrategy when make is called with no arguments, us-
ing your implementation and the testStrategy.c file that you can find in
/c/cs223/Hwk7/sourceFiles/testStrategy.c.

7.7.2 Explanation of the testing program

The testStrategy program implements one of four rules for when you can play
from each pile. The arguments to testStrategy are a character indicating which
rule to apply, the number of cards to deal (which can be pretty big), and the
number of piles (which is much more limited, because testStrategy.c tracks
the pile each card is in using a char to save space). The actual cards dealt are
generated deterministically and will be the same in every execution with the
same arguments. The test files in /c/cs223/Hwk7/testFiles give the expected

497
output when testStrategy is run with the arguments specified in the filename
(after removing the - characters); this will always be the value, in hexadecimal,
of the smallest card in each pile, starting with the top pile.
For example, running the harmonic rule h with 1000 cards and 4 piles (not
counting the 0 pile) gives the output
$ ./testStrategy h 1000 4
5462035faf0d6fa1
501ebb6268d39af3
25732b5fee7c8ad7
301e0f608d124ede
This output would appear in a filename h-1000-4, if this particular combination
of parameters were one of the test cases.

7.7.3 Submitting your assignment

Submit your assignment as usual with /c/cs223/bin/submit 7. You should

submit your source file(s), your Makefile, and any other files needed to build your
program other than strategy.h and testStrategy.c, which will be supplied
by the test script. You can test your submission using the public test script
in /c/cs223/Hwk7/test.public using the command /c/cs223/bin/testit 7
public.

7.7.4 Sample solution

I implemented a heap with uint64_t elements as a separate module (heap.h,

heap.c) and then used in in a main module strategy.c that allocates a separate
heap for each pile and manages the translation between strategyDeal and
strategyPlay and the heap functions. The Makefile is pretty much the usual.

7.8 Assignment 8, due Wednesday 2015-04-08, at 11:00pm

7.8.1 An ordered set

For this assignment, you are to implement an ordered set data type for holding
null-terminated strings. The interface to this data type is given in the file
orderedSet.h, shown below.
/*
* Ordered set data structure.
*/

/* Make a new empty set */

498
struct orderedSet *orderedSetCreate(void);

/* Destroy a set */
void orderedSetDestroy(struct orderedSet *);

/* How many elements in this set? */

size_t orderedSetSize(const struct orderedSet *);

/* Insert a new element. Has no effect if element is already present. */

void orderedSetInsert(struct orderedSet *, const char *);

/* Delete an element. Has no effect if element is not already present. */

void orderedSetDelete(struct orderedSet *, const char *);

/* Return a new ordered set containing all elements e

* for which predicate(arg, x) != 0.
* The predicate function should be applied to the elements in increasing order. */
struct orderedSet *orderedSetFilter(const struct orderedSet *, int (*predicate)(void *arg, c
examples/2015/hw/8/orderedSet.h
In addition to the usual create and destroy functions, an ordered set supports
inserting and deleting elements, counting the number of distinct elements in the
set, and filtering the set based on a predicate function passed in as an argument.
This filtering operation does not modify the input set, but instead generates a
new ordered set containing only those elements on which the predicate returns a
nonzero value.
The filtering operation is where most of the excitement happens; because the
predicate function takes an argument of type void * that is also passed to the
filter function, it is possible for the predicate to compute an arbitrary function
on the elements of the set as a side-effect of processing each element to decide
whether to put it in the output set. This allows predicates to be abused to
perform all sorts of computations, including printing out all elements of the set
or computing a hash of all the strings in the set concatenated together. These
features are used by the test program testOrderedSet.c that we have provided.
To ensure that these traversals give consistent results, it is required that when
your implementation of orderedSetFilter is executed, it calls the predicate
function exactly once on each element of the set in increasing order as determined
by strcmp.

7.8.2 The testOrderedSet wrapper

The test program is a fairly thin wrapper over the implementation that allows
you to call the various functions using one-line commands on standard input.
A command is given as the first character of the line, and the rest of the line

499
contains the argument to the command if needed. The + and - commands add
or remove an element from the set, respectively, while the p, s, and h commands
print the contents of the set, the size of the set, and a hash of the set (these
commands ignore any argument). The f command removes all elements of the
set that do not contain a particular substring.
Here is a simple input to the program that inserts four strings, filters out the
ones that don’t contain ee, then prints various information about the results.
+feed
+the
+bees
+please
fee
s
h
p
This should produce the output
2
15082778b3db8cb3
bees
feed

7.8.3 Submitting your assignment

Submit, with the usual /c/cs223/bin/submit 8 filename, your Makefile

and any supporting files needed to build the program testOrderedSet from
testOrderedSet.c and orderedSet.h when make is called with no arguments.
These last two files will be provided by the test script and you do not need to
submit them.
You can test your submission against the public test script in /c/cs223/Hwk8/test.public
with /c/cs223/bin/testit 8.

7.8.4 Sample solution

There were a lot of ways to do this. For the sample solution, I decided to do
something unusual, and store the set as a hash table. This is not ordered, but
since the only operation that requires the set to be ordered is orderedSetFilter,
which will take Ω(n) time no matter how you implement it, the O(n log n) cost
to call qsort to sort the elements as needed does not add much overhead.
#include <stdio.h>
#include <stdlib.h>
#include <assert.h>

500
#include <stdint.h>
#include <string.h>

#include "orderedSet.h"

/* We'll use a hash table with linear probing.

* This is not actually ordered, but the only operations that
* depend on order a linear-time anyway, so we can afford to sort as needed */
struct orderedSet {
size_t n; /* number of elements */
size_t size; /* size of the table */
char **table; /* hash table */
};

#define INITIAL_SIZE (16)

#define MAX_ALPHA (0.75)

/* Make a new empty set with given size */

static struct orderedSet *
orderedSetCreateInternal(size_t size)
{
struct orderedSet *s;

s = malloc(sizeof(*s));
assert(s);

s->n = 0;
s->size = size;
s->table = calloc(s->size, sizeof(char *));

return s;
}

struct orderedSet *
orderedSetCreate(void)
{
return orderedSetCreateInternal(INITIAL_SIZE);
}

/* Destroy a set */
void
orderedSetDestroy(struct orderedSet *s)
{
size_t i;

for(i = 0; i < s->size; i++) {

501
 if(s->table[i]) {
free(s->table[i]);
}
}

free(s->table);
free(s);
}

/* How many elements in this set? */

size_t
orderedSetSize(const struct orderedSet *s)
{
return s->n;
}

static size_t
hash(const char *s)
{
size_t h;

/* usual crummy hash function */

for(h = 0; *s; h = h * 97 + *s++);

return h;
}

static char *
strMalloc(const char *s)
{
char *s2;

s2 = malloc(strlen(s)+1);
strcpy(s2, s);

return s2;
}

/* Insert and element without doing size check or malloc */

/* Frees element if already present */
static void
orderedSetInsertInternal(struct orderedSet *s, char *elt)
{
size_t h;

assert(elt);

502
 /* skip over non-empty slots with different values */
for(h = hash(elt) % s->size; s->table[h] && strcmp(s->table[h], elt); h = (h+1) % s->siz

/* check if not already present */

if(s->table[h] == 0) {
s->table[h] = elt;
s->n++;
} else {
free(elt);
}
}

/* Insert a new element. Has no effect if element is already present. */

void
orderedSetInsert(struct orderedSet *s, const char *elt)
{
size_t h;
struct orderedSet *s2;

if(s->n >= s->size * MAX_ALPHA) {

/* rebuild the table */
s2 = orderedSetCreateInternal(s->size * 2);

/* copy all the elements */

for(h = 0; h < s->size; h++) {
if(s->table[h]) {
orderedSetInsertInternal(s2, s->table[h]);
}
}

/* free the table and then do a brain transplant */

free(s->table);
*s = *s2;
free(s2);
}

orderedSetInsertInternal(s, strMalloc(elt));
}

/* Delete an element. Has no effect if element is not already present. */

void
orderedSetDelete(struct orderedSet *s, const char *elt)
{
size_t h;
char *later;

503
 /* skip over non-empty slots with different values */
for(h = hash(elt) % s->size; s->table[h] && strcmp(s->table[h], elt); h = (h+1) % s->siz

/* if we reached a nonempty slot, it must be our target */

if(s->table[h] != 0) {
/* remove the initial element */
free(s->table[h]);
s->table[h] = 0;
s->n--;

/* remove and reinsert any elements up to the next hole, in case they wanted to be e
for(h = (h+1) % s->size; s->table[h] ; h = (h+1) % s->size) {
later = s->table[h];
s->table[h] = 0;
s->n--;
orderedSetInsertInternal(s, later);
}
}
}

static int
compare(const void *s1, const void *s2)
{
return strcmp(*((const char **) s1), *((const char **) s2));
}

/* Return a new ordered set containing all elements e

* for which predicate(arg, x) != 0.
* The predicate function should be applied to the elements in increasing order. */
struct orderedSet *
orderedSetFilter(const struct orderedSet *s, int (*predicate)(void *arg, const char *), void
{
size_t h;
const char **a; /* temporary array to sort */
size_t top; /* where to put things in a */
size_t i;
struct orderedSet *s2;

a = malloc(sizeof(const char *) * s->size);

assert(a);

top = 0;

for(h = 0; h < s->size; h++) {

if(s->table[h]) {

504
 a[top++] = s->table[h];
}
}

qsort(a, top, sizeof(const char *), compare);

s2 = orderedSetCreate();

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

if(predicate(arg, a[i])) {
orderedSetInsert(s2, a[i]);
}
}

free(a);

return s2;
}
examples/2015/hw/8/orderedSet.c
Makefile{examples/2015/hw/8/Makefile}

7.9 Assignment 9, due Wednesday 2015-04-15, at 11:00pm

7.9.1 Finding a cycle in a maze

For this problem, you are given a rectangular maze consisting of wall squares
(represented by 0) and path squares (represented by 1). Two path squares are
considered to be adjacent if they are at most one square away orthogonally or
diagonally; in chess terms, two path squares are adjacent if a king can move
from one to the other in one turn. The input to your program is a maze in which
the graph consisting of all path squares is connected and contains at most one
cycle, where a cycle is a sequence of distinct squares s1 , s2 , . . . , sk where each si
is adjacent to si+1 and sn is adjacent to s1 . Your job is to write a program maze
that finds this cycle if it exists, and marks all of its squares as cycle squares
(represented by 2).
For example, here is a picture of a 200-by-100 maze that contains a small cycle:
and here is the same maze with the cycle highlighted in white:

7.9.2 Input and output format

The input to your program should be taken from stdin, in a restricted version
of raw PGM format, an old image format designed to be particularly easy to

505
 Figure 9: 200 by 100 maze

Figure 10: 200 by 100 maze, showing cycle

506
parse. The input file header will be a line that looks like it was generated by the
printf conversion string "P5 %d %d 255\n", where the first int value is the
width of the image in columns and the second is the height of the image in rows;
the same conversion string can be given to scanf to parse this line. Following
the newline will be a long sequence of bytes, each representing one pixel of the
image, with each row following immediately after the previous one. These bytes
will be either 0 or 1 depending on whether that position in the maze is a wall or
a path.
The output to your program should be in the same format, with the difference
that now some of the bytes in the image data may be 2, indicating the cycle.
If there is no cycle, the output should be identical to the input. Your program
is not required to detect or respond in any particular way to input mazes that
violate the format or do not contain a connected graph of path squares, although
you are encouraged to put in reasonable error checking for your own benefit
during testing.
For example, the maze depicted above is stored in the file 200-100-4.in.pgm;
the corresponding output is stored in the file 200-100-4.out.pgm. Other sample
inputs and outputs can be found in /c/cs223/Hwk9/testFiles.
This file format is hard to read with the naked eye, even after loading into a text
editor. The script /c/cs223/Hwk9/toPng will generate a PNG file that doubles
the pixel size and rescales the 0, 1, 2 pixel values to more reasonable values for
display. This can be called as /c/cs223/Hwk9/toPng filename.pgm to produce
a new file filename.pgm.png. This works best if filename.pgm is already in a
directory you can write to. PNG files can be displayed using most web browsers
and image manipulation tools.

7.9.3 Submitting and testing your program

Submit whatever files you need to build maze (including a Makefile that gen-
erates maze when called with no arguments) using /c/cs223/bin/submit 9.
You can apply the public test script in /c/cs223/Hwk9/test.public to your
submitted files using /c/cs223/bin/testit 9 public.

7.9.4 Sample solution

This uses breadth-first search, which makes the search a bit simpler than depth-
first search but requires some more effort to compute the cycle. The program
also includes code for generating random mazes.
#include <stdio.h>
#include <stdlib.h>
#include <assert.h>
#include <math.h>

507
#include <limits.h>

struct direction {
signed char x;
signed char y;
};

#define DIRECTIONS (8)

const struct direction directions[DIRECTIONS] = {

{ -1, -1 },
{ -1, 0 },
{ -1, 1 },
{ 0, -1 },
{ 0, 1 },
{ 1, -1 },
{ 1, 0 },
{ 1, 1 }
};

struct position {
int x;
int y;
};

const struct position NO_POSITION = { -1, -1 };

static inline int

eqPosition(struct position p, struct position q)
{
return p.x == q.x && p.y == q.y;
}

#define WALL (0)

#define PATH (1)
#define CYCLE (2)

struct square {
int contents;
struct position parent; /* used by search routine */
};

struct maze {
struct position size; /* rows = size.x, columns = size.y */
struct square *a; /* packed array of squares */
};

508
/* look up a position in a maze */
#define Mref(m, pos) ((m)->a[(pos).y * (m)->size.x + (pos).x])
#define Mget(m, pos) (assert((pos).x >= 0 && (pos).y >= 0 && (pos).x < (m)->size.x && (pos).

/* add direction to source to get target */

/* returns 1 if target is in range */
int
offset(const struct maze *m, struct position *target, struct position source, struct directi
{
target->x = source.x + dir.x;
target->y = source.y + dir.y;

return target->x >= 0 && target->y >= 0 && target->x < m->size.x && target->y < m->size.
}

/* free a maze */
void
destroyMaze(struct maze *m)
{
free(m->a);
free(m);
}

/* load a maze in restricted PGM format */

struct maze *
loadMaze(FILE *f)
{
struct maze *m;
struct position i;

m = malloc(sizeof(*m));
assert(m);

fscanf(f, "P5 %d %d 255\n", &m->size.x, &m->size.y);

m->a = malloc(sizeof(struct square) * m->size.y * m->size.x);

for(i.y = 0; i.y < m->size.y; i.y++) {

for(i.x = 0; i.x < m->size.x; i.x++) {
Mref(m, i).contents = getchar();
assert(Mref(m, i).contents == 0 || Mref(m, i).contents == 1);
}
}

return m;

509
}

void
saveMaze(struct maze *m, FILE *f)
{
struct position i;

fprintf(f, "P5 %d %d 255\n", m->size.x, m->size.y);

for(i.y = 0; i.y < m->size.y; i.y++) {

for(i.x = 0; i.x < m->size.x; i.x++) {
putc(Mref(m, i).contents, f);
}
}
}

/* how many neighbors of position are PATH? */

int
countNeighbors(const struct maze *m, struct position p)
{
struct position q;
int i;
int count = 0;

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

if(offset(m, &q, p, directions[i]) && Mget(m, q).contents == PATH) {
count++;
}
}

return count;
}

struct position
randomPosition(const struct maze *m)
{
struct position r;

r.x = rand() % m->size.x;

r.y = rand() % m->size.y;

return r;
}

#define PATIENCE_MULTIPLIER (4)

510
/* generate a random connected maze with no cycles */
struct maze *
generateMaze(struct position size)
{
struct maze *m;
struct position r;
struct position i;
size_t countdown; /* how long to run before we get tired of not making progress */
size_t maxCountdown; /* value to reset countdown to when we make progress */

m = malloc(sizeof(struct maze));
assert(m);

m->size = size;
m->a = malloc(sizeof(struct square) * m->size.x * m->size.y);
assert(m->a);

/* start with all WALL */

for(i.y = 0; i.y < m->size.y; i.y++) {
for(i.x = 0; i.x < m->size.x; i.x++) {
Mref(m, i).contents = WALL;
}
}

/* place a PATH on a random square */

r = randomPosition(m);
Mref(m, r).contents = PATH;

maxCountdown = PATIENCE_MULTIPLIER * size.x * size.y * log(size.x * size.y);

for(countdown = maxCountdown; countdown > 0; countdown--) {

/* pick a random square */
r = randomPosition(m);

/* add if we have exactly one neighbor already in the maze */

if(Mget(m, r).contents == WALL && countNeighbors(m, r) == 1) {
Mref(m, r).contents = PATH;

/* reset countdown */
countdown = maxCountdown;
}
}

return m;
}

511
/* create a cycle by adding one extra PATH square
* that connects two existing squares */
void
mazeAddCycle(struct maze *m)
{
struct position r;

do {
r = randomPosition(m);
} while(Mget(m, r).contents != WALL || countNeighbors(m, r) != 2);

Mref(m, r).contents = PATH;

}

/* Search for a cycle of PATH nodes.

* If found, mark all nodes on the cycle as CYCLE. */
void
mazeSearchForCycle(struct maze *m)
{
struct position root; /* root of tree */
struct position current; /* what we just popped */
struct position parent ; /* current's parent */
struct position neighbor; /* neighbor to push */
struct position ancestor; /* for filling in CYCLE */
int i;
struct position *queue;
size_t head; /* where to dequeue */
size_t tail; /* where to enqueue */

/* this is probably more space than we need */

queue = malloc(sizeof(struct position) * m->size.x * m->size.y);
assert(queue);

head = tail = 0;

/* clear out bookkeeping data */

for(current.y = 0; current.y < m->size.y; current.y++) {
for(current.x = 0; current.x < m->size.x; current.x++) {
Mref(m, current).parent = NO_POSITION;

/* probably not necessary but will avoid trouble

* if somebody calls this twice */
if(Mget(m, current).contents != WALL) {
Mref(m, current).contents = PATH;
}
}

512
}

/* find a root */
/* we don't care what this is, but it can't be a WALL */
do {
root = randomPosition(m);
} while(Mget(m, root).contents != PATH);

/* push root */
Mref(m, root).parent = root;
queue[tail++] = root;

/* now perform the BFS */

/* if we ever find a neighbor that is already in the tree and not our parent,
* we have found our cycle */
while(head < tail) {
current = queue[head++];
parent = Mget(m, current).parent;

/* push all neighbors not already in tree */

/* if one is in the tree, we win */
for(i = 0; i < DIRECTIONS; i++) {
if(offset(m, &neighbor, current, directions[i]) && Mget(m, neighbor).contents ==
/* is it already in the tree? */
if(!eqPosition(Mget(m, neighbor).parent, NO_POSITION)) {
/* we win */
/* cycle consists of all ancestors of neighbor and current
* up to common ancestor */
for(ancestor = neighbor; !eqPosition(ancestor, root); ancestor = Mget(m,
Mref(m, ancestor).contents = CYCLE;
}

/* also mark root */

Mref(m, root).contents = CYCLE;

/* now work up from current */

for(ancestor = current; !eqPosition(ancestor, root); ancestor = Mget(m,
if(Mget(m, ancestor).contents == PATH) {
/* add to the cycle */
Mref(m, ancestor).contents = CYCLE;
} else {
/* this is the common ancestor, which is not root */
/* mark all proper ancestors as PATH */
do {
ancestor = Mget(m, ancestor).parent;
Mref(m, ancestor).contents = PATH;

513
 } while(!eqPosition(ancestor, root));

/* can't just break, too many loops */

goto doneWithSearch;
}
}
} else {
Mref(m, neighbor).parent = current;
queue[tail++] = neighbor;
}
}
}
}

doneWithSearch:
free(queue);
}

int
main(int argc, char **argv)
{
struct maze *m;
struct position size = { 80, 60 };
int seed;

switch(argc) {
case 1:
/* sample solution for the assignment */
m = loadMaze(stdin);
mazeSearchForCycle(m);
saveMaze(m, stdout);
destroyMaze(m);
break;
case 4:
/* generate a new test image */
/* usage is ./maze width height seed */
/* if seed is negative, use absolute value and don't put in cycle */
size.x = atoi(argv[1]);
size.y = atoi(argv[2]);
seed = atoi(argv[3]);

srand(seed < 0 ? -seed : seed);

m = generateMaze(size);
if(seed >= 0) { mazeAddCycle(m); }
saveMaze(m, stdout);
destroyMaze(m);

514
 break;
default:
fprintf(stderr, "Usage %s or %s width height seed\n", argv[0], argv[0]);
return 1;
}

return 0;
}
examples/2015/hw/9/maze.c
And the Makefile.

8 Common C coding and debugging issues

Here are some notes from a helpful Zoo denizen about debugging programs in
223. (Note: these have been edited slightly from the original.)
Date: Thu, 10 Feb 2005 06:02:23 -0500 (EST)
From: James Terry <[email protected]>
Subject: 223 coding feedback

Hi Jim,

Several of your students for 223 were up late last night in the Zoo
working on their assignments, and they seemed to be getting hung up on
some coding issues. They were pretty frustrated with some standard
language/debugging issues, so I helped them get the type-checker and
Valgrind to stop yelling at them. I noticed some recurring problems and I
thought I'd pass them on to you. They're pretty standard mistakes, and
I've made most of them myself at some point, either in your class or in
Stan's. It occurred to me that there might be more confused people than
were around last night, and they'd probably appreciate it if someone told
them about these sort of things. I'm not trying to intrude on how you
teach the class; I just thought this feedback would be helpful and I
wasn't sure that it would find its way to you otherwise. I'm sure you've
already taught them several of these, and I understand that sometimes
students just don't pay attention. Still, these seem like good points to
hammer down:

Recurring debugging/coding problems:

1. If you want a debugger/Valgrind to give you line numbers, you must

compile with debugging info turned on, i. e. using the -g[level] flag.
2. On the Zoo, pointers and int's are 4 bytes; char's are 1. (Some

515
people didn't seem to realize that a char* is 4 bytes rather than 1.)
3. I think it would be helpful if you explained why, when using
realloc(), it's a good idea to increase the allocated size
multiplicatively rather than additively. Besides, everyone loves the
"tearing down the hotel" metaphor. :)
4. If they use call-by-reference, they had better make sure that they
keep the same reference. So if they pass in a pointer as an argument to a
function, they shouldn't call malloc() or realloc() on that function.
(Mention the double pointer as another option.) Most people will make
this mistake eventually if no one warns them about it. When I was
learning C, I sort of viewed malloc() and realloc() as magical
memory-increasing functions; that is to say, I didn't think very hard
about the meaning of assigning a pointer to malloc()'s return value. I
suspect some of your students would benefit from having the details
spelled out. (Or spelled out again, if you've already done that.)
5. It's possible to get through a lot (but not all) of the CS major
without learning basic Unix shell syntax, but that's really just wasted
time. Pipes, backgrounding, man, scp, and grep really help even at the
intro level. I realize the purpose of the class isn't to teach Unix, but
in past years I think there was a TA help session on these things. They
don't need to know how to write their own Emacs modes, but the basics
would definitely be helpful.
6. malloc/free -- If Valgrind/gdb reports a problem inside of malloc() or
free(), chances are that the student has *not* discovered a bug in gcc.
(I just heard how one of Zhong's students' proved the correctness of the
libraries for his thesis; that's pretty cool.) Explain why you can't
malloc() twice on the same pointer. Explain how with multidimensional
pointers, you must malloc/free each dimension separately. Drill down the
one-to-one correspondence between malloc'ing and free'ing.
7. Null characters: It's not obvious to newbies that some library functions
require them, particularly null-terminated strings. Tell them that
char*'s must be null terminated in order for <string.h>
functions to work.
8. Off-by-one errors: Tell people that when all else fails, take a hard
look at their comparison operators; i. e. make sure that > shouldn't
really be a >=.
9. This is probably another thing for a help session or workshop, but I
feel almost everyone could benefit from basic software engineering
methodology. Stylistic awkwardness I noticed:
--Using a mess of if-then-else's instead of nested control
structures.
--Using while-loops with iterators that get initialized right
before the beginning of the loop and get incremented with each iteration,
when they could be using for-loops.
--Doing the setup work for a loop right before the beginning of
the loop and then at the end of every iteration, instead of at the

516
beginning of every iteration. Conversely: doing the cleanup work at the
beginning of every iteration and then after the loop has completed.
10. Tell them to use assert(). (Frequently.) When you cover binary
search, using placement of debugging statements in code in order to pin
down an error might be an instructive example.
11. Tell them to use either printf statements or a debugger to debug. I
think they can figure out how to do this on their own, they just need to
be told it's a good idea.

Hopefully some of these suggestions will be helpful. I won't be offended

if you don't pass them on to your students, and I understand if you put a
higher teaching priority on non-coding subjects, but all the things I
mentioned were things I wish someone had told me. I'd be happy to run a
help session on this stuff if you and the TAs are too busy, but otherwise
I wouldn't presume.

Best,
Jim

517