want to modify a vertex. Therefore, the insight at that time was that you can combine high order techniques when building generic algorithms with disciplined use of side effects. Side effects are not necessarily bad; they are bad only when they are misused.
In the summer of 1985 I was invited back to GE Research to teach a course on high order programming. I demonstrated how you can construct complex algorithms using this technique. One of the people who attended was Art Chen, then the manager of the Information Systems Laboratory. He was sufficiently impressed to ask me if I could produce an industrial strength library using these techniques in Ada, provided that I would get support. Being a poor assistant professor, I said yes, even though I didn't know any Ada at the time. I collaborated with Dave Musser in building this Ada library. It was an important undertaking, because switching from a dynamically typed language, such as Scheme, to a strongly typed language, such as Ada, allowed me to realize the importance of strong typing. Everybody realizes that strong typing helps in catching errors. I discovered that strong typing, in the context of Ada generics, was also an instrument of capturing designs. It was not just a tool to catch bugs. It was also a tool to think. That work led to the idea of orthogonal decomposition of a component space. I realized that software components belong to different categories. Object-oriented programming aficionados think that everything is an object. When I was working on the Ada generic library, I realized that this wasn't so. There are things that are objects. Things that have state and change their state are objects. And then there are things that are not objects. A binary search is not an object. It is an algorithm. Moreover, I realized that by decomposing the component space into several orthogonal dimensions, we can reduce the number of components, and, more importantly, we can provide a conceptual framework of how to design things.
Then I was offered a job at Bell Laboratories working in the C++ group on C++ libraries. They asked me whether I could do it in C++. Of course, I didn't know C++ and, of course, I said I could. But I couldn't do it in C++, because in 1987 C++ didn't have templates, which are essential for enabling this style of programming. Inheritance was the only mechanism to obtain genericity and it was not sufficient.
Even now C++ inheritance is not of much use for generic programming. Let's discuss why. Many people have attempted to use inheritance to implement data structures and container classes. As we know now, there were few if any successful attempts. C++ inheritance, and the programming style associated with it are dramatically limited. It is impossible to implement a design which includes as trivial a thing as equality using it. If you start with a base class X at the root of your hierarchy and define a virtual equality operator on this class which takes an argument of the type X, then derive class Y from class X. What is the interface of the equality? It has equality which compares Y with X. Using animals as an example (OO people love animals), define mammal and derive giraffe from mammal. Then define a member function mate, where animal mates with animal and returns an animal. Then you derive giraffe from animal and, of course, it has a function mate where giraffe mates with animal and returns an animal. It's definitely not what you want. While mating may not be very important for C++ programmers, equality is. I do not know a single algorithm where equality of some kind is not used.
You need templates to deal with such problems. You can have template class animal which has member function mate which takes animal and returns animal. When you instantiate giraffe, mate will do the right thing. The template is a more powerful mechanism in that respect.
However, I was able to build a rather large library of algorithms, which later became part of the Unix System Laboratory Standard Component Library. I learned a lot at Bell Labs by talking to people like Andy Koenig and Bjarne Stroustrup about programming. I realized that C/C++ is an important programming language with some fundamental paradigms that cannot be ignored. In particular I learned that pointers are very good. I don't mean dangling pointers. I don't mean pointers to the stack. But I mean that the general notion of pointer is a powerful tool. The notion of address is universally used. It is incorrectly believed that pointers make our thinking sequential. That is not so. Without some kind of address we cannot describe any parallel algorithm. If you attempt to describe an addition of n numbers in parallel, you cannot do it unless you can talk about the first number being added to the second number, while the third number is added to the fourth number. You need some kind of indexing. You need some kind of address to describe any kind of algorithm, sequential or parallel. The notion of an address or a location is fundamental in our conceptualizing computational processes---algorithms.
Let's consider now why C is a great language. It is commonly believed that C is a hack which was successful because Unix was written in it. I disagree. Over a long period of time computer architectures evolved, not because of some clever people figuring how to evolve architectures---as a matter of fact, clever people were pushing tagged architectures during that period of time---but because of the demands of different programmers to solve real problems. Computers that were able to deal just with numbers evolved into computers with byte- addressable memory, flat address spaces, and pointers. This was a natural evolution reflecting the growing set of problems that people were solving. C, reflecting the genius of Dennis Ritchie, provided a minimal model of the computer that had evolved over 30 years. C was not a quick hack. As computers evolved to handle all kinds of problems, C, being the minimal model of such a computer, became a very powerful language to solve all kinds of problems in different domains very effectively. This is the secret of C's portability: it is the best representation of an abstract computer that we have. Of course, the abstraction is done over the set of real computers, not some imaginary computational devices. Moreover, people could understand the machine model behind C. It is much easier for an average engineer to understand the machine model behind C than the machine model behind Ada or even Scheme. C succeeded because it was doing the right thing, not because of AT&T promoting it or Unix being written with it.
C++ is successful because instead of trying to come up with some machine model invented by just contemplating one's navel, Bjarne started with C and tried to evolve C further, allowing more general programming techniques but within the framework of this machine model. The machine model of C is very simple. You have the memory where things reside. You have pointers to the consecutive elements of the memory. It's very easy to understand. C++ keeps this model, but makes things that reside in the memory more extensive than in the C machine, because C has a limited set of data types. It has structures that allow a sort of an extensible type system, but it does not allow you to define operations on structures. This limits the extensibility of the type system. C++ moved C's machine model much further toward a truly extensible type system.
In 1988 I moved to HP Labs where I was hired to work on generic libraries. For several years, instead of doing that I worked on disk drives, which was exciting but was totally orthogonal to this area of research. I returned to generic library development in 1992 when Bill Worley, who was my lab director established an algorithms project with me being its manager. C++ had templates by then. I discovered that Bjarne had done a marvelous job at designing templates. I had participated in several discussions early on at Bell Labs about designing templates and argued rather violently with Bjarne that he should make C++ templates as close to Ada generics as possible. I think that I argued so violently that he decided against that. I realized the importance of having template functions in C++ and not just template classes, as some people believed. I thought, however, that template functions should work like Ada generics, that is, that they should be explicitly instantiated. Bjarne did not listen to me and he designed a template function mechanism where templates are instantiated implicitly using an overloading mechanism. This particular technique became crucial for my work because I discovered that it allowed me to do many things that were not possible in Ada. I view this particular design by Bjarne as a marvelous piece of work and I'm very happy that he didn't follow my advice.
When did you first conceive of the STL and what was its original purpose?
In 1992, when the project was formed, there were eight people in it. Gradually the group diminished, eventually becoming two people, me and Meng Lee. While Meng was new to the area---she was doing compilers for most of her professional life---she accepted the overall vision of generic programming research, and believed that it could lead to changing software development at the point when very few people shared this belief. I do not think that I would be able to build STL without her help. (After all, STL stands for Stepanov and Lee...) We wrote a huge library, a lot of code with a lot of data structures and algorithms, function objects, adaptors, and so on. There was a lot of code, but no documentation. Our work was viewed as a research project with the goal of demonstrating that you can have algorithms defined as generically as possible and still extremely efficient. We spent a lot of time taking measurements, and we found that we can make these algorithms as generic as they can be, and still be as efficient as hand-written code. There is no performance penalty for this style of programming! The library was growing, but it wasn't clear where it was heading as a project. It took several fortunate events to lead it toward STL.
When and why did you decide to propose STL as part of the ANSI/ISO Standard C++ definition?
