Dr. Carlo Pescio
Interview with Bjarne Stroustrup


This is the original English text of an interview with Bjarne Stroustrup, later translated in Italian and published in Computer Programming No 50, September 1996. Follow this link for the translated version. This article has also been translated to Serbo-Croatian by Web Geeks.

In the following text, CP stands for Carlo Pescio and BS for Bjarne Stroustrup.

CP: Are you satisfied by the results and pace of the C++ standardization effort? I mean, on the one hand, it is a proof of the success of the language, but on the other hand it is adding a lot of bureaucracy; do you miss the initial freedom you had on core language issues?

BS: Naturally, my patience has been sorely tested but by and large, I'm happy with the outcome of the C++ standardization effort. With one minor exceptions, ISO will have the features I feel it really needs and none that I find harmful. ISO C++ is a much more powerful and coherent language than the early versions of C++ was - and no major features was added that I didn't work on and approve of. Some details of ISO C++ may show signs of "design by committee," but the overall set of facilities is a better match for my original view of what C++ should be than earlier versions of the language.

The primary reason I took part of the standards effort was that the committee was convened before I had been able to complete the language. C++ without templates and exceptions would have been unacceptable, and C++ without namespaces and run-time type information would have left us much poorer than we are today.

It is easy to exaggerate my "initial freedom." I decided early on that I wanted my language to be compatible with C and that I wanted to support real users. This placed constraints on what I could do. I think the alternative would have been to design yet another cult language. I can't wait for the standard to become complete. The C++ community needs a solid standard and I need the time I have had to devote to the standards effort over the last six years or so. I think the language has benefitted significantly from my efforts in the standards committee. I was the chair of the extensions sub-committee that processed all requests for extensions and other major changes to the language and was generally involved in most major issues - including the design of the standard library. Standardization is necessary for a language with C++'s number of users. However, I wouldn't call standardization fun. Much credit - much more credit than is usually accorded to "faceless committees" - should go to the dozens of people who volunteered their time and very considerable efforts to the standard year in and year out. I documented many of their names and efforts in my recent book "The Design and Evolution of C++" (usually called D&E).

CP: Lot of readers would like to know if there is a "the c++ language - 3rd edition" and/or a 2nd edition of the ARM under way...

BS: I'm working on a 3rd edition and thinking of something to replace the ARM. However, I don't see a point in writing a new book unless I have a lot of new things to say, so progress is slow. My aim is to make the 3rd edition as much an improvement over my 2nd as the 2nd is over the 1st. In other words, I want to produce something even more approachable that the 2nd yet containing information that will be new and exhiting to essentially every C++ programmer. I cannot predict when I complete, but noone who needs a good textbook should wait for the 3rd edition. The 2nd edition is still one of the most complete, thorough, and up-to-date textbooks available. Naturally, I cannot publish an ARM replacement until there is a standard to annotate. Until I get those books written, D&E is the best source for information about new features in particular and the reasons behind the design of C++ in general.


CP: Are there any technical decisions that in retrospective you'd like to change in C++ - not something you had no way to get accepted, but something that further experience proved to be somehow flawed, and that you'd like to change if you had the opportunity to (even if today is not possible for compatibility reasons).

BS: There are lots of details, I'd like to tinker with, but there are no major features I would like to delete (even if I could) or major features that I both like and know how to add. When people ask this kind of question, they usual have some feature such as MI or RTTI in mind, so let me say that C++ without one of those would be a poorer language. I use both extensively and the workarounds I would have to use if they weren't there are not pretty sights. There are of course examples and explanations in D&E.


CP: Actually I mostly had [changes to] virtual functions in mind: for instance, I'm not very fond on the fact that a virtual function can be redefined in a *privately* derived class. The function is not accessible - yet is redefinable; this conjure a bit on the usefulness and safety of private inheritance.

To a larger extent, you probably remember Sakkinen's papers on the "inheritance principles of C++": he made some fine points, and I particularly liked the fact that [under more restrictive rules] one would not have to extend the liability of invoking the constructor of a virtual base class farther than "necessary" in the inheritance graph. In fact, while I understand the reasons behind the current rules for VBC in C++, I have to admit that in some sense, invoking the constructor of a VBC weakens the encapsulation provided by intermediate classes. On the other hand, we know the strong and weak points of C++ through years of use - and probably Sakkinen's proposals would have lead just to other weak points, although I sort of like them as far as "programming style" is concerned.

BS: Fundamentally, I consider these these issues purely academic and of no interest to practical programming. It is possibly to design a constructor to maliciously access information, but a constructor written to initialize its own data is fundamentally harmless. Furthermore, if the virtual base class is public - as I always recommend - all classes are supposed to know it exists and should expect its constructor to be invoked. In this, virtual base classes doesn't differ from other bases. If the virtual base is declared private somewhere and public elsewhere and if the most derived class initializes it in a way that surprises the writer of the private virtual base, I guess "the encapsulation provided by intermediate classes" has been weakened. However, if that is someone's worst problem that someone must be truly blessed.

I see no problem with overriding a virtual function by a private function or from a base in which the base is private. If you consider a base class an interface, it is of no concern of the users of that interface how its implementors (the derived classes) provide their implementations. I find it easy to imagine a derived class that makes its overriding functions private exactly to prevent use of the derived class except through the intented base class interface and to inhibit further derivation. If a base is private, it may still be accessible to friends, or the derived class may be handing out pointers to its (private) base on request. For example, a derived class could return its base class interface as the result of an operation that performed system-level acccess control checks.

class A
  {
  virtual void f() ;
  } ;

class B : private A
  {
  void f(); // implementation 
  A* get_A(Rights& r) { /* check rights */ return (*A)this; }
  } ;

CP: On the other hand, private inheritance is not transitive, so in a case like:

class A
  {
  virtual void f() ;
  } ; 

class B : private A
  {
  void h() { f() ; }
  } ; 

class C : public B
  {
  virtual void f() ;
  } 

class A is not an interface for class C, A :: f() is not accessible in C, but is redefinable in class C. We could just say that the culprit is class C implementor, who based his code on an implementation detail of class B (the fact that class B is-implemented-using class A). Some may like to have that forbidden from the language...

BS: Yes, that's a messy example. However, it is not easy to craft a set of rules that outlaws every messy example without also doing damage by outlawing examples that some consider messy, yet others deem essential for their work. In general, I'm against restrictions for which I don't see a practical purpose. I do not consider orthogonality a primary design aim, but certainly I prefer it whenever there are no first-order reasons for non-ortogonality. The C++ access rules are very orthogonal (with naming rules, overriding rules, etc.) and I see no strong reason for them not to be. People can be surprised by the rules, but so they could by less orthogonal rules.

CP: I understand your opinion and I agree to a large extent. That's why I said this kind of "rule" may be good as far as "programming style" is concerned. There are certainly times when something like that may be useful (for instance in reusing a library without source code) and in that case one may "violate the rule" and do it, because C++ is not a motherly, restrictive language.

BS: I would have liked C++ to be more capable of catching logical errors made by programmers. After all, many of the C++ features has exactly that effect when compared to what you'd have to write in C to perform an equivalent action. However, I do not think that safety should be bought at the cost of complicating the expression of good solutions to real-life problems. This - and compatibility concerns - limits what can be done to make the language itself cleaner. I recommend the use libraries to make use of C++ safer for specific applications. For example, if someone worry about the fact that C arrays aren't range checked, they should use a range-checked vector class. I do that most of the time myself, especially for debugging.

CP: Let's go back to the retrospective on C++...

BS: For good and bad, I have maintained a high degrees of C compatibility since the earliest days of C++ and the standards committee has followed my ``as close to C as possible - but no closer'' policy. Many things in C++ could be better from a language-technical point of view, but that is not realistic. When I started I decided on compatibility and tried the best I could to live with second-order flaws, fixing only problems relating to the type system. The alternative would have been to build a yet another cult language: Beautiful in the eyes of its adherents and eliciting nothing more than ayawn from almost all who program to get results. Had C not been there to be compatible with, I would have chosen some other language else to be compatible with. I was - and am - convinced that my time would not have been well spent inventing yet another way of writing a loop.

Concurrency is an issue that keeps coming up. Most people like some form of concurrency and would like to see it directly supported in C++. However, there is no one form of concurrency that serves more than a small fraction of the C++ programmers who need concurrency really well. OS writers needs one form of concurrency, database users another, and network application implementers yet another. Thus, I decided not to include specific features supporting concurrency in C++. People who wants some form of concurrency can - and does - support it through language extensions of (preferably) libraries. The standard committee backed me on that issue also; we knew many attractive concurrency schemens, but not one that could serve a large majority of C++ users. The range of areas where C++ is used it truly amazing.

I was asked to write an article on C++ for ACM's second conference on the history of programming languages (they have one every 15 years!). For that paper, I was asked what I considered my greatest mistake with C++. In my mind there was just one candidate for the title of "worst mistake:" I failed to produce an acceptable foundation library and ship it with release 1.0 in 1985. My excuse was that I didn't know how to write a good enough library and that I needed templates to provide efficient, flexible, and type-safe containers. The result of that mistake is the current mess of of incompatible foundation libraries of widely differing philosophies and qualities.

Fortunately, the standards committee was able to agree on an excellent standard library. We now have what I failed to produce and something that I did not know how to design and implement ten years ago. The framework of containers and fundamental algorithms that is a major part of the standard library was primarily the work of Alex Stepanov. It actually has some roots in our search for principles and techniques for good library components back in the years before and after release 1.0, so the delay wasn't completely wasted.

CP: As you know there are proposals to add some of the C++ features to C9x, for instance a limited support for classes. I had the impression of a naive approach, e.g. they don't have constructors, assignment overloading, and so on. What's your position on the "new ANSI C"?

BS: The features did look naive from my perspective and the explanations that acompanied them seems to reflect a lack of appreciation of how C++ was evolved in response to feedback. I had (and still have) an overall view of what the language is supposed to do, but within that overall view, I was careful to seek feedback on the language as it evolved to adjust its features to real experience and needs.

Attempting to pick a small subset of features from C++ to provide "a much simpler language with almost all of the power of C++" is in my opinion doomed to failure. Unless very carefully chosen based on real experience, the features will only partially support coherent programming styles and one feature will lead to another - only when guided by an overall view of what design and programming styles are to be directly supported will the additions lead to a coherent language.

It would not be reasonable for me to try to tell the C community how they should standardize their language. If I did, my advice would probably not be wellcome anyway because the people who would value my advice most would be programming in C++ anyway. The C committee will do what it wants to with C and presumably they know best for their core community. any extensions as compatible with C++ as at all possible. The programming community doesn't need yet another incompatible C dialect. K&R C isn't dead yet, current ANSI C will persist long after an new standard is promulgated, and various vintages of C++ will also exists well after the C++ standard becomes official. Things always takes longer than we think and that we would like. New features in C9x will necessarily become an extra source of instability in the C and C++ community.

CP: Talking of C++'s sons, I cannot avoid to mention Java... how big in your opinion is the bandwagon effect, and were (if anywhere) you see Java's real strengths. I know that as a language designer, you are very careful on criticizing languages - sort of "any language has a niche" - but what are your gut feelings?

BS: Java has a C++-like syntax, but is a fundamentally different language supporting a different culture and a different (rather narrow) range of programming styles. Java certainly isn't the C++-like language I would have designed in the absence of compatibility constraints. Currently Java rides amazingly high on expectation, Sun marketing money, and its integration with the Web. Time will tell how it will fare as a general-purpose language, and how the majority of programmers and managers will react when they discover that the "security" of Java and in particular of Javascript leaves much to be desired. People confuse programming language type safety (which a correct Java implementation provides) with security; that is, the preservation of system integrity and privacy (which can be seriously compromised using Java). Our security people jokingly refers to Java as "the virus implementation language" and strongly recommend that we run with Java and Javascript disabled in our netbrowsers. Back to programming language issues:

C++

- is a better C

- supports data abstraction

- supports object-oriented programming

- supports generic programming

Of these, Java covers the object-oriented part only, and it does so in ways that differ significantly from C++.

CP: Some of the newest C++ additions have been useful but small details - explicit, mutable - or "escape from C" features like the new-style casts. Do you see any future for more design-oriented features in the language? For instance, maybe there are some valuable inspirations in projects like Annotated C++ or Larch-C++, leaning more toward specifications and less at the coding level, or some ways to add some more semantics - for instance make object-sharing decisions more explicit. Or is your intention to strengthen the language were is already strong - in the embedded / system programming / performance-critical areas?

BS: C++ is part of a general drift towards a more declarative styles of programming. However, as far at the official definition of the language is concerned changes to the language and standard libray now has to stop to give implementers, users, tool builders, teachers, etc. a chance to catch up. Naturally, experimentation will continue (though most likely not by me), but in my opinion the C++ developer community needs stability more than anything else. C++ is now complete and coherent at what it does; further work will have to stay in sub-communities (such as academia) for a few years now.

I think most people yet have to appreciate the strenghts of the template mechanism for various forms of specification. For example, here is the outline of how one can define a list so that a single implementation is shared by all lists of pointers:

// general list<T>:

template<class T> class list { /* ... */ };

// specialization for lists of void*:

template<> class list<void*> { /* ... */ };

// general list of pointers (implemented using list<void*>):

template<class T> class list<T*> : list<void*> { /* ... */ };

The specialization mechanism used here allows different implemetations to be chosen (using type deduction) while still providing the user with a single general interface. One important aspect of this is that it strengthens the declarative nature of C++ programming while simplifying the user's interface and improving run-time efficency. Such techniques in the standard library allows us to provide a single general sort() routine that for real examples have outperformed to C standard library qsort() by a factor of seven!

CP: In IEEE Computer, February 95, Prof. Wirth labeled C++ as "a language that discourages structured thinking and disciplined programming construction". I cannot say I agree, or that I find Oberon encouraging more structure and discipline than C++, but is there anything you'd like to concede to the purists/academics out there, that have to decide between teaching C++ because is useful in real world, and not teaching it because it is too distant from the formal, specification-based approach often used in CS teaching... so they end up teaching Eiffel because of all the "programming by contract" promotion or Smalltalk because "is pure OO"...

BS: Professor Wirth isn't known for being generous with praise for languages he hasn't designed himself, so I cannot say I'm surprised by his eveluation. On the other hand, I think he is flat wrong. C++ is a more-than-adequate tool for good design, for industrial scale programming, and for precise thinking about serious problems. I guess this would be a good place to express my gratitude to the designers of Simula and C for providing a solid base for me to build C++ on and for being such genuine nice people. I also learned a fair bit from many other languages. If you know where to look, you can find traces of Algol68, ML, Ada, and BCPL in C++. There are a lot of great languages around. Everyone should aim at being proficient in more than one language - that holds for both programming languages and natural languages. Another language adds significantly to one's world view and abilities. There are lots that could be better in C++. However, that is true for every language in real use. Even those that claim to be "pure." In my experience the problems with C++ are not serious for teaching nor for real use. Naturally, a student can fail to learn and a teacher can take an approach that makes learning unnecessarily painful. However that can and does happen in every language. C++ has the advantage that its use scales to real world problems in many diverse application areas. Much of the ease of learning cleaner/newer languages comes by simplifications that force its users to abandon the language when they hit an application outside the domain where the "clean" language is a reasonable choice. Naturally, this can happen to C++ users also, but only rarely in any field that somehow touches upon systems programming. C++ has clean subsets and the complexity comes when people starts playing with features and programming styles (``paradigms'') that require more extensive understanding. This is where users of "cleaner" languages often has to resort to alternative, lower-level, and "unclean" languages - usually C or C++. In my opinion, C++ should be taught in stages and with a strong emphasis on concepts.

CP: Good point, really. Still, one might imagine an even cleaner subset, for instance a "student-C++" where an array is not cast to a pointer without the user asking for that, and some more limiting factors on the same spirit. Do you think it could be a useful teaching tool (and perhaps a production tool for peoples less bound by C compatibility), or just a source of confusion?

BS: Actually, I'd like to see a ``student C++'' where built-in arrays wasn't used at all. Instead students would use vector, list, and string classes from a teacher-provided library (based on the standard library no doubt). That is easily done and easily enforced in a teaching situation - even without compiler changes (just decrease the grade given in a built-in array was used). Similarly, a teacher would find it easy to ban explicit casts; they have no place in the kind of code a beginning student should be writing. The difficult part of learning C++ - or any other language - is learning the new programming and design techniques, not the language features used to express them.

Far too often, people gets obsessed by language features. Far too often, programmers get lost in a futile attempt to understand every aspect of a rich language without sufficient background to understand the techniques the features exist to support. It is worth noting that even C++ is orders of magnitudes simpler than the environments, frameworks, and major applications

we work with in real-world application development.

In teaching, C++ has been hurt by its close - and valuable - relationship with C. Because C++ is (almost) a superset of C, many think that they must learn (almost) all C features and techniques before approaching C++. This is not so, C++ is in many ways better behaved than C, and libraries can be used to avoid having the student face the complexities of C pointer manipulation, casting, arrays, etc. until the basics have been learned in an environment containing proper vectors, strings, etc. C++ can be - and occationally is - an excellent language for teaching programming, programming styles, design, etc. However, we must distinguish the teaching of programming from the teaching of programming languages. That done, we might make some progress and might even avoid many of the silly language wars that too often wastes our time.

One strength of C++ as a teaching language is that it lends itself to the teaching of a variety of design and programming techniques. The alternative is to teach a variety of "cleaner" languages to illustrate the same range of techniques. What I consider flat wrong is the present one design and programming style - usually embodied in a single programming language - as the one and only true style. A professional programmer or computer scientist should eventually be comfortable with C++, Smalltalk, ML, Lisp, and Eiffel - just to mention a few. Naturally, few people can be real experts in more than one or two or those at one time, but the ideal must be to be acquainted with all and over time try out every one in some real project.

CP: C++ exceptions have been criticized as hard to use correctly - cfr the Cargill's article in C++ Report, the "need" to introduce auto_ptr in the standard library to make pointers more manageable under the presence of exceptions, the mismatch between templates and exception specifications. It also seems to be somehow difficult to implement: Borland had serious problems linking together DLLs with and without exception handling. Not to mention the "retry" debate, that you covered well in your D&E. So, considering that exception will also make the learning curve for C++ steeper, do you think that - as they are in C++ - exceptions are paying off?

BS: Every new feature is deemed hard to use, expensive, and unnecessary by some when it first appears. Of the many "problems" pointed out, few are real problems in real programs. I find that exceptions significantly simplify my code. Like all really worthwhile features, they require some new thinking and some new ways of organizing code (if not, how could the feature be a significant improvement?), but I think that they are emminently worthwhile. I consider the so-called "mismatch between exceptions and templates" bogus. Exceptions are for building firewalls against error conditions. That is, you choose a specific interface and decide to let only a subset of error conditions through. Almost all templates are poor candidates for firewalls. Templates are deliberately designed to interact very closely with user-defined types and if you have any sense you'd not try to build a firewall right through closely interwoven code. If that's a mismatch, so be it. However, I see it simply as independence of the concepts. They do differnt things and they can be used in combination.

CP: to some extent, there is a mismatch between templates and *exception specification*; since templates uses the actual arguments in some way, it is almost impossible to end up with a correct specification, even on innocent-looking, simple code (e.g. a template function for comparison)... I'd say in general that if there is any dark side in exceptions, is that they makes innocent-looking code not so innocent anymore.

BS: Actually, a lot of such "innocent looking code" was never so innocent. Such simple code is often full of unchecked error conditions and subject to being bypassed by C-style setjmp/longjmp. Thus, exceptions focus attention on a problem that many prefer to ignore, but the complexity is there already; it is not introduced by exception handling. I don't think it makes sense to add exception specifications to template functions - at least only to rather unusual template functions. The reason is - as you correctly point out - that the exceptions potentially thrown are the ones thrown by the template plus the ones thrown by the template arguments. That is one reason that I say that templates are usually not good candidates for firewalls. I do not believe that it makes sense to make every little piece of code bulletproof. Instead, I prefer to express systems in terms of sub-systems and make the sub-system boundaries the firewalls. That is where I use exception specifications.

CP: An area where C++ seems still weak is object persistency... there are a number of libraries/tools, but in most cases you end up needing a custom preprocessor or some handcoded function. RTTI seems a promising way to go, but unless there is some standard, it would be just another nonportable extension...

BS: I'm not at all sure that persistence belongs in a general-purpose programming language. Different people need different types of persistent data with radically different requirements on performance, reliability, access control, nature of quiries, etc. I think it is right to leave this issue to library vendors and database vendors. I prefer to limit the use of preprocessors and extra-linguistic tools, but sometime they are needed. In my opinion, a programming language should not try to do everything. It cannot do everything well anyway. And yes, RTTI can be of considerable help to implementors of various persistence and data base services.

CP: You are the designer of one of the most successful languages ever - do you have any suggestions to the countless peoples in academy who crunch out yet-another-language every other day?

BS: Be guided by problems. A useful language is a solution to a well-understood set of problems rather than simply something that fullfils all the currently fashionable criteria for what a programming language should look like. If you don't have a serious problem that cannot be handled reasonably with any existing language, don't even think of designing yet another language. Language design is a field with an almost 100% failure rate. No sensible person would enter that field if there was an alternative. So, look for serious programming problems without acceptable solutions, and try hard not to get involved in language design. If you must design a new language, borrow as much as you can - with acknowledgements - from an existing language. Be prepared for failure, and for extraordinary amounts of work should you beat the odds and succeed.

CP: The other most successful language is visual basic. Some said that it delivers where OOP/C++ promises, that is, lot of pluggable components, true reuse, maybe at the expense of some underengineering. Do you feel it right that interoperability between different C++ compilers is left to third-party products like SOM, and is not part of the standard? Of course any binary standard would limit the freedom of the compilers writers, but that's true for *any* standardized issue. One may squeeze out some more performance by giving up multiple inheritance support, but that's not a good reason to remove MI from C++. Why is a binary standard any different? [naturally the binary standard is only one step toward "true" software components]

BS: C++ delivers what it promised. It cannot be expected to meet the hype of every language and system claiming to be object-oriented or whatever. C++ is a programming language, not a module specification language or an operating system. It - like any other language - cannot be everything to everybody. You can build "pluggable components" using C++, but that is not the primary aim of C++ and it takes work. Interoperability is a hard problem. People generally don't appreciate that only through lots of work an lots of agreements between competing organizations can even interoperability of C program fragments compiled by different compilers be achieved. There has to be agreement on function calling sequences, data layout, floating point arithmetic details, etc. C++ is harder than C, but not much, because almost all of the hard problems are political rather than technical. The issue of multiple inheritance is completely separate from any issue of binary standards and interoperability of C++ compilers. I believe that the absence of MI in at least early versions of SOM reflects nothing but a leaning towards the Smallatlk and Objective C among the initial SOM designers. In a language like C++, that relies on static type checking of interfaces, some form of multiple inheritance is essential. The alternative is warped code, unsafe interfaces, or both.

CP: Do you have any new concept, idea, or feature that you are thinking of for a "next generation C++" and that you'd like to anticipate to us? [I understand you want C++ to be stable for a while; I guess that does not stop you from thinking to improvements, maybe even improvements on some implementation issues for existing features].

BS: My feeling is that when it comes to programming languages, people pay lipservice to experimentation and try to treat the field as a branch of mathematics or philosophy. However, I feel that the next generation C++ should come from real problems in real applications and from experimentation rather than speculation and polishing of the existing language. I find it much easier to describe what I have done and what I think of things I understand some of than trying to predict the future. I love Science Fiction, but not when it masquarades as technical articles. I do think, though, that we have too many "true believers" and too few experimentalists in our field. To improve our computer systems we must have lots of good experiments and tons of reliable data. From that can come the insights that allows us to determine what are the real problems and how to solve them. Far too often, we just sit around philosophising about our feelings, opinions, and theories instead of making real progress.

Reader's Map
Several visitors who have read
this article have also read:

Biography

Carlo Pescio holds a doctoral degree in Computer Science and is a consultant and mentor for various European companies and corporations, including the Directorate of the European Commission. He specializes in object oriented technologies and is a member of IEEE Computer Society, the ACM, and the New York Academy of Sciences. He lives in Savona, Italy and can be contacted at pescio@eptacom.net.