Stroustrup on the Future of C++

/ASCII writes "Bjarne Stroustrup, the father of C++, has written an essay [PDF] on the features of the upcoming C++0x standard. In his essay, he argues that new features should whenever possible go into the standard library and not into the language, but that the language needs to shave of a few rough corners to make it easier to use for novices."

Comments

Dynamics in programming languages

[jurt1235]

It is interesting to see how an abstract language like a programming language evolves through time. I would think that in a language like C++ the more userfriendlyness should be in the libraries and not in the language, so I agree with the author on this point. Putting it in the libraries makes it better backwards compatible, and distributable.

a 'few' rough edges

[leomekenkamp]

Shaving off a 'few' rough edges to make it more easy for newbies... It's that what Sun tried to do? (hint: java)

Re:a 'few' rough edges

[leomekenkamp]

Yes, I know these are fundamental changes, that's why I put the word 'few' in between quotes.

Point of fact stays however: c++ has an enormous learning curve. Complexity in the language is imho comparable to Latin: you can almost recognise the author by lookin at the code. That complexity is far, far less in java.

Especially the gc makes life a lot easier for even seasoned programmers.

And by the way: you can be almost as un-OO in java as you can be in c++. Introspection in java is mostly used in 'frameworks', like for instance the serialization API; normally you almost never use it.

Re:a 'few' rough edges

[orthogonal]

Shaving off a 'few' rough edges to make it more easy for newbies... It's that what Sun tried to do? (hint: java)

Java's ok, but a few mistakes were made. Among the worst: over-reliance on the instanceof operator, and the need for something called a MethodNotSupported exception.

Now think about what MethodNotSupported means in the context of Barbara Liskov's Substitution Principle:

We've got a language that supports inheritance. Inheritance makes sense only insofar as Liskov's Principle is followed. Put simply, if class D inherits from class B, any "promises" that B made -- it's interface, both syntactically, in terms of public member names, and semantically, that the member functions really do do some particular thing -- any of those promises must also be kept by D, because we are saying, as Liskov tells us, that D is-a B.

If I hand you an instance of D, and you plug it in place of am instance of B, anything that the B instance does, the D instance must at least do. Otherwise, how can you safely and with confidence substitute a the D in where the B used to be?

But along comes Java, and the MethodNotSupported exception -- a method used by some of Java's core library, including, dismayingly enough, commonly used container classes. It's used so that a derived class can suddenly and at run-time throw it, telling you, "Whoops! Ha ha! Trickled you, I can't in fact really do what B promises! I claim I'm a B, but really, I'm not!"

Now imagine you had a component stereo, and you replaced your tuner with a "better" tuner with all the same jacks as the original. You'd plug it in and expect it to do the same work as the original tuner, just with better reception or fidelity or whatever. Now what if you plugged in the new tuner to the old connections, and suddenly its display read: "Sure, I have a two-plug RCA input, but it's not actually connected to anything, so you can plug in an RCA jack but no sound will come out". That's what having a method that throws MethodNotSupport is like: the jack -- or member function name -- is there, but the implementation is not.

D inherits ("extends") B, it says so right there on the label, but no sound comes out. With Liskov undermined, you don't really have an object oriented language at all. You don't really have true inheritance. What you have is a sham, a sham that can break down on you at run-time because an object is lying to you.

And the upshot is that you don't see enough real object oriented programming by client Java coders. If the core library can get it that wrong, why should clients of that library bother to adhere to a standard?

So you get lots of java code that does RTTI using instanceof and if( p instanceof Foo ) ... else if( p instanceof Bar ) ladders rather than calling p.virtualFunction(). Because coders won't take seriously OO that the language and its core libraries don't take seriously.

Re:a 'few' rough edges

[orthogonal]

"That exception lets you know when you're breaking an API contract -- though a List may have add() methods, if you want that List to be immutable, is it preferable to throw an exception from a mutator or remove the mutator from the subclass completely? Or should we just take add() out of the List interface?"

If those are the motivations, we should have an ImmutableList class without an add function, and derive the List with an add function from it (in C++, probably by private inheritance, that is, inheritance of implementation). Only the List class with the (working) add should publically inherit from the interface with an add function.

Problem solved.

"I could read stupid C++/D/C/Ruby/whatever code and then blame the language too, you know."

Sure, sure. Don't be touchy. I code Java too, in fact I'm Sun Certified.

What I'm saying is, you tell me you'd "fight" unnecessary uses of instanceof, and I believe you. I'm just trying to make your fight easier. The point is, if your core libraries don't take liberties, your fellow client coders will be less likely to -- or at least less likely to use the bad example ("but the core library does it, why can't I?") when a good java programmer, like you, points out the problem.

Is threading going to be abstraced out ?

[Gopal.V]

Stuff that went into Boost should be in the standard library from now on... Also anyone who has had to use g++filt will agree with me that C++ STL error messages need to look less like the dog's breakfast :)

A better wheel

[Anonymous Coward]

Am I the only one who wonders why we need a successor to C or C++?

I've sat through the past 10 years and watched things like Java and D and Objective C come, but meanwhile most serious OS level and game development continues to be in C or C++. Doesn't this demonstrate that new language are merely a distraction to developers who haven't fully exploited the current set?

Re:A better wheel

[onnel]

No, it's more a signifier that different problems need different solutions. If all you have is a hammer, everything may look like a nail, but if you know c++, perl, php, java and a few others, your toolbox will contain an appropriate solution for a broad range of problems. I write my games in c++, my websites in php and my text parsing utilities in php. Java and python also serve their niches. Anyone who believes that a single language is the best solution to all of the broad range of problems faced by programmers is sorely mistaken. Onnel

What we need is "Smoothed out" people...

[brockbr]

Every time a cost of entry is reduced and the expected prerequisites are thrown aside, the result is a truly less capable *person*, not a more capable tool.

Agreed - There are some things that need smoothing, but I for one am tired of dealing with people who lack a fundamental understanding of the *systems* as a whole. Examples of this are the folks coming out with CS degrees who aren't even capable of following a thought, let alone starting an actual career designing and developing software.

For them, a tool like Java offers an entry level that is acceptable given there current capabilities - A tool geared towards THAT user (and a fine one at that).

But C (and C++) can be leveraged by people who know the tool and *use* the tool for what it can do, even with it's high(er, intellectually) cost of entry.

Syntactic candy.

[mcc]

This is mere unnecessary conveniences. What exactly are they giving us here? Well, if I'm reading this right, "C++0x"* will give us:

• Some cases where you can type ; where previously {} was necessary, saving as many as two keystrokes.
• The lexer ambiguity where a<b<c>> and a<b<c> > meant entirely different things is fixed, saving as many as one keystroke.
• EVEN MORE "free" constructors which may or may not work quite the way you wanted them to in practice.
• More "fancy cast" operators, which sounds nice until you remember that if there's ever a time you find yourself using the C++ "fancy cast" operators, then there is a damn good chance that it's because you're doing something unwise enough it really would be a better idea to refactor the code to make the "fancy cast" unnecessary.

...in the meantime the fundamental issues with C++ remain not only unresolved, but unaddressed. The template system is still not a generics system, but an ugly cut & paste macro system which can incidentally be used for generics, with some caveats. The class system is still fundamentally brittle and unfriendly to simple things such as upgrading a DLL or determining at runtime if two objects are of the same type. The syntax is still a forest of features whose features interact in ways so crazy and unpredictable it approaches Perl in its chaos; references are still gimped; the distinctions in behavior and use between static and dynamic objects remain awkward and newbie-unfriendly. The features that people obviously desire to have in the language as demonstrated by their tendency to hack them in with third-party libraries (like BOOST) are-- they tell us-- a good thing, and they tell us we should continue to hack them in with libraries (like BOOST). That's nice. You know, that would be a lot easier though if we had a macro system** capable of anything smarter than blind code cut and paste-- or at least a macro system** fundamentally designed to be used for anything at all other than generics.

Meanwhile, it appears if I'm reading this right that the most important differences in C++0x will be changes to the standard library. Great. The STL was defined how many years ago, and it's only just in the last few years that compliant implementations have become commonplace? How many decades will it be before the "C++0x" library changes have become common in a cross-platform compatible way?

C++ is an extremely useful language, and making an update to C++ which changes as little as possible so as to follow some kind of "if it aint broken don't fix it" principle is an idea which makes a whole lot of sense. However it seems likely to me from reading this that C++0x will offer so little significant difference from C++ as to make itself simply redundant.

* ("C++0x". Were they specifically trying to come up with a name less convenient than "C#"? Ah well, I guess we can call it "COX" or "cocks" for short.)
** "Template system"

How about a module system?

[putko]

More powerful languages like ML, CAML and some Schemes have a module system. This allows the programmer to control the meaning of identifiers (e.g. what function they mean). E.g. what gets imported (and with what name), what gets exported.

Often there is a separate notation for the modules -- it says what files contain the things, what's exported/imported etc.

That's very useful for encapsulation.

C++'s namespaces are a crappy attempt to get the benefits of a module system -- that's likely where he got the idea from.

Hygienic macros would be a good step too.

Re:C++... always the ugly step-kid

[NDSalerno]

Here here! In the realm of native application development I have programmed in C++, Delphi, and Eiffel. Each language has its pros and cons. C++ is no doubt the most popular and most successful of the three. However, I have been most productive and had the most fun with Eiffel and Delphi. Some might be quick to say I am not experienced enough with C++ or that I haven't learned the idioms. I have programmed in C++ for many years. I have read all the books written by Stroustrup, Scott Myers, Herb Sutter, Andrei Alexandrescu, etc. I regularly read the newsgroups and paid attention to Sutter's Guru of the Week articles. With all that under my belt I still make mistakes. Why? In Eiffel and Delphi a statement of code can do one or two things. In C++ a statement of code can do several things, some of which incur side effects of their own. If you don't understand these behaviors then you write buggy code. Or, in other words, C++ is complex.

The hardcore crowd will tell you that ultimately you are better off going with C++ because of speed and power (implying that every other is a step above Visual BASIC). NONSENSE! Delphi is on par with C++. In certain situations C++ is more efficient than Delphi. In other situations Delphi is more efficient than C++. In the long run they are both the same. The only thing Delphi can't do is metaprogamming. Big deal. I can write my number crunching algorithm in C++ templates and wrap it in a library that Dephi can call. Problem solved. You know what C++ can't do elegantly? GUI development. It's tortuous. C++ Builder doesn't count because the language is modified and is not ANSI C++, which should really tell you that there is truth in the statement that C++ lacks some features. Qt has always been my favorite ANSI compliant GUI library. However, even Qt is a testament to the needed addition to C++. The meta-object compiler?? As if my compile times weren't long enough?

I have debated on the C++ newsgroups about accepting Borland's C++ extensions to the language, namely properties, closures, and reflection (advanced RTTI). I admit that boost does have a way to implement closures as template classes. They work and are efficient. My only dislike is that the syntax is too cumbersome. But I'll let that go. However, it has been argued and proven that properties CAN NOT be implemented as a class library without overhead. They simply must be a feature of the language. Combine that with better RTTI support (reflection capabilities) and you will have some powerful new tools to make elegant designs. And no, I am not advocating using RTTI for general development. Obviously if you have to test your base class pointer to discover the actual child type then your design is flawed. But that does not mean RTTI is as evil as goto statements. Properties and reflection is superior in the domain of GUI frameworks. GTK++, MFC, Qt, etc., should be proof enough that GUI libraries are too cumbersome without properties and reflection.

A lot of people ask for C++ to have some kind of GUI library but I do agree with Stroustrup's rationale for why that is not possible. There is no common denominator of design that would make sense for every platform and would appease everyone (whom has their own opinions on the design of a GUI framework, which are radically different). It is better for C++ to rely on third party libraries that deal with GUI development. That is fine for me. What I ask for is that we improve the language to allow better development for these third party GUI libraries.

Outside of GUI development, properties can be useful. The hardcore group, however, always argues against syntactic sugar because it is somehow pointless. This may be true for computer scientists. It is not true for software engineers. I always welcome any kind of help in reducing bugs. Syntactic sugar can actually help. To say it briefly, Delphi and Eiffel read like English. C++ reads like machine language. Guess which group of developers is going to have a higher defect rate?

Nicholas

Re:The future of C++...

[Nasarius]

... is extinction. Just stating a fact here.

Obviously. The same goes for any language. But what do you propose to replace it? The vast majority of "modern" languages are not compiled to machine code. For the moment, C++ is a good balance between the efficiency of C and the user-friendliness of higher-level languages. It fills a niche (resource-gobbling 3D games, operating systems, compatibility with C/C++ libraries, etc) that no other language does.

Re:How about...

[CptNerd]

..we wait for people to actually understand amd implement the current C++ standard?

Unfortunately, because the spec is so convoluted, you can implement some aspects in completely incompatible ways, and still be "implementing the standard."

I've been in the profession for over 20 years, and I've seen the same thing happen to certain "standards". Something new and innovative will come along, and within a decade, features will be added that will make it unusable, and foster the demand for "something like X but easier/simpler/faster/cleaner". Most of the feature creep is driven by large corporate interests and the demands of their programming staffs to add features that they claim they can't develop without.

Not many people remember X Window 10R4, but the entire Xlib documentation took up a half-inch thick stack of single-sided pages, and described a basic networked window system that was useful enough to develop distributed systems with. The protocol was easy to implement, straightforward, and clean. But that wasn't good enough for the wonks at Apollo, DEC, HP, Sun, AT&T, and the researchers at MIT. So, once everyone added their "must have" requirements to X Window (we have to have font servers or we can't write good systems!) you ended up with a monstrosity of an Xlib spec that took up a thousand or so pages of documentation, added thousands of new points of failure, and spawned an entire bookshelf from O'Reilly just to try to explain how to use X.

I saw the same thing happen to C++. In the early days it was what it set out to be, a simple, easy to implement extension to C that let you do basic object oriented programming. Sure, it wasn't Smalltalk or Scheme, but it was good enough for 90% of the tasks that really needed simple lightweight objects. But, certain vested software interests demanded that the spec be extended to include things like templates and virtual this and that, all designed to save some corporate programming group some design time, at the expense of creating a spec that required Talmudic scholars to interpret.

I've been on the outside of Java, watching it since it was first described, and I've seen the changes and additions to the language spec over the years. The same thing is happening to Java that I saw happen to X and C++, and I predict that in 10 years or so someone will come out with a language "that does what Java does without all the overhead". I knew Java was doomed to follow the C++/X path when they announced namespace support in the language. Sorry, Javaheads, but that was the tipping point, and you're all on that long slide into obfuscation and bloat that many derided C++ for.

I know many who actually read this post won't believe it, will argue that "Java is different" somehow, but mark my words, Java will bloat into uselessness within 10 years, just like C++ did.

Re:wait a moment...

[Kurt Granroth]

Not when it's God talking

Speaking of which, did anybody else notice the author description in that article?

"Bjarne Stroustrup is the College of Engineering Professor in Computer Science at Texas A&M University"

What!? Isn't that like saying the George Bush is a businessman from Texas? It's technically true, but completely misses the point on what matters.

This makes even less sense when you try to find the target audience. If the person reading the article already knows who Bjarne is (by which I mean every single C++ programmer), then I could see them omitting the entire "creator of C++" description because, well, they already know... but then why have a description at all?

If the person doesn't already know who this "Bjarne" is and they are reading the article, don't you think they'd want to know that he created the thing and therefore might have some relevant opinions on that matter?

Very odd...

Strostrup is in denial

[Animats]

C++ is in deep trouble, and Strostrup is in denial about it.

C++ is the only major language with extensive abstraction but without memory safety. All other major languages are either memory-safe or don't hide the underlying machinery. (Java, C#, VB, Perl, Python, etc. all have automatic memory management. Some use garbage collection; some use reference counts. C is unsafe, but hides nothing.) This fact is responsible for millions of program crashes every day. Most security holes in C++ code come from this problem. Java and C# were invented primarily to eliminate the safety problems of C++. The open source community has generally stayed with C, where at least you can see by examination what's going on. C++ is losing market share to Java.

And Strostrup denies this is a problem.

This has happened before. Last time, it was Wirth. Wirth designed Pascal, Modula, and Modula II, but refused to admit that each had serious problems. He fought external compilation in Pascal. He fought extensions to the language. He even fought compile-time arithmetic. In the end, he took Pascal from a major language to a historical footnote.

Serious systems programming was once done in Pascal, but not in Wirth's version of it. The original Macintosh and Lisa software was written in nonstandard versions of Pascal. And much of the DOS era was built on Turbo Pascal. But proliferating nonstandard versions of Pascal caused another set of problems.

C++ has been in decline for years. "Evans Data has found that the percentage of developers using C++ has steadily declined over the last six years--from 76 percent in the spring 1998 to 46 percent in fall 2004." Strostrup also denies that.

The C++ committee has been taken over by template fanatics. Most of the committee's effort revolves around obscure template features that few will use, and which no responsible programming manager would allow on a mission-critical project. There's very little interest in language safety, and a vocal minority that insists language safety is undesirable or impossible.

All is not well in the C++ world. Claming otherwise is irresponsible.

Balance between ease of use and features

[Mingco]

The perfect balance of ease of use and richness of features for C++ is defined thusly:

If any novice is able to pick up Bjarne Stroustrups' The C++ Language and beat him within an inch of his life... but no more, then the book is perfectly balanced between ease of use and richness of features.

As it currently stands, I believe that every second or third novice would easily be able to beat Bjarne to a messy, bloody pulp almost as ugly as the code his language engenders.

It's far too late

[pammon]

"If you think C++ is not overly complicated, just what is a protected abstract virtual base pure virtual private destructor and when was the last time you needed one?" -Tom Cargill

The largest problem with C++ is its complexity. It is not just too complicated, it is *unmanageably complicated.* Some of the symptoms are:

• The STL has inexplicable omissions. For example, there's no portable way to seed the built-in PRNG in the random_shuffle algorithm, rendering it useless.
• Guruhood isn't good enough. Consider the seemingly simple task of creating a stack that works correctly with exceptions. It's extraordinarily difficult even for a guru.
• Language features interact in nonintuitive ways, producing a combinatorial explosion. For example, if you overload a function in a class, you don't have to use the scope resolution operator, and if you override a function, you don't have to use it either - but if you do both, you DO have to use the scope resolution operator or else you get a compile error!

Insofar as the new C++ standard adds stuff, instead of simplifying, this will only get worse. Since it's unreasonable to expect a new standard to remove features, the problem is unfixable. The result will be that programmers carve out their separate comfort zones, compiler vendors will not implement all features, and the monolithic C++ language will fragment into a Venn diagram of sublanguages. More so than today.