Stroustrup Says C++ Education Needs To Improve

simoniker writes "Over at Dr. Dobb's, C++ creator Bjarne Stroustrup has given an in-depth interview dealing with, among other things, the upcoming C++0x programming standard, as well as his views on the past and future of C++. He comments in particular on some of the difficulties in educating people on C++: 'In the early days of C++, I worried a lot about "not being able to teach teachers fast enough." I had reason to worry because much of the obvious poor use of C++ can be traced to fundamental misunderstandings among educators. I obviously failed to articulate my ideals and principles sufficiently.' Stroustrup also notes, 'Given that the problems are not restricted to C++, I'm not alone in that. As far as I can see, every large programming community suffers, so the problem is one of scale.' We've discussed Stroustrup's views on C++ in the past."

Comments

more to it

[g2devi]

That may be so, but there's more to it. It's only been recently fully implemented and a few years ago Stroustrup himself commented that he's constantly surprised that some things (e.g. template recursion) are even possible in C++.

The language is overly complex. The key advice any C++ expert is "restrict yourself to a specific subset of C++". That's the bulk of the difficulty. If C++ were simplified to include only that subset, you'd have a lot less need for training,

Re:more to it

[shutdown -p now]

Unfortunately, even if you keep away from the more advanced stuff such as template metaprogramming, there's still plenty of surprising (in a bad sense) behavior in the most basic C++ constructs (mostly inherited from C, but exacerbated by C++ pretense at high-levelness, and with some new stuff thrown in for good measure). Pointer arithmetic, unsigned types, silent implicit narrowing conversions, PODs vs non-PODs (which in turn separate into aggregate types and everything else) and the subtle differences in initialization of those, ability to delete void* and pointers to incomplete types (and the wonders of virtual destructors), etc... The language is a mess, really. It is very powerful, and the existing code generators are very good at optimizing stuff, but writing in C++, or any subset thereof, requires a very good understanding of subtle details.

A simple example: given a vector of ints, calculate the average value, using standard idiomatic C++. Let's give it a try:

#include <algorithm>
#include <iomanip>
#include <iostream>
#include <istream>
#include <iterator>
#include <numeric>
#include <ostream>
#include <vector>

int main()
{
std::vector<int> v;
std::copy(std::istream_iterator<int>(std::cin), std::istream_iterator<int>(), std::back_inserter(v));
int avg = std::accumulate(v.begin(), v.end(), 0) / v.size();
std::cout << avg << std::endl;
}

A simple task, and a clean implementation, using the standard library as much as possible. No fancy language features used - no template metaprogramming, no pointers, no virtual inheritance. The kind of stuff a new C++ programmer might write after reading Stroustrop. And yet, it gives wrong result for certain input data. Now, who'll be the first one to spot and explain the problem here, preferrably without actually running it? Bonus points for explaining why we need "istream", "ostream" and "iomanip" headers here in addition to "iostream" (and we do, if we want this to be portable).

Re:more to it

[truedfx]

I'll admit I had to run the program, and even then it took me a while to see it.

v.size() has vector<int>::size_type as its return type. Depending on which specific type, either that type will be converted to int, and all is fine, or the result of std::accumulate(v.begin(), v.end(), 0) is converted to an unsigned type, then divided, and then converted back to int. In the latter case, when you have a negative sum, the division gives the wrong result.

I think the <istream> and <ostream> headers are required because the types of std::cin and std::cout might otherwise be incomplete, meaning operator<< might be undeclared. I know the <ostream> header is also required to that std::endl is declared. I'm not seeing what could be in <iomanip> that's of interest of your program. Could you explain that last one?

Re:more to it

[shutdown -p now]

I know the header is also required to that std::endl is declared. I'm not seeing what could be in that's of interest of your program. Could you explain that last one?.

Ironically, it's my mistake - std::endl is also a stream manipulator, so I figured it is properly supposed to be in "iomanip" as well, and didn't bother to double-check (let's face it, in real world we usually just include "iostream" or other stream header as needed, and that's it - the rest gets dragged in, even if the Standard doesn't guarantee that it will). But you're absolutely right, std::endl is in "ostream", shame on me. And I've been coding fancy C++ stuff (with templates, STL, and all that) for five years now... *sigh* I guess it sort of serves to reinforce the point I've originally made :)

Re:I completely agree

[schmidt349]

Yeah, but that doesn't sound like the sort of problem that gets fixed by learning C++, which has a nasty habit of adding loads of complexity in all the wrong places. If you want tight, fast code and your project doesn't need OOP, use C.

As for programming pedagogy, I think we'd do a lot better if the faculty of CS departments would migrate away from using Java/C++ as the introductory programming model because so much of what gets said initially just goes in one ear and out the other. I will admit to not remembering at all how typedefs or templates in C++ work, and I can't say it's harmed me much.

Python would be a much better choice in my view for a variety of reasons (and I say this though I'm a Perl nut!), or hell, if you teach them Lisp they'll be horribly screwed up for the rest of their lives but at least they'll understand how registers and OOP work.

In short, novice programmers are not going to learn anything useful if you use C++ as the prescriptive model for how a well-written computer program should look -- they're just going to hit the bottle earlier in life.

Education, even at Universities, needs to Improve

[Lord Byron II]

I'm continually amazed that the CS majors at my major US accredited university can't program using pointers, object oriented techniques, or parallelism. I think that there is an attitude among the professors that this material is difficult, so it shouldn't be worth a lot of points. As a result, the students don't bother to learn it.

I work in another department and sadly, without formal CS experience, I'm a better programmer than many (if not most) of the CS department's graduates. I don't think, however, that this problem is unique to my school. I've visited other US universities where the situation is very similar.

In fact, I recently took an informal survey of about a dozen CS seniors and found that none (yes, none) of them knew what K&R, the "white book", or the "Art of Computer Programming" were.

Re:Education, even at Universities, needs to Impro

[IamTheRealMike]

My university course spent about half an hour on pointers in a 3 year course. Most of that half hour was factually wrong: the slides were full of code samples that wouldn't compile or would always crash.

They did, however, spend two terms teaching Hoare logic. Or rather, they spent one term teaching it, and then repeated the same material in another term with a different lecturer, because their communication was so poor they never realised they had duplicated their teching.

Friends at other universities reported similar stupidities, though not always on the same scale.

C++ is a rather complex language, but simplifying it won't help. The problem is that low quality education is rampant.

Strostrup is the problem

[Animats]

The big problem with C++ is Strostrup. He's in denial about the fact that the language is fundamentally broken. But he's still influential in C++ circles. Thus, no one else can fix the mess at the bottom.

The fundamental problem with C++ is that it has hiding ("abstraction") without memory safety. This is the cause of most of the world's buffer overflows. No other major language has that problem. C has neither hiding nor memory safety, so it is still vulnerable to buffer overflows, but they're to some extent visible at the place they occur. Pascal, Modula, Ada, Java, C#, and all the interpreted "scripting languages" have memory safety. C++ stands alone as a language where you can't see what's going on, and the compiler doesn't have enough information to check subscripts.

The reaction of the C++ standards committee has been to try to paper over the problems at the bottom with a template layer. That didn't work. The template classes just hide the mess underneath; they don't make the language memory safe. There are too many places that raw pointers leak out and break any protection provided by the templates. The template language itself is deeply flawed, and attempts to fix it have resulted in a collection of "l33t features" understood and used by few, and too dangerous to use in production code.

The fundamental cause of the trouble comes from C's "pointer=array" equivalence. That was a terrible mistake, borrowed from BCPL. The trouble is that the compiler knows neither which variables are arrays nor how big the arrays are. You can't even talk about arrays properly. I mean, of course,
int read(int fd, char* buf, size_t len);
That's just trouble waiting to happen. "read" has no information about how big "buf" is.

C++ added references to C, and should have added syntax like
int read(int fd, char& buf[len], size_t len);
to go along with it, so that arrays became first-class objects with sizes. But it didn't. There are some other things that have to be done to the language to make this concept work, but this is the general idea. This is the elephant in the living room of C++, and Strostrup is in denial about it.

Every time you have another crash from a buffer overflow, every time you install another patch to fix a buffer overflow, every time you have a security break-in from a buffer overflow, think of this.

So long its not Matlab

[Latent Heat]

The trend we need to fight is not teaching programming using Java, C, or C++, but defaulting to Matlab as a programming language of choice. This is what is happening in Engineering colleges and other non-CS places at the U.

Granted, Engineering always went for things that CS considered "brain dead" -- Basic, PC's, DOS, Windows. But Matlab is more brain dead than most.

What happened is that a lot of the current generation of Engineering profs cut their teeth on FORTRAN -- their Intro to Programming was in FORTRAN, whatever industrial job they had before getting a PhD had them compute things in FORTRAN. Few of them were ever comfortable in it and most of them spent hours in the computing center debugging programs dumped to massive punch card decks.

When Matlab came around, it was numerical Nirvana. It had this massive numeric library that you didn't have to write your own Q-R linear equation solver or SVD subroutine, and you didn't have to go searching for this stuff either, it was all there. It had a command prompt to performed immediate execution along with reasonably friendly error messages. And it acquired a thoroughly feature-full graphics package.

Don't get me wrong, Matlab is a very capable numerical applications language and even turns out to be one of the better Java scripting languages of all things. But it really falls down in terms of extensibility of its type system, and as far as what Mathworks tacked on for object-oriented programming, fuggedaboutit. It is also the Swiss Army knife of software for a whole bunch of people, and forget about introducing them to a socket wrench and handle that can apply serious torque to a bolt when they think they can get by with the pliers tool.

While people who know what they are doing can benefit from the convenience of the numeric and graphics libraries, the immediate mode, the verbose error handling and rare instances of complete crashes, if you don't know what you are doing (i.e. you are just learning), it can lead to as many hour-gobbling skull-cracking debug sessions as anything else. Our required Numerical Methods course is in CS, it uses Matlab, our faculty is complaining that the students are complaining that they hate the course because they are spinning their wheels trying to get programs to run (in Matlab of all things), and we have guys in our department we want to teach Numerical Methods (in Matlab, of course), in the context of a watered-down Intro to Engineering offering.

What the community needs right now is a Python distro with enough of a numerics and graphics package rolled in to do 90 percent of what is in Matlab (Are the Python people still hashing out that Numerics/Numpy divide? Is there an engineering graphics library that is Numerics/Numpy compatible? 99 percent of what you do in Matlab is that you have a Leatherman Tool of a 2-D array type (Matlab, Matrix Lab) along with all of the libraries being compatible with that type.) CS departments could teach their Intro to Programming along with their Numerical Methods courses using that Python distro, and we can save a generation of engineers from brain damage.

Re:C++ has issues

[SageMusings]

Real Mike,
you have my full attention. Please, support your assertion that Java and C# suck rocks.

C++ can be fast at execution time but the development time is prohibitive in many applications where you need to be agile and actually ship code in a hurry. I try not to get hung up on all the esoteric points of different programming languages, although I am quite amused to read other's comments. Yet, I will hazard a post on this topic.

I learned C++, not all of it to be sure, but the portions 85% of us might need in a given project. It may be intellectually stimulating to code an app form a "purest perspective" but many of us have to earn a living and produce a lot of code in short order. C++ does not fit this bill. Most applications just have to work and work today, not next quarter. Then we have to extend the app after a few months. Since C++ is quite a bit harder to read and I have to learn code I did not write in short order to perform this maintenance, I enjoy Java and C# apps a lot more than ones coded in C++.

Please, tell me why Java and C# suck compared to C++ in the practical world. Nearly all of us are not writing low-level, time-critical code. Most of us write apps for business transactions. I happen to write business software that is widely distributed and the C++ performance boost is nullified by the latency of remote calls to distant servers.

Please, tell me the advantage of writing an app in a year vice 6 months.

There are many languages because there are many problem domains. C++ is not the best language. There is no best language, period.

Nearly all serious desktop software

Finally, it has been my observation most "serious" code is no longer constrained to the desktop.

Is it because it is just a language?

[dwarfking]

Reading the threads, many people are discussing the relative merits of the other programming languages/environments - Java/C#/Python/etc - but what do all of those have that C++ does not have?

More complete environments.

When you install a C++ compiler, you get a C++ compiler and the standard library. When you install Java or C# or Python you get libraries to support simplified Networking, IO, Database access, GUIs, Memory Management, Threading and more.

Now it is possible to find all that for C++, but they are all separate components that the developer needs to decide on and download. And the number of choices for each is large. Do you use wxWidgets or FLTK or GTK+ for GUI, for example.

The other environments actually reduce your options, and for projects on a timeline the less time you spend on determining what you need to accomplish the task, the sooner you finish. Yes you can bring in replacement libraries in Java or Python or C#, but few people do. The folks that wrote those libraries did a pretty reasonable job on them, and since they are bundled with the standard installers, unless there are really specific needs, there's rarely a reason to replace them.

Look as an example of this at the Mono project. It is an attempt to provide the C#/.Net environment outside of Windows, but it does not have as much traction as .Net on Windows, why? in part because the .Net frameworks are more complete on Windows than in Mono. I not many .Net developers that use WindowsForms in every project. Without that piece of the eco-system already available, their project would take much longer. Mono basically provides C# for Linux, just another programming language.

I've watched over the years as some folks tried to assemble Java-like libraries for C++, but they didn't really take off.

This appears to me as why C++ has the reputation of being so hard to build applications in. The developer has to do so much extra work just to get to the point of assembling the program that the Java or Ruby or C# or Python crowd gets out of the box. Is this the fault of C++? Not the language, but perhaps it is something the steering committe should address. As someone pointed out in an earlier thread, the C++ standard group likes to make the comment that a particular given feature is not part of the language. Perhaps they should rethink that stand.

As point of background, I started working with C++ when it first appeared as a pre-processor that created C code that was compiled by a C compiler (when you had to use the keyword Overload). I later moved into Java and have made a good living doing Java development. Recently though I have gotten deep into programming in 3D graphics with OpenGL. I'm doing it both in Java (using jogl) and C++ (direct gl calls as well as engines). This is one area where there is not a clear choice for any platform, but because in the Java world I have the Networking and Threading, I was able to put a system together much quicker than I could in C++. Of course the Java approach has it's own problems because of the sheer volume of objects created/destroyed (imagine a 3D model made of Vector3D objects), so I end up using C++ approaches using float[] arrays (also an object, but only one).

Sorry for the ramble. Anyway, the point is, I personally think C++ would be more acceptable if it really was an eco-system and not just a programming language.