Why should I have written ZeroMQ in C, not C++ (part II)

http://www.250bpm.com/blog:8

by rumcajz • 14 years ago

128 156 14 years ago

> Rather it is a deficiency in the design of C++ language.

It's a deficiency of the C++ standard library, not the language itself.

You can find an implementation of intrusive lists in C++ in boost. http://www.boost.org/doc/libs/1_51_0/doc/html/intrusive.html

icefox2• 14 years ago

100% agree. It almost feels insulting to say that C++ programmer will only use std::list and a C programmer will only use a hand made doubly linked list. This almost feels like an interview question: > Which is faster and uses less memory? A generic double linked list or a hand made double linked list?

This is a data structure, optimization, software development problem. The reason we use any 3rd party generic (in C OR C++) doubly linked list is because it helps code the problem faster. Any time you use a component you need to be aware of its overhead. After profiling my C++ code if I found that the std::list was taking up all my memory and all my cpu time I would then evaluate the algorithms I am using and pick the data structure that best fits that be that a custom link list or something else such as the above mentioned boost library.

You even have to nitpick the example about erase being expensive. When using std::list why isn't the code passing around the iterators rather than the person object? And for the memory of std::list overhead why didn't he use std::list<person>?

There is no problem using a hand made C style doubly link list in your C++ code for the core part of your application for performance reasons. Being a C++ program doesn't mean you can't use C style code or even have asm snippets. Following the same logic as the blog ZeroMQ should have been written in ASM and not C because it could have been faster, or the reverse it should have been written in Ruby because it would have been coded in a quarter of the time (but at the trade of slower runtime)

As for the conclusion I would say that the inefficiency is still in ZeroMQ code as the author doesn't fully understand C++.

rumcajz14• 14 years ago

Why not pass iterators? Consider the case where the object is contained in two lists. You would have to pass a pair of iterators. What if it is in 3 lists? Etc.

rumcajz14• 14 years ago

Yup. It's possible to do that. My point was that if you want to do that you break the encapsulation, i.e. contained object would have to know about its container(s). This applies to Boost intrusive lists as well AFAICS.

scott_s6• 14 years ago

Which is entirely true, and is often the case when performance is a goal - abstractions must be broken. But that is not an argument for or against C++.

ch2• 14 years ago

Having no experience with C++ the next statement will just further show my ignorance of the language.

I was under the impression that an optimizing C++ compiler would be able to inline the container object and the contained object into one when working with template code, so that you would end up with something exactly like the C version but without the manual bookeeping.

At least I thought it could do this for certain types of classes, much like an optimizing C compiler can selectively inline functions based on heuristics.

megrimlock• 14 years ago

You are on the right track here, but consider what the contained object is:

    std::list <person*> people;

You're right that the instantiated list entry (what the article calls "helper") will directly include a value, but the value here is a pointer to a person, not a person.

zanny• 14 years ago

There is nothing stopping him from having a list of person and avoiding the memory overhead of references though.

ch2• 14 years ago

I had overlooked that he was using a pointer to the struct, and so I was interpreting things to be that C++ was actually doing under the hood allocations. I'm glad I asked though, cause I was really worried about C++ for a second there :)

So it seems like if the author had wanted he could have used C++ like a safer version of C macros, which if I ever need to use C++ will likely be my approach too.

sharth• 14 years ago

It's not using heuristics. It's how the code is actually written.

An implementation of std::list<T> is going to contain multiple objects of type __node<T>.

__node<T> would be defined as:

    template <typename T> struct __node {
        __node *prev;
        __node *next;
        T value;
    };

(Note that this doesn't change what anyone else said though, T is a pointer in this particular case, so there's still another indirection)

stonemetal• 14 years ago

It wouldn't even require an optimizing compiler just one that properly implemented templates. As megrimlock points out that would be a list of person not a list of pointers to person. Also his delete performance issue would not be an issue if done correctly. It would be a one liner of person_list.erase(person_iter).

So far as I can tell he is either trolling or has internalized procedural programming so much that he can't see any other path. I wonder what his thoughts on FP are.

fein• 14 years ago

While I am an advocate of straight C programming, I wonder if his issues with his person list in c++ vs a person dlink list implementation was an incorrect comparison.

stl vector would have been the proper "optimized" way of implementing that list of people in c++, as vectors don't deal with the heap. If your vector's memory space can fit entirely in L2, we're looking at IMMENSE performance increases.

tptacek13• 14 years ago

vectors don't deal with the heap? Huh? Yes they do.

astrodust3• 14 years ago

I think the comment was a bad phrasing of "a vector is a contiguous memory structure, whereas a list is a series of independent allocations at various points in the heap".

Daniel_Newby• 14 years ago

I think fein meant heap allocators and deallocators, which vectors rarely use if you are careful.

wheels2• 14 years ago

I stopped reading after the first example, which once again (as in the last post in the series) demonstrated that the author doesn't really know what he's talking about. The code would be written as:

  std::list<person>

Not:

  std::list<person *>

...which avoids the double malloc and presents a nicer API to boot. Not to mention that in that case, a C++ programmer would either use a struct or would need to write accessors.

C++ has a lot of warts, and the author seems to have written some non-trivial software in it, but his critiques of it reveal that there are a lot of gaps in his knowledge of the language and a lot of his criticisms are based on falsehoods. (However, a totally valid criticism of C++ is the amount of idiosyncrasies the language has and how relatively difficult it is to master.)

epsylon• 14 years ago

In practice lists have such poor memory locality that they should almost never be used. Use vectors.

EDIT: See for example Stroustrup's discussion on the matter at the 44th minute of this video : http://channel9.msdn.com/Events/GoingNative/GoingNative-2012...

On modern architectures, cache misses are performance killers, and a list is a cache-miss machine. Moving memory is cheaper than doing a linear search through a list. What used to be true 20 years ago isn't true anymore, let's stop using those lists already !

rumcajz14• 14 years ago

With vectors erase is O(1) anyway.

luriel2• 14 years ago

That most C++ programmers are so confused about how to properly use their language says quite a lot about C++ itself.

tptacek13• 14 years ago

Is your point that programmers in other languages automatically make good choices about data structures and memory usage? Because that seems, respectfully, to be a dumb point. They do not.

forgotusername• 14 years ago

A reasonable response, although the problem with this is that it's impossible to manipulate persons without causing their values to be copied (which might not even be possible given the other members of the type - which might not even be under the author's control, e.g. 3rd party libs). Additionally without extra complexity, simple pointer swapping tasks like moving a person between list suddenly starts to involve memory allocation, etc.

agwa4• 14 years ago

std::list<T>::splice() lets you move elements between lists without copies or allocations.

Before C++11, it was a problem that inserting an element into a std::list required a copy (that's the only place though - once in the list there are no more copies). C++11 adds move semantics and also std::list<T>::emplace(), which eliminate the need to copy.

justinsb3• 14 years ago

Right, but the same trade-offs apply to C; C++ is actually superior here: if something shouldn't be copied, you can enforce that in C++, but not in C.

The author was arguing that writing in C++/STL meant your code was necessarily less efficient than C; that just seems incorrect.

Edit: If you're referring to copying the person values when they're returned from the list: a reference is returned; the compiler will optimize that "pointer" away. Most C++ compilers can even do this when the value itself is returned and avoid the copy. My rule with C++ is that the compiler is stupid, but never where you expect. You need to optimize based on a profiler.

430gj9j• 14 years ago

No copying here:

  struct person { int age; };
  std::list<person> people;
  people.push_back(person());
  people.back().age = 42;

tptacek13• 14 years ago

I had to fight the urge to stop there too; it's also more idiomatic and probably both more economic and performant to use a vector in this situation; "lists" in C++, at least on the projects I worked on, tended to be a code smell.

I think this is a very big problem with C++: most of the funnel of new developers for C++ are C programmers, and transliterating C idioms to C++ (and, worse, the C++ standard library) usually produces pessimal C++ code.

ajross• 14 years ago

To the first part: I guess it's true that if you see someone using a STL list for "just a bucket of stuff" that this is bad form. But they're general purpose doubly-linked lists and absolutely should be there in the API for situations where they're needed (i.e. something where order is important, the size is likely to be large, and you need to mutate in the middle).

The second point I think is just isomorphic to "breadth is good". A Ruby or Python programmer coming to C++ is going to avoid a lot of the mistakes of the C programmer, but will likewise miss a lot of the core ideas and generate equally pessimal code, just for different reasons.

itsboring3• 14 years ago

I've been putting myself through C++ "boot camp". I got a copy of Bjarne's "The C++ Programming Language" and I'm doing every exercise. I'm up to 302 at last count, and only a couple more chapters to go.

It's been grueling, but very well worth the time. It helps me avoid a lot of mistakes I would have made had I just jumped in from C and C#.

Unfortunately, the 4th edition (which presumably covers C++11) doesn't come out until Feb.

dkersten10• 14 years ago

I've been programming in C++ for about ten years, first as a hobbyist, then as a student and finally professionally and it is extremely rare that I've seen an std::list needed over an std::vector or std::deque.

The general rules I follow are roughly: If the collection only grows one way and has lots of random access (or even lots of iterating over elements... linked lists aren't exactly the most cache friendly data structures), then use a vector. If the collection needs to grow from both ends and has lots of random access, then use a deque. Only if neither of those are suitable, should a list be considered and the main reason, IMHO, to use an std::list is if inserting or removing elements in the middle of the list is a common operation. But in the case of removing, if copying is cheap and order does not matter, then a vector may still be better and removal can be done by swapping the element to remove with the last element and then removing the last element. In general, if ordering is not important, a std::vector is probably a good choice.

So really, I see the use case for std::list as a real but very rare one.

JoachimSchipper2• 14 years ago

Are linked lists good C, though? Arrays with some double-on-full code - that is, vectors - perform much better and are often easier to work with.

rumcajz14• 14 years ago

That works only for simple datatypes. Any not trivial object (containing large amount of memory, handles, file descriptors etc.) is non-copyable. Thus list<person> won't work (according to STL, "person" has to be model of Assignable concept.)

dkhenry• 14 years ago

before C++11 you would use <person *> to prevent additional allocations on insertion. C++11 is a little bit smarter about not wasting allocations for temporary objects, but requires you to have move operators for any classes you make. In C land no one cares since there are no temporary objects to even worry about, you would be passing around the memory anyway.

Maybe you should read the rest of his article.

agwa4• 14 years ago

> requires you to have move operators for any classes you make

Actually, in the specific case of std::list<T>, T can be both unmovable and uncopyable. Instead of using push_back() and insert() to insert elements, you use emplace_back(), and emplace() to construct the elements in place.

Also, the compiler is pretty good about auto-generating move constructors when all the members of a class are safely movable, so classes you write will probably be movable without you needing to do anything.

qdog2• 14 years ago

Just the discussion of std::list<person> vs std::list<person *> that follows your comment makes my head hurt. It's like not only is c++ a language from a committee, any line of c++ code could be analyzed and argued over by a committee.

I think I'll tend to stick to C most of the time, thanks.

Even the C++ people probably can easily analyze and reason out the C implementation without much effort or dissent, apparently not so with C++.

pjkundert• 14 years ago

This is the argument that comes up every time a C vs. C++ discussion occurs. It is faulty, because it assumes a false premise -- that the C and C++ implementations will be largely of the same complexity.

Unfortunately, in the C implementation of anything beyond a trivial application with no large-scale data structures, the majority of the "reasoning" being done will revolve around basic data structure implementations! The very containers, memory allocation structures, and such that require virtually NO reasoning in the C++ implementation, because they are proven.

After you've implemented the same basic structures again, and again for the hundredth time in large-scale C programs, C++ is a breath of fresh air.

Unfortunately, many people approach C++ before they have many miles on their programming chassis; this is like handing an infant a revolver. Bad things are likely to happen.

This is not a problem with the revolver; it is a parenting issue. Don't let a junior programmer use a language requiring significant reasoning skills, before they have earned the right -- by developing those skills, repeatedly implementing the very structures they'll be accessing in the C++ implementation!

We wouldn't let a first-year civil engineering student develop a bridge design, without first gaining many, many years of experience developing -- and failing -- on much simpler designs. Why should we expect wide-spread success, allowing junior programmers access to a spectacularly powerful and complex programming tool?

batista2• 14 years ago

>I stopped reading after the first example

I stopped reading your comment at "I stopped reading" too.

I've _never_ read an intelligent comment starting with "I stopped reading". These kind of comments always try to show how "clever" the commenter is, and how he doesn't tolerate fools, etc, while 90% of the time they miss the point.

Have you continued, you would have seen:

"Assume that the object to be contained in the list is non-Assignable as is the case with any non-trivial objects, for example those holding large memory buffers, file descriptors, handles etc. If the object is Assignable simple std::list<person> would do and there is no problem."

And even if you read the post _before_ he added the above clarification, if you didn't have the knee jerk reaction you could probably have seen that this (object being non assignable) could be a potential problem in your solution too.

wheels2• 14 years ago

He did add that clarification afterwards; that was before the first example. I read up to the first example.

However, his clarification, and your proviso are also invalid.

First, your proviso, that it's non-assignable -- his type was unusable; that it's non-assignable simply seemed to be an error. In C++ the default permissions for a class are private (not so in a struct, as I hinted at in my post, in fact, such is the only difference between a class and a struct in C++).

Now, onto his clarification:

The baked in assumption is that in a non-trivial C project that one would not be using a standard set of container abstractions as well. Such is generally not the case. See, for example, GList [1]. The replacement for a template based container in C++ is generally going to be a macro based container abstraction in C, whereas the replacement for a container of pointers in C++ is going to be a void * based generic container implementation (like GList) in C. Both sets of generic implementations are going to have broadly the same performance characteristics (specifically, they're going to have the same number of mallocs of about the same size).

In both languages, if one needs specific performance characteristics -- specifically if mallocs are so critical that they must be restricted as much as possible -- one will use a special purpose data structure. That however is a somewhat rare case for structs which are composed of non-POD types.

[1] http://developer.gnome.org/glib/2.33/glib-Doubly-Linked-List...

okamiueru• 14 years ago

I'm just sitting back with beer in one hand, popcorn in the other, and enjoying the irony show.

jongraehl• 14 years ago

He's not a great C++ programmer. Whether or not it's worth becomming a great C++ programmer is debatable, but it's worked for me.

He could also use http://www.boost.org/doc/libs/1_51_0/doc/html/intrusive/usag... for exactly the "C" layout he wants.

greesil• 14 years ago

Funny, I was going to say the reason to not use C++ was to make sure no one starts bloating up your code size by using boost.

lttlrck• 14 years ago

It might bloat up a code base, and the resulting binary (if you are using it just for the sake of it), and increase the build-time (if you aren't careful/don't know what you are doing), but it definitely doesn't bloat up the size of the code you are writing...

kingkilr• 14 years ago

Ugh, the two solutions to this problem he proposes are totally different. The C version allows a person to only be a member of a single list.

DHowett• 14 years ago

In addition to that, he's effectively comparing apples and steaks in the case of std::list.erase() vs linked-list removal. If you want O(1) removal of a random element, you use a data structure that guarantees O(1) removal of a random element (caveat: not all lists are created equal, and not all lists have O(1) removal.)

That said, it's not even removal that's O(n) here, though he may characterize it this way: it's finding the element to remove, and list searches are nearly always[1] O(n) - even with doubly-linked lists.

[1]: Except in the case of CFArray/NSArray, which is occasionally a non-list masquerading as an array - this is, of course, not a true exception as it's not a true list. (http://ridiculousfish.com/blog/posts/array.html)

ori_b• 14 years ago

> If you want O(1) removal of a random element, you use a data structure that guarantees O(1) removal of a random element

Like std::list.

It's guaranteed that insertion and removal is done in constant time. It does mean that you need an iterator pointing to the element you want to insert in front of ahead of time.

If you want to be more specific, it guarantees is that given a list of elements to insert, the insertion and deletion will be linear in the number of elements inserted or deleted (ie, independent of the size of the list being inserted into; the insertion or removal of each value passed to insert() or erase() is required to be O(1))

(That's one thing that I like about the C++ STL: It usually guarantees a certain complexity for the data structures it provides)

kombine3• 14 years ago

Incapsulation is a very questionable concept as well. However what C++ does provide is the generic programming: in C you will have to copy and paste implementation of List for every data structure as well as the algorithms. As suggested in the comment there are no inherent deficiencies in C++ and there is an implementation of intrusive lists in Boost. You don't have to use incapsulation if you don't want to.

bunderbunder2• 14 years ago

Incapsulation is a very questionable concept as well.

One can live without it on a smaller scale. But for anything that will have any lifetime whatsoever it's absolutely essential. Distinguishing between a module's intended interface and its implementation details is what makes it possible to amend the implementation should the need arise. Without proper encapsulation every change must be assumed to be a breaking change, software maintenance costs go through the roof, module interactions become unpredictable, and that upstart kid who wears Chucks to work but knows how to write testable code eats your Wheaties.

kombine3• 14 years ago

Well, incapsulating algorithms is great. But incapsulating of the data is not always. Languages like Haskell completely go away without it.

maximilianburke• 14 years ago

Anything you can do in C you can do in C++. It is not that the author couldn't implement more memory efficient containers in C++ but that he chose not to even though he was aware of the drawbacks.

justinsb3• 14 years ago

In my opinion, this is just wrong:

1) He uses std::list<person*>, but std::list<person> is the equivalent to his C example, and would produce identical results to C, with much less code. It would also be safer e.g. in terms of memory leaks.

2) Templates are the C++ magic, which get round some of those theoretical limitations of OO programming. The compiler can evaluate the composition of the various objects, to see if any optimizations are possible. With C, this would be have to be done by hand. Yes, it's slower than compiling C, but it's much faster than having a human do it.

3) Making use of #2, it is then easy to swap out the implementation. Want to try a pooled memory allocator (because # of mallocs seems to be the axis we're optimizing on): just change the declared type. Want to try a different data structure (e.g. sort by age for fast lookups by age): just change the declared type. The compiler "expands" the templates, substituting in the various types, and produces code that is reasonably optimized. For example, it will use inline integer comparison for sorting by age, rather than calling out to a comparator function.

I'm sure there are valid reasons to preferring C to C++, but as a user of ZeroMQ, I wish they would spend their efforts documenting their protocol rather than bashing their tools.

VikingCoder2• 14 years ago

1) No, it does not produce identical results.

Most of the STD containers use copying, not pointers, as the article actually points out.

That doesn't make his larger point valid (I disagree with it, actually), but your argument is not technically correct.

justinsb3• 14 years ago

Things that look like they use copying in C++ often don't use copying in the final output. Values are usually returned by reference, and many C++ compilers can often avoid a copy even when that's not the case (e.g. Return Value Optimization).

Even if the compiler doesn't optimize everything away, copying is sometimes faster, always safer and always easier than passing around pointers.

It's a big lesson I learned for C++ - write correct code first; profile; fix any real performance problems, rather than obsessing about my preconceived notions of what would be slow.

agwa4• 14 years ago

To get constant-time erasure of list elements, the author should be referring to the Person elements with a std::list<Person>::iterator instead of a Person* . Passing an iterator to erase gives you constant-time erasure.

An iterator is equivalent to a pointer (* and -> get you a Person), but it also lets the implementation access the linked list node.

jbert2• 14 years ago

If you always reference the person in it's people list (i.e. the object is "owned" by the list), you never need to pass around a ptr to the object of type:

    person *

Instead you can pass an list iterator which references the "person in the list" of type

    std::list<person *>::iterator

Then you don't need to do an O(n) walk of the list to find the entry for erasure, you can just people.erase(it) directly.

rumcajz14• 14 years ago

I've modified the text to address this comment:

EDIT: A lot of people point out that iterator should be used instead of pointer. However, imagine the object is contained in 10 different lists. You would have to pass structure containing 10 iterators around instead of the pointer. Morever, it doesn't solve the encapsulation problem, just moves it elsewhere. Instead of modifying "person" object every time you would want to add it to a new type of container you would have to modify the iterator tuple structure.

plorkyeran• 14 years ago

The C version doesn't support having an object in multiple lists at all, so having it merely be awkward to have an object in multiple lists in the C++ version is still an advantage.

jbert2• 14 years ago

If the object is contained in 10 different lists, wouldn't the C version have 20 ptrs (10x prev and next)? That seems to me to be roughly equivalent complexity to your proposed "10x iterator" in C++, if I've understood correctly.

Basically, something needs to track your references if you have multiple lists. With C prev/next ptrs, the tracking is explicit in the object (well, the adjacent objects in the lists, to be more precise). With C++ containers, the tracking is iterator based.

With C++ you also have the option of using a smart pointer to do your usage tracking, which is probably simpler than either of the two approaches.

vampirechicken2• 14 years ago

This is a well known trade-off - your data structures affect the design and therefore run time profile of your algorithm.

A tightly coupled data structure is more efficient in time and memory, but harder to reuse, and makes it harder modify the code. A more general data structure is easier to reuse and makes it easier to modify the code, at the expense of time and memory at run time.

We use a combination of Moore's Law, and the cost of a programmer to justify the less efficient general solution in the near term, and explain that we'll optimize hot spots later. This mostly works. The Author has found himself lamenting this choice in piece of high-performance software.

If he has few lists, then coding the doubly-linked list pointer directly in to his structs is the way to go. If he has many lists, I'd suggest using the cpp or m4 (or some other templating tool) to statically generate doubly-linked lists at complile time.

rumcajz14• 14 years ago

My point was that the trade-off is not something unavoidable. A better language could allow you to syntactically decouple the tightly coupled objects. See the example at the end of the article.

dkersten10• 14 years ago

A better language could allow you to...

And yet C is not this language.

vampirechicken2• 14 years ago

I was actually agreeing with you, and elaborating on what you talked about.

I went back and read the bottom of the article. The Design Pattern people might call that a decorator. The ruby folks would use a mixin to monkey patch the person class. You're in good company.

The trade-offs are still code complexity + programmer time vs. memory use + run time. The more general you make something, the more overhead at run time.

aidenn02• 14 years ago

Okay, I hate C++ as much as the next guy, but this is ridiculous. There are dozens of ways to get similar code to what the author mentions.

So lets say a decent C++ programmer writes it with std::list<person *> then does some profiling and determines that heap fragmentation is becoming an issue. They can then refactor the code to use a more intrusive list type.

A crappy C++ programmer won't even know its a problem, but presumably we are comparing programmers of similar expertise?

qdog2• 14 years ago

A relatively new C programmer would probably come up with the C code. From the discussion of this topic it appears the C++ solution needs quite a bit more expertise to get right (I would not have known what exact C++ method to use, but I don't pretend to be a C++ expert at this point).

Hmm, actually I would expect most of the C++ methods not to have a problem with heap fragmentation, if they are copying the list when manipulating. The C one doesn't know when it will alloc a new bit of memory, so you might potentially have a cache miss for every single node in the C list, making traversal the worst case scenario.

X-Istence• 14 years ago

What happened to using:

  mylist->remove(object);

Let the underlying algorithm take care of it. No need to write out the loop. Also, this makes it easy to then switch to std::vector or a hash based list or something else where walking the list may not be required.

pmr_3• 14 years ago

What strikes me is the assumption that C++ is an object-oriented language. Yes, there is some support for the paradigm but it is neither great nor is it the main focus of the language. Developers have come up with reusable intrusive data-structure solutions, well knowing that they are bad for encapsulation and they are widely accepted. Most C++ developers wouldn't even shrug when you write a C-style list (in absence of a good library) precisely for the reasons stated in his article.

tjaerv• 14 years ago

This brings to mind Alan Kay's quip that "I made up the term 'object-oriented', and I can tell you I did not have C++ in mind."

pmr_3• 14 years ago

I was thinking the same after re-reading my post. I always thought it was a negative statement but it simply confirms what Stroustrup has always been saying about C++: "It's a multiparadigm programming language." Surely the support for one or the other paradigm could be better but this just reflects that things are shifting away a little from pure object-orientedness across languages.

batista2• 14 years ago

Really? Because it started as "C with classes" IIRC. That you can used different paradigms, from procedural a la C to template-based programming etc, doesn't mean that it's not OO, or that it's not used for OOP 99% of the time.

pmr_3• 14 years ago

Yes, that's how it started. But C with classes (or even C++ before standardization) is a really different language from what is there now.

As I said in another child of my post: It is a multiparadigm language. Of course a lot of people are using it for OOP. I don't dare to say what is the prevalent style of C++ nowadays. I don't know and I don't think anybody really does given the widespread use and different domains.

YZF2• 14 years ago

std::list (used correctly) is IMHO a great example of why you should write in C++, not in C.

Most large C projects I've seen end up with multiple implementations of lists. You will have some combination of: void* for a generic list, performance issues because everything is a function call to another C file which won't get inlined, memory leaks, subtle bugs, thread safety issues because it's often unclear what guarantees the homebrew version provides.

Do you want to spend your time on reinventing the list or on things that provide value?

(Edited for formatting)

mangamadaiyan• 14 years ago

Hrm. Check out sys/queue.h in any BSD (Linux too - though Linux kernel code IIRC uses a different list implementation). You don't necessarily need void * for a generic list, and you most certainly don't have to reinvent the wheel.

YZF2• 14 years ago

queue.h is an improvement. You still need to somehow know this is in BSD and you'll need to copy it to build on different systems.

The thing is that in C, most people do reinvent the wheel, including the person in the original article and as you say, the Linux kernel. That's been my experience.

std::list is standard and cleaner IMO.

mkhalil• 14 years ago

For a person with OCD, this title really bothers me :/ I want it to say "Why I should have written ZeroMQ in C, not C++ (part II)".

akkartik2• 14 years ago

I'm amazed nobody is talking about his 'private in' idea.

tedunangst5• 14 years ago

Perhaps someone could explain how it differs and compares to the friend keyword.

akkartik2• 14 years ago

It has more to do with layout than with access control. Fields belonging to one class get stored and laid out with another object. They're accessed as part of the other object, but the locks they acquire are for the original object.

I see faint echoes of ruby's open classes if I squint..

rumcajz14• 14 years ago

Friend only allows others to access your private members. It doesn't allow you to define members in other classes though.

scott_s6• 14 years ago

I did; we can achieve the same functionality through mixins: http://news.ycombinator.com/item?id=4455823

jbp• 14 years ago

At least for 2nd problem mentioned wouldn't

  std::unordered_map<person *, person *>

make sense?

ChristianMarks• 14 years ago

The level of disagreement and misunderstanding surrounding C++, its typing and implementation is a sign that the language and its community openly welcome extraneous cognitive load. Or its a sign of my own confirmation bias or both (or neither). Other languages have their quirks as well, but C++ stands out for me. I haven't done a study--perhaps there are studies on this, so my opinion amounts to so much line noise.

Why should I have written ZeroMQ in C, not C++ (part II)

Related Stories