0 Members and 3 Guests are viewing this topic.
How can people overuse vectors or iterators in C++? STL containers are more or less meant to be used instead of old-fashioned plain array data structures, and iterators are meant to be used for iterating through them. It's kind of like std::string vs. C-string thing; in C++ using good ol' C strings is more or less considered bad practice(error prone, not always clear, POINTERS yuck!) because you've got std::string. Of course old C strings are perfectly valid, but I'd guess that it's more because of better backward compatibility with C and it's legacy rather than as an practical language feature.
The results stay the same even after swapping them with GCC 4.45, that is not the reason. As I explained they boil down to simple pointers, THAT is the reason why they're faster.
for (int* i = va; i < &va[100000000]; ++i) { *i += 10; // do something so loop doesn't get optimized away }
And you can easily change your code to:Code: [Select] for (int* i = va; i < &va[100000000]; ++i) { *i += 10; // do something so loop doesn't get optimized away }I find this more readable than the iterator code.But it is all personal preference, so don't take what I say as argument, just as discussion
for (auto i = v.begin(); i < v.end(); i++) { ... }
// Result stored here vector<std::string> result; // Give return result; // Store actual servers here vector<server_entry> servers;
Will auto acquire some new usefulness in C++0x?
auto func = [](int x, int y) { return x + y; };
Quote from: pubby8 on April 09, 2011, 01:36:48 pmAnd you can easily change your code to:Code: [Select] for (int* i = va; i < &va[100000000]; ++i) { *i += 10; // do something so loop doesn't get optimized away }I find this more readable than the iterator code.But it is all personal preference, so don't take what I say as argument, just as discussion That is just terrible, no offense. No C++ programmer in their right mind would code like that, that looks like something a C programmer that just started C++ would do. And while we're on the subject of aesthetics, in C++0x you can actually do this with iterators, which is much nicer:
for (int& x: va) { x += 10; // we're iterating by reference}
2) That is actually defined as undefined behavior per standard when you run that code; you're accessing the array outside of its boundaries. It most likely works on most compilers, but the fact is that accessing outside the array's boundary. The proper way would be i < &va[0]+100000000 or i <= &va[99999999].
3) You're using a magic number, you probably meant to use va.size() instead of that 100000000.
4) In case you want it, you're not taking advantage of your compiler's some features, for instance MSVC has a feature called checked iterators that add extra security to them. This is arguable though, since checked iterators actually have a bit of an overhead to them, so if you care more about speed than safety, you would turn these off anyways.
So basically, iterators are generally safer and more consistent. If you get used to iterators, like you should, then you don't need to guess "how should I iterate this type of container", since every container would be iterated through similarly. There's no other way to iterate through some container's than the standard library iterators, take for example a list or a set or a map. Get used to the iterators everywhere.
Oh, I think there is misunderstanding.That code was being used on a normal array, not on a vector (that is why I renamed it "va" (a for array) and not "v")
I see what you are saying, but &va[0]+100000000 could also be problematic if it happened to overflow (very very unlikely)
Here is another reason why I dislike vectors - they store additional values such as size and reserved size. 90% of arrays I need are a constant size. If it was actual code, of course there wouldn't be a magic number in there, but it was a test snippet so I hope you understand the purpose of it was not about use of magic numbers.
If iterators are so safe and consistent, why are there "checked iterators" then?
You should be thinking about "how should I iterate this container," as it can be very critical to your performance. Especially since STL wasn't designed for threading.
I've used the STL a great deal a while ago, but I found it limiting. I rewrote my code that was using it and actually got ~7 times performance improvement. It still remained readable, it was largely OO, and I didn't have to write too much code. You don't need to have some oversimplified helper library to hold your hand to get stuff done in a good way.
That obviously depends on the size of the array.. If it has 100000000 elements, then doing &va[0]+100000000 is same as &va[100000000], there's no difference, it is not "very very unlikely" that one would overflow but other would not. I don't understand how this is relevant to the discussion at hand in anyway.
In terms of additional overhead things like keeping track of the size; it's not much. What's expensive though is that if the vector goes over its reserved capacity, since it'll need to copy the data to a new location in that case (C++0x has a nifty feature called move semantics, that make copying larger objects that are being destructed after the copy much faster!) Once again a good C++ programmer is aware of this, and knows to reserve space for his vector when necessary.
They're mostly useful for finding bugs in your code. For example normally if you try to use an uninitialized iterator or one that has gone beyond the bounds of the container you'll get undefined behavior; you might spot the error or you might not. It might not happen for you, but it might happen to some of your users. With checked iterators you catch bugs like those much easier in your code, as they're checked and exceptions are thrown (it's mainly useful to keep on in debugging builds, for release builds and maximal speed you usually want to turn it off if you want the extra speed).
I don't believe that. I would like to see some benchmarks and code where you custom written containers/code is 7 times faster. What is more likely is that you misused the standard library and used wrong containers for wrong things (it's essential to understand how each container works so you can use it for the purpose it was meant for), or perhaps you did your benchmarks using debug builds instead of release builds (as I demonstrated before, the containers/iterators are quite a bit faster in release builds due to optimizations taking place). Also note that the C++0x has new containers that are better for certain things, such as std::unordered_map (basically a hashmap), before C++0x most people prefer to use boost to get access to these useful containers (why reinvent the wheel right? boost containers follow the format of stdlib containers as in they're iteretable and compatible with the stdlib algorithms, and they are thoroughly tested and optimized). But I'm willing to be proven wrong, but I don't take your word for it. I want to see code and benchmarks, I'm sure you understand my skepticism of you being able to write ~7 times faster code than the people who write these compilers for living and know C++ and inside and out, plus how to optimize their code for their own compiler.
While vector obviously isn't 7 times slower, stuff like list and map are not optimal in most cases. One algorithm I got big speed improvements was a LRU caching one - I originally used a list, and also tried a map+list, but by rewriting it using a custom hash table that uses a linked list with 4 pointers per node, and a custom stack-based allocator, I got massive improvements.
You can't really compare that. You wrote an entirely different kind of container and compared it to another kind of container.
Btw std::map insertions are a lot faster if you know where in the container you insert it, and give a hint iterator to the .insert() function, and by a lot I mean it can easily be up to 5-10 times faster, or even more, depending on the size of the map.
