提升STL使用效率的50个具体策略

需积分: 29 148 浏览量更新于2024-07-28 收藏 1.46MB PDF 举报

"《Effective STL：50个具体方法改进STL使用》是计算机科学领域的一本经典著作，作者通过50个具体的建议，详细阐述了如何更有效地使用STL（Standard Template Library，标准模板库）。书中涵盖了STL中的容器、迭代器、算法和分配器等方面，旨在帮助读者提升对STL的理解和应用能力。" 以下是书中的部分知识点详解： 1. **选择合适的容器**：不同的STL容器有不同的特性和用途，如vector适合动态增长数组，list适合频繁插入和删除元素，map提供键值对映射。选择时应根据实际需求来决定。 2. **避免容器独立代码的幻觉**：尽管STL提供了统一的接口，但不同容器的实现和性能差异可能导致代码优化上的问题，因此编写代码时要考虑特定容器的特性。 3. **使容器内的对象复制廉价且正确**：确保对象复制操作有效率，并且在容器复制时能保持数据一致性。 4. **调用empty而非检查size()是否为零**：直接使用empty()可以更清晰地表达意图，且可能具有更好的性能。 5. **优先使用范围成员函数**：如`for_each`，`find_if`等，它们通常比单元素版本更高效，因为它们可以利用迭代器范围来减少边界检查。 6. **警惕C++的最令人困扰的解析**：避免因语法歧义导致的错误，例如构造函数与普通函数之间的混淆。 7. **处理new创建的指针容器**：当使用容器存储new创建的指针时，记得在容器销毁前释放这些指针，以防止内存泄漏。 8. **不要在容器中使用auto_ptr**：auto_ptr不适用于容器，因为它不遵循STL的常规所有权规则，可能导致意外的行为。 9. **谨慎选择删除选项**：在容器中删除元素时，了解不同方式（如erase，clear）的差异和影响。 10. **理解分配器的约定和限制**：分配器影响内存管理，了解其工作原理有助于优化STL容器的性能。 11. **定制分配器的合理使用**：在特殊场景下，如内存池或特定硬件需求，可能需要自定义分配器，但需谨慎，因为它们增加了复杂性。 12. **对STL容器的线程安全性有现实期望**：默认情况下，STL容器不是线程安全的，需要额外的同步机制来保证并发访问的安全。 13. **推荐使用vector和string**：这两个容器在大多数情况下优于动态分配的数组，因为它们提供了更多的便利功能和优化。 14. **使用reserve避免不必要的重分配**：预分配足够的空间可以避免vector或string在插入时的动态扩展，从而提高效率。 15. **注意string实现的差异**：不同编译器的string实现可能略有不同，要考虑到这些差异可能带来的兼容性问题。 16. **将vector和string数据传递给遗留API**：了解如何正确地将这些容器的数据传递给不支持STL的旧式API。 17. **使用“交换技巧”修剪过剩容量**：通过交换技巧，可以快速释放vector或string的多余容量，避免浪费内存。这些知识点揭示了STL使用中的最佳实践，帮助开发者编写出更高效、更易于维护的代码。

Compared to arrays. STL containers are much more civilized. They create (by

copying) only as many objects as you ask for, they do it only when you direct them to,

and they use a default constructor only when you say they should. Yes, STL containers

make copies, and yes, you need to understand that, but don't lose sight of the fact that

they're still a big step up from arrays.

Item 4. Call empty instead of checking size() against zero.

For any container c, writing

if (c.size() == 0)...

is essentially equivalent to writing

if (c.empty())...

That being the case, you might wonder why one construct should be preferred to the

other, especially in view of the fact that empty is typically implemented as an inline

function that simply returns whether size returns 0.

You should prefer the construct using empty, and the reason is simple: empty is a

constant-time operation for all standard containers, but for some list implementations,

size takes linear time.

But what makes list so troublesome? Why can't it, too. offer a constant-time size? The

answer has much to do with list's unique splicing functions. Consider this code:

list<int> list1;

list<int> Iist2;

...

list1.splice( // move ail nodes in Iist2

list1.end(), Iist2, // from the first occurrence

find(list2.begin(), Iist2.end(), 5), // of 5 through the last

find(list2.rbegin(), Iist2.rend(), 10).base() // occurrence of 10 to the

); //end of list1. See Item 28

// for info on the "base()" call

This code won't work unless Iist2 contains a 10 somewhere beyond a 5, but let's

assume that's not a problem. Instead, let's focus on this question: how many elements

are in list1 after the splice? Clearly, list1 after the splice has as many elements as it did

before the splice plus however many elements were spliced into it. But how many

elements were spliced into it? As many as were in the range defined by

find(list2.begin(), Iist2.end(), 5) and find(list2.rbegin(), Iist2.rend(), 10).base(). Okay,

how many is that? Without traversing the range and counting them, there's no way to

know. And therein lies the problem.

Suppose you're responsible for implementing list, list isn't just any container, it's a

standard container, so you know your class will be widely used. You naturally want

your implementation to be as efficient as possible. You figure that clients will

commonly want to find out how many elements are in a list, so you'd like to make size

a constant- time operation. You'd thus like to design list so it always knows how many

elements it contains.

At the same time, you know that of all the standard containers, only list offers the

ability to splice elements from one place to another without copying any data. You

reason that many list clients will choose list specifically because it offers high-

efficiency splicing. They know that splicing a range from one list to another can be

accomplished in constant time, and you know that they know it, so you certainly want

to meet their expectation that splice is a constant-time member function.

This puts you in a quandary. If size is to be a constant-time operation, each list

member function must update the sizes of the lists on which it operates. That includes

splice. But the only way for splice to update the sizes of the lists it modifies is for it to

count the number of elements being spliced, and doing that would prevent splice from

achieving the constant-time performance you want for it. If you eliminate the

requirement that splice update the sizes of the lists it's modifying, splice can be made

constant-time, but then size becomes a linear-time operation. In general, it will have to

traverse its entire data structure to see how many elements it contains. No matter how

you look at it, something — size or splice — has to give. One or the other can be a

constant-time operation, but not both.

Different list implementations resolve this conflict in different ways, depending on

whether their authors choose to maximize the efficiency of size or splice. If you

happen to be using a list implementation where a constant-time splice was given

higher priority than a constant-time size, you'll be better off calling empty than size,

because empty is always a constant-time operation. Even if you're not using such an

implementation, you might find yourself using such an implementation in the future.

For example, you might port your code to a different platform where a different

implementation of the STL is available, or you might just decide to switch to a

different STL implementation for your current platform.

No matter what happens, you can't go wrong if you call empty instead of checking to

see if size() == 0. So call empty whenever you need to know whether a container has

zero elements.

Item 5. Prefer range member functions to their single-element

counterparts.

Quick! Given two vectors, v1 and v2, what's the easiest way to make v1’s contents be

the same as the second half of v2's? Don't agonize over the definition of "half when v2

has an odd number of elements, just do something reasonable.

Time's up! If your answer was

v1.assign(v2.begin() + v2.size() /2, v2.end());

or something quite similar, you get full credit and a gold star. If your answer involved

more than one function call, but didn't use any kind of loop, you get nearly full credit,

but no gold star. If your answer involved a loop, you've got some room for

improvement, and if your answer involved multiple loops, well, let's just say that you

really need this book.

By the way, if your response to the answer to the question included "Huh?", pay close

attention, because you're going to learn something really useful.

This quiz is designed to do two things. First, it affords me an opportunity to remind

you of the existence of the assign member function, a convenient beast that too many

programmers overlook. It's available for all the standard sequence containers (vector,

string, deque, and list). Whenever you have to completely replace the contents of a

container, you should think of assignment. If you're just copying one container to

another of the same type, operator= is the assignment function of choice, but as this

example demonstrates, assign is available for the times when you want to give a

container a completely new set of values, but operator= won't do what you want.

The second reason for the quiz is to demonstrate why range member functions are

superior to their single-element alternatives. A range member function is a member

function that, like STL algorithms, uses two iterator parameters to specify a range of

elements over which something should be done. Without using a range member

function to solve this Item's opening problem, you'd have to write an explicit loop,

probably something like this:

vector<Widget> v1, v2; // assume v1 and v2 are vectors

//of Widgets

v1.clear();

for ( vector<Widget>::const_iterator ci = v2.begin() + v2.size() / 2;

ci != v2.end();

++ci)

v1.push_back(*ci);

Item 43 examines in detail why you should try to avoid writing explicit loops, but you

don't need to read that Item to recognize that writing this code is a lot more work than

is writing the call to assign. As we'll see shortly, the loop also happens to impose an

efficiency penalty, but we'll deal with that in a moment.

One way to avoid the loop is to follow the advice of Item 43 and employ an algorithm

instead:

v1.clear();

copy(v2.begin() + v2.size() / 2, v2.end(), back_inserter(v1 ));

Writing this is still more work than writing the call to assign. Furthermore, though no

loop is present in this code, one certainly exists inside copy (see Item 43). As a result,

the efficiency penalty remains. Again. I'll discuss that below. At this point, I want to

digress briefly to observe that almost all uses of copy where the destination range is

specified using an insert iterator (i.e.. via inserter, back_inserter. or front_inserter) can

be — should be — replaced with calls to range member functions. Here, for example,

the call to copy can be replaced with a range version of insert:

v1 .insert(v1 .end(), v2.begin() + v2.size() / 2, v2.end());

This involves slightly less typing than the call to copy, but it also says more directly

what is happening: data is being inserted into v1. The call to copy expresses that, too,

but less directly. It puts the emphasis in the wrong place. The interesting aspect of

what is happening is not that elements are being copied, it's that v1 is having new data

added to it. The insert member function makes that clear. The use of copy obscures it.

There's nothing interesting about the fact that things are being copied, because the STL

is built on the assumption that things will be copied. Copying is so fundamental to the

STL. it's the topic of Item 3 in this book!

Too many STL programmers overuse copy, so the advice I just gave bears repeating:

Almost all uses of copy where the destination range is specified using an insert iterator

should be replaced with calls to range member functions.

•

Returning to our assign example, we've already identified two reasons to prefer

range member functions to their single-element counterparts:

•

It's generally less work to write the code using the range member functions.

•

Range member functions tend to lead to code that is clearer and more

straightforward.

In short, range member functions yield code that is easier to write and easier to

understand. What's not to like'.'

Alas, some will dismiss these arguments as matters of programming style, and

developers enjoy arguing about style issues almost as much as they enjoy arguing

about which is the One True Editor. (As if there's any doubt. It's Emacs.) It would be

helpful to have a more universally agreed-upon criterion for establishing the

superiority of range member functions to their single-element counterparts. For the

standard sequence containers, we have one: efficiency. When dealing with the

standard sequence containers, application of single-element member functions makes

more demands on memory allocators, copies objects more frequently, and/or performs

redundant operations compared to range member functions that achieve the same end.

For example, suppose you'd like to copy an array of ints into the front of a vector. (The

data might be in an array instead of a vector in the first place, because the data came

from a legacy C API. For a discussion of the issues that arise when mixing STL

containers and C APIs, see Item 16.) Using the vector range insert function, it's

honestly trivial:

int data[numValues]; // assume numValues is

// defined elsewhere

vector<int> v;

...

v.insert(v.begin(), data, data + numValues); // insert the ints in data

//into v at the front

Using iterative calls to insert in an explicit loop, it would probably look more or less

like this:

vector<int>::iterator insertLoc(v.begin());

for (int i = 0; i < numValues; ++i) {

insertLoc = v.insert(insertLoc, data[i]);

}

Notice how we have to be careful to save the return value of insert for the next loop

iteration. If we didn't update insertLoc after each insertion, we'd have two problems.

First, all loop iterations after the first would yield undefined behavior, because each

insert call would invalidate insertLoc. Second, even if insertLoc remained valid, we'd

always insert at the front of the vector (i.e., at v.begin()), and the result would be that

the ints copied into v would end up in reverse order.

If we follow the lead of Item 43 and replace the loop with a call to copy, we get

something like this:

copy(data. data + numValues, inserter(v, v.begin()));

By the time the copy template has been instantiated, the code based on copy and the

code using the explicit loop will be almost identical, so for purposes of an efficiency

analysis, we'll focus on the explicit loop, keeping in mind that the analysis is equally

valid for the code employing copy. Looking at the explicit loop just makes it easier to

understand where the efficiency hits come from. Yes, that's "hits." plural, because the

code using the single-element version of insert levies up to three different performance

taxes on you, none of which you pay if you use the range version of insert.

The first tax consists of unnecessary function calls. Inserting numValues elements into

v one at a time naturally costs you numValues calls to insert. Using the range form of

insert, you pay for only one function call, a savings of numValues-1 calls. Of course,

it's possible that inlining will save you from this tax, but then again, it's possible that it

won't. Only one thing is sure. With the range form of insert, you definitely won't pay

it.

Inlining won't save you from the second tax, which is the cost of inefficiently moving

the existing elements in v to their final post-insertion positions. Each time insert is

called to add a new value to v. every element above the insertion point must be moved

up one position to make room for the new element. So the element at position p must

be moved up to position p+1, etc. In our example, we're inserting numValues elements

at the front of v. That means that each element in v prior to the insertions will have to

剩余197页未读，继续阅读

xyjcn85

粉丝: 1

提升STL使用效率的50个具体策略

Effective STL

Effective C++ 3rd Edition 55 Specific Ways to Improve Your Programs and Designs (PDF)

Effective STL 中文版（含中文+英文）（STL专辑）

C++四书五经 - 05. 标准库

C++四书五经 - 07. 杂项-02

C++四书五经 - 07. 杂项-01

提升STL使用技巧：《Effective STL》50个具体建议

Effective STL使用指南：优化标准模板库的50个具体方法

提升STL使用效率的50个技巧

提升STL使用效率的50个具体方法

最新资源