不要做过早的假设，因为其他应用程序代码和使用可能会影响您的功能。因此，孤立地看待函数是没有用的。

如果你是认真的应用程序，那么VTune它或使用任何类似的分析工具，并查看热点。

糯米

您可以为特定大小的对象编写一个非常高性能的特殊堆分配器。但是，一般的堆分配器性能不是特别好。

我也同意Torbjörn Gyllebring关于对象的预期生命期的观点。好点!

c++语言特有的关注点

首先，c++中没有所谓的“堆栈”或“堆”分配。如果你谈论的是块作用域中的自动对象，它们甚至没有被“分配”。(顺便说一下，C语言中的自动存储时间肯定与“分配”不一样;后者在c++中是“动态的”。)动态分配的内存在自由存储区上，而不一定在“堆”上，尽管后者通常是(默认的)实现。

Although as per the abstract machine semantic rules, automatic objects still occupy memory, a conforming C++ implementation is allowed to ignore this fact when it can prove this does not matter (when it does not change the observable behavior of the program). This permission is granted by the as-if rule in ISO C++, which is also the general clause enabling the usual optimizations (and there is also an almost same rule in ISO C). Besides the as-if rule, ISO C++ also has copy elision rules to allow omission of specific creations of objects. The constructor and destructor calls involved are thereby omitted. As a result, the automatic objects (if any) in these constructors and destructors are also eliminated, compared to naive abstract semantics implied by the source code.

另一方面，免费商店的分配绝对是设计上的“分配”。在ISO c++规则下，这样的分配可以通过调用分配函数来实现。然而，自ISO c++ 14以来，有一个新的(非as-if)规则允许在特定情况下合并全局分配函数(即::operator new)调用。因此，部分动态分配操作也可以像自动对象一样是无操作的。

分配函数用于分配内存资源。可以使用分配器根据分配进一步分配对象。对于自动对象，它们是直接呈现的——尽管底层内存可以被访问，并被用来为其他对象提供内存(通过放置new)，但这作为自由存储没有太大意义，因为没有办法将资源移动到其他地方。

所有其他问题都超出了c++的范围。尽管如此，它们仍然是重要的。

c++的实现

C++ does not expose reified activation records or some sorts of first-class continuations (e.g. by the famous call/cc), there is no way to directly manipulate the activation record frames - where the implementation need to place the automatic objects to. Once there is no (non-portable) interoperations with the underlying implementation ("native" non-portable code, such as inline assembly code), an omission of the underlying allocation of the frames can be quite trivial. For example, when the called function is inlined, the frames can be effectively merged into others, so there is no way to show what is the "allocation".

However, once interops are respected, things are getting complex. A typical implementation of C++ will expose the ability of interop on ISA (instruction-set architecture) with some calling conventions as the binary boundary shared with the native (ISA-level machine) code. This would be explicitly costly, notably, when maintaining the stack pointer, which is often directly held by an ISA-level register (with probably specific machine instructions to access). The stack pointer indicates the boundary of the top frame of the (currently active) function call. When a function call is entered, a new frame is needed and the stack pointer is added or subtracted (depending on the convention of ISA) by a value not less than the required frame size. The frame is then said allocated when the stack pointer after the operations. Parameters of functions may be passed onto the stack frame as well, depending on the calling convention used for the call. The frame can hold the memory of automatic objects (probably including the parameters) specified by the C++ source code. In the sense of such implementations, these objects are "allocated". When the control exits the function call, the frame is no longer needed, it is usually released by restoring the stack pointer back to the state before the call (saved previously according to the calling convention). This can be viewed as "deallocation". These operations make the activation record effectively a LIFO data structure, so it is often called "the (call) stack". The stack pointer effectively indicates the top position of the stack.

Because most C++ implementations (particularly the ones targeting ISA-level native code and using the assembly language as its immediate output) use similar strategies like this, such a confusing "allocation" scheme is popular. Such allocations (as well as deallocations) do spend machine cycles, and it can be expensive when the (non-optimized) calls occur frequently, even though modern CPU microarchitectures can have complex optimizations implemented by hardware for the common code pattern (like using a stack engine in implementing PUSH/POP instructions).

But anyway, in general, it is true that the cost of stack frame allocation is significantly less than a call to an allocation function operating the free store (unless it is totally optimized away), which itself can have hundreds of (if not millions of :-) operations to maintain the stack pointer and other states. Allocation functions are typically based on API provided by the hosted environment (e.g. runtime provided by the OS). Different to the purpose of holding automatic objects for functions calls, such allocations are general-purpose, so they will not have frame structure like a stack. Traditionally, they allocate space from the pool storage called the heap (or several heaps). Different from the "stack", the concept "heap" here does not indicate the data structure being used; it is derived from early language implementations decades ago. (BTW, the call stack is usually allocated with fixed or user-specified size from the heap by the environment in program/thread startup.) The nature of use cases makes allocations and deallocations from a heap far more complicated (than pushing/poppoing of stack frames), and hardly possible to be directly optimized by hardware.

对内存访问的影响

The usual stack allocation always puts the new frame on the top, so it has a quite good locality. This is friendly to the cache. OTOH, memory allocated randomly in the free store has no such property. Since ISO C++17, there are pool resource templates provided by <memory_resource>. The direct purpose of such an interface is to allow the results of consecutive allocations being close together in memory. This acknowledges the fact that this strategy is generally good for performance with contemporary implementations, e.g. being friendly to cache in modern architectures. This is about the performance of access rather than allocation, though.

并发性

Expectation of concurrent access to memory can have different effects between the stack and heaps. A call stack is usually exclusively owned by one thread of execution in a typical C++ implementation. OTOH, heaps are often shared among the threads in a process. For such heaps, the allocation and deallocation functions have to protect the shared internal administrative data structure from the data race. As a result, heap allocations and deallocations may have additional overhead due to internal synchronization operations.

空间效率

由于用例和内部数据结构的性质，堆可能会受到内部内存碎片的影响，而堆栈则不会。这对内存分配的性能没有直接影响，但在虚拟内存的系统中，低空间效率可能会降低内存访问的整体性能。当HDD被用作物理内存交换时，这种情况尤其糟糕。它会导致相当长的延迟——有时是数十亿个周期。

堆栈分配的限制

尽管在现实中，堆栈分配在性能上通常优于堆分配，但这并不意味着堆栈分配总是可以取代堆分配。

首先，在ISO c++中无法以可移植的方式在运行时指定大小的堆栈上分配空间。诸如alloca和g++的VLA(变长数组)等实现提供了扩展，但是有理由避免使用它们。(IIRC, Linux源代码最近删除了VLA的使用)(还要注意ISO C99确实有强制的VLA，但ISO C11是可选的支持。)

Second, there is no reliable and portable way to detect stack space exhaustion. This is often called stack overflow (hmm, the etymology of this site), but probably more accurately, stack overrun. In reality, this often causes invalid memory access, and the state of the program is then corrupted (... or maybe worse, a security hole). In fact, ISO C++ has no concept of "the stack" and makes it undefined behavior when the resource is exhausted. Be cautious about how much room should be left for automatic objects.

如果堆栈空间用完，则堆栈中分配的对象太多，这可能是由于过多的活动函数调用或不恰当地使用自动对象造成的。这种情况可能表明存在错误，例如没有正确退出条件的递归函数调用。

Nevertheless, deep recursive calls are sometimes desired. In implementations of languages requiring support of unbound active calls (where the call depth only limited by total memory), it is impossible to use the (contemporary) native call stack directly as the target language activation record like typical C++ implementations. To work around the problem, alternative ways of the construction of activation records are needed. For example, SML/NJ explicitly allocates frames on the heap and uses cactus stacks. The complicated allocation of such activation record frames is usually not as fast as the call stack frames. However, if such languages are implemented further with the guarantee of proper tail recursion, the direct stack allocation in the object language (that is, the "object" in the language does not stored as references, but native primitive values which can be one-to-one mapped to unshared C++ objects) is even more complicated with more performance penalty in general. When using C++ to implement such languages, it is difficult to estimate the performance impacts.

正如其他人所说，堆栈分配通常要快得多。

但是，如果复制对象的代价很高，那么如果不小心，在堆栈上分配可能会导致以后使用对象时的巨大性能损失。

例如，如果你在堆栈上分配了一些东西，然后将其放入容器中，那么在堆上分配并将指针存储在容器中会更好(例如使用std::shared_ptr<>)。同样的情况也适用于按值传递或返回对象，以及其他类似的情况。

重点是，尽管在许多情况下堆栈分配通常比堆分配更好，但有时如果你在不最适合计算模型的情况下费尽脑汁进行堆栈分配，它可能会导致比它解决的问题更多的问题。

堆栈分配要快得多，因为它所做的只是移动堆栈指针。 使用内存池，您可以从堆分配中获得类似的性能，但这会略微增加复杂性，并带来令人头痛的问题。

此外，堆栈与堆不仅是性能方面的考虑;它还告诉您许多关于对象的预期生存期的信息。

老实说，写一个程序来比较性能是很简单的:

#include <ctime>
#include <iostream>

namespace {
    class empty { }; // even empty classes take up 1 byte of space, minimum
}

int main()
{
    std::clock_t start = std::clock();
    for (int i = 0; i < 100000; ++i)
        empty e;
    std::clock_t duration = std::clock() - start;
    std::cout << "stack allocation took " << duration << " clock ticks\n";
    start = std::clock();
    for (int i = 0; i < 100000; ++i) {
        empty* e = new empty;
        delete e;
    };
    duration = std::clock() - start;
    std::cout << "heap allocation took " << duration << " clock ticks\n";
}

有人说，愚蠢的一致性是小心眼的妖怪。显然，优化编译器是许多程序员心中的妖怪。这个讨论曾经在答案的底部，但人们显然不想读那么远，所以我把它移到这里，以避免遇到我已经回答过的问题。

优化编译器可能会注意到这段代码什么都不做，并可能将其全部优化。这样做是优化器的工作，与优化器斗争是徒劳的。

我建议在编译此代码时关闭优化，因为没有好方法可以欺骗当前正在使用或将来将使用的每个优化器。

任何打开优化器，然后抱怨与它斗争的人都应该受到公众的嘲笑。

如果我关心纳秒精度，我就不会使用std::clock()。如果我想把这些结果作为博士论文发表，我会对此做更大的研究，我可能会比较GCC、Tendra/Ten15、LLVM、Watcom、Borland、Visual c++、Digital Mars、ICC和其他编译器。实际上，堆分配所花费的时间是堆栈分配的数百倍，我认为进一步研究这个问题没有任何用处。

优化器的任务是去除我正在测试的代码。我不认为有任何理由告诉优化器运行，然后试图欺骗优化器不进行实际优化。但如果我看到这样做的价值，我会做以下一项或多项:

Add a data member to empty, and access that data member in the loop; but if I only ever read from the data member the optimizer can do constant folding and remove the loop; if I only ever write to the data member, the optimizer may skip all but the very last iteration of the loop. Additionally, the question wasn't "stack allocation and data access vs. heap allocation and data access." Declare e volatile, but volatile is often compiled incorrectly (PDF). Take the address of e inside the loop (and maybe assign it to a variable that is declared extern and defined in another file). But even in this case, the compiler may notice that -- on the stack at least -- e will always be allocated at the same memory address, and then do constant folding like in (1) above. I get all iterations of the loop, but the object is never actually allocated.

Beyond the obvious, this test is flawed in that it measures both allocation and deallocation, and the original question didn't ask about deallocation. Of course variables allocated on the stack are automatically deallocated at the end of their scope, so not calling delete would (1) skew the numbers (stack deallocation is included in the numbers about stack allocation, so it's only fair to measure heap deallocation) and (2) cause a pretty bad memory leak, unless we keep a reference to the new pointer and call delete after we've got our time measurement.

在我的机器上，在Windows上使用g++ 3.4.4，对于任何小于100000个分配的堆栈和堆分配，我都得到“0个时钟滴答”，即使这样，对于堆栈分配，我也得到“0个时钟滴答”，对于堆分配，我得到“15个时钟滴答”。当我测量10,000,000个分配时，堆栈分配需要31个时钟滴答，堆分配需要1562个时钟滴答。

是的，优化编译器可以省略创建空对象。如果我理解正确的话，它甚至可以省略整个第一个循环。当我将迭代次数增加到10,000,000次时，堆栈分配花费了31个时钟节拍，堆分配花费了1562个时钟节拍。我认为可以肯定地说，在没有告诉g++优化可执行文件的情况下，g++并没有省略构造函数。

在我写这篇文章之后的几年里，Stack Overflow的首选是发布优化构建的性能。总的来说，我认为这是正确的。然而，我仍然认为，当你实际上不希望代码被优化时，让编译器去优化代码是愚蠢的。在我看来，这很像给代客泊车额外付费，却拒绝交出钥匙。在这个特殊情况下，我不希望优化器运行。

使用稍微修改过的基准测试版本(以解决原始程序在每次循环时都没有在堆栈上分配一些东西的问题)，并在不进行优化的情况下编译，但链接到发布库(以解决我们不希望包括任何由于链接到调试库而导致的放缓的问题):

#include <cstdio>
#include <chrono>

namespace {
    void on_stack()
    {
        int i;
    }

    void on_heap()
    {
        int* i = new int;
        delete i;
    }
}

int main()
{
    auto begin = std::chrono::system_clock::now();
    for (int i = 0; i < 1000000000; ++i)
        on_stack();
    auto end = std::chrono::system_clock::now();

    std::printf("on_stack took %f seconds\n", std::chrono::duration<double>(end - begin).count());

    begin = std::chrono::system_clock::now();
    for (int i = 0; i < 1000000000; ++i)
        on_heap();
    end = std::chrono::system_clock::now();

    std::printf("on_heap took %f seconds\n", std::chrono::duration<double>(end - begin).count());
    return 0;
}

显示:

on_stack took 2.070003 seconds
on_heap took 57.980081 seconds

在我的系统上，当用命令行编译cl foo。cc /Od /MT /EHsc。

您可能不同意我获得非优化构建的方法。这很好:您可以随意修改基准测试。当我打开优化，我得到:

on_stack took 0.000000 seconds
on_heap took 51.608723 seconds

这并不是因为堆栈分配实际上是瞬时的，而是因为任何半像样的编译器都能注意到on_stack没有做任何有用的事情，可以进行优化。我的Linux笔记本电脑上的GCC也注意到on_heap没有做任何有用的事情，并优化了它:

on_stack took 0.000003 seconds
on_heap took 0.000002 seconds