为什么Clang和VS2013接受移动大括号初始化的默认参数,但不接受GCC 4.8或4.9?

时间:2014-01-07 21:36:58

标签: c++ gcc c++11 list-initialization

就像标题所暗示的那样,我有一个简短的演示程序可以编译所有这些编译器,但在使用gcc 4.8和gcc 4.9编译后运行核心转储:

关于为什么的任何想法?

#include <unordered_map>

struct Foo : std::unordered_map<int,int> {
    using std::unordered_map<int, int>::unordered_map;
    // ~Foo() = default; // adding this allows it to work
};

struct Bar {
    Bar(Foo f = {}) : _f(std::move(f)) {}
    // using any of the following constructors fixes the problem:
    // Bar(Foo f = Foo()) : _f(std::move(f)) {}
    // Bar(Foo f = {}) : _f(f) {}

    Foo _f;
};

int main() {
    Bar b;

    // the following code works as expected
    // Foo f1 = {};
    // Foo f2 = std::move(f1);
}

我的编辑设置:

g++ --std=c++11 main.cpp

以下是GDB的回溯:

#0  0x00007fff95d50866 in __pthread_kill ()
#1  0x00007fff90ba435c in pthread_kill ()
#2  0x00007fff8e7d1bba in abort ()
#3  0x00007fff9682e093 in free ()
#4  0x0000000100002108 in __gnu_cxx::new_allocator<std::__detail::_Hash_node_base*>::deallocate ()
#5  0x0000000100001e7d in std::allocator_traits<std::allocator<std::__detail::_Hash_node_base*> >::deallocate ()
#6  0x0000000100001adc in std::__detail::_Hashtable_alloc<std::allocator<std::__detail::_Hash_node<std::pair<int const, int>, false> > >::_M_deallocate_buckets ()
#7  0x000000010000182e in std::_Hashtable<int, std::pair<int const, int>, std::allocator<std::pair<int const, int> >, std::__detail::_Select1st, std::equal_to<int>, std::hash<int>, std::__detail::_Mod_range_hashing, std::__detail::_Default_ranged_hash, std::__detail::_Prime_rehash_policy, std::__detail::_Hashtable_traits<false, false, true> >::_M_deallocate_buckets ()
#8  0x000000010000155a in std::_Hashtable<int, std::pair<int const, int>, std::allocator<std::pair<int const, int> >, std::__detail::_Select1st, std::equal_to<int>, std::hash<int>, std::__detail::_Mod_range_hashing, std::__detail::_Default_ranged_hash, std::__detail::_Prime_rehash_policy, std::__detail::_Hashtable_traits<false, false, true> >::~_Hashtable ()
#9  0x000000010000135c in std::unordered_map<int, int, std::hash<int>, std::equal_to<int>, std::allocator<std::pair<int const, int> > >::~unordered_map ()
#10 0x00000001000013de in Foo::~Foo ()
#11 0x0000000100001482 in Bar::~Bar ()
#12 0x0000000100001294 in main ()

*** error for object 0x1003038a0: pointer being freed was not allocated ***

1 个答案:

答案 0 :(得分:11)

更新

它显示为fix for the problem has been checked in


有趣的问题。这似乎是GCC如何处理= {}初始化默认参数的错误,这是late addition to the standard。问题可以通过一个非常简单的类代替std::unordered_map<int,int>

来重现
#include <utility>

struct PtrClass
{
    int *p = nullptr;

    PtrClass()
    {
        p = new int;
    }

    PtrClass(PtrClass&& rhs) : p(rhs.p)
    {
        rhs.p = nullptr;
    }

    ~PtrClass()
    {
        delete p;
    }
};

void DefArgFunc(PtrClass x = {})
{
    PtrClass x2{std::move(x)};
}

int main()
{
    DefArgFunc();
    return 0;
}

Compiled with g++ (Ubuntu 4.8.1-2ubuntu1~12.04) 4.8.1,它显示了同样的问题:

*** glibc detected *** ./a.out: double free or corruption (fasttop): 0x0000000001aa9010 ***
======= Backtrace: =========
/lib/x86_64-linux-gnu/libc.so.6(+0x7eb96)[0x7fc2cd196b96]
./a.out[0x400721]
./a.out[0x4006ac]
/lib/x86_64-linux-gnu/libc.so.6(__libc_start_main+0xed)[0x7fc2cd13976d]
./a.out[0x400559]
======= Memory map: ========
bash: line 7:  2916 Aborted                 (core dumped) ./a.out

进一步深入,GCC似乎在使用这种语法时创建了一个额外的对象(尽管它只调用构造函数和析构函数):

#include <utility>
#include <iostream>

struct SimpleClass
{    
    SimpleClass()
    {
        std::cout << "In constructor: " << this << std::endl;
    }

    ~SimpleClass()
    {
        std::cout  << "In destructor: " << this << std::endl;
    }
};

void DefArgFunc(SimpleClass x = {})
{
        std::cout << "In DefArgFunc: " << &x << std::endl;
}

int main()
{
    DefArgFunc();
    return 0;
}

Output

In constructor: 0x7fffbf873ebf
In DefArgFunc: 0x7fffbf873ea0
In destructor: 0x7fffbf873ebf

将默认参数从SimpleClass x = {}更改为SimpleClass x = SimpleClass{}会产生

In constructor: 0x7fffdde483bf
In DefArgFunc: 0x7fffdde483bf
In destructor: 0x7fffdde483bf

正如所料。

似乎正在发生的是创建一个对象,调用默认构造函数,然后执行类似于memcpy的操作。这个“鬼对象”是传递给移动构造函数并被修改的内容。但是,析构函数在原始的未修改对象上调用,该对象现在与移动构造的对象共享一些指针。最终两者都试图释放它,导致问题。

根据上述说明,您注意到的四个更改修复了问题:

// 1
// adding the destructor inhibits the compiler generated move constructor for Foo,
// so the copy constructor is called instead and the moved-to object gets a new
// pointer that it doesn't share with the "ghost object", hence no double-free
~Foo() = default;

// 2
// No  `= {}` default argument, GCC bug isn't triggered, no "ghost object"
Bar(Foo f = Foo()) : _f(std::move(f)) {}

// 3
// The copy constructor is called instead of the move constructor
Bar(Foo f = {}) : _f(f) {}

// 4
// No  `= {}` default argument, GCC bug isn't triggered, no "ghost object"
Foo f1 = {};
Foo f2 = std::move(f1);

将参数传递给构造函数(Bar b(Foo{});)而不是使用默认参数也可以解决问题。