Initializing from an initializer list, but without

问题:

I recently stumbles across some problem with initializer lists. Consider a program that stores map-like data

struct MyMapLike {
  MyMapLike(std::map<std::string, int> data)
    :data(std::move(data))
  { }

private:
  std::map<std::string, int> data;
};

That looks straight forward. But when initializing it, it becomes ugly. I want to let it look like

MyMapLike maps = { { "One", 1 }, { "Two", 2 } };

But the compiler doesn't want to accept this, because the above means that it should look for a two-parameter constructor that can accept { "One", 1 } and { "Two", 2 } respectively. I need to add extra braces, to make it look like a single-parameter constructor accepting the { { ... }, { ... } }

MyMapLike maps = { { { "One", 1 }, { "Two", 2 } } };

I would not like to write it like that. Since I have a map-like class, and the initializer has the abstract value of a mapping-list, I would like to use the former version, and be independent of any such implementation details like level of nesting of constructors.

One work around is to declare an initializer-list constructor

struct MyMapLike {
  MyMapLike(std::initializer_list< 
    std::map<std::string, int>::value_type
    > vals)
    :data(vals.begin(), vals.end())
  { }

  MyMapLike(std::map<std::string, int> data)
    :data(std::move(data))
  { }

private:
  std::map<std::string, int> data;
};

Now I can use the former, because when I have an initializer-list constructor, the whole initializer list is treated as one element instead of being splitted into elements. But I think this separate need of the constructor is dead ugly.

I'm looking for guidance:

• What do you think about the former and latter form of initialization? Does it make sense to be required to have extra braces in this case?
• Do you consider the requirement for addition of an initializer list constructor in this case bad?

If you agree with me on that the former way of initialization is nicer, what solutions can you think of?

回答1:

Since I have a map-like class, and the initializer has the abstract value of a mapping-list, I would like to use the former version

And herin lies the problem: it's up to you to supply the constructors that allow your class to be treated like a map. You called your solution a work-around, but there's nothing to work around. :)

But I think this separate need of the constructor is dead ugly.

It is, but unfortunately since it's your class, you have to specify how the initializer lists work.

回答2:

What do you think about the former and latter form of initialization? Does it make sense to be required to have extra braces in this case?

I think so. I think it would permit too much ambiguity to allow a constructor be called not just when the syntax matches that constructor, but when the syntax matches some constructor of the single argument to the constructor, and so on recursively.

struct A { int i; };
struct B { B(A) {}  B(int) {} };
struct C { C(B) {} };

C c{1};

Do you consider the requirement for addition of an initializer list constructor in this case bad?

No. It lets you get the syntax you want but without creating the problems that arise if we make the compiler search harder for a constructor to use.

回答3:

Wouldn't something like this give the desired effect (MyMapLike can be constructed in any way that std::map can, but does not implicitly convert to std::map)?

struct MyMapLike : private std::map<std::string, int>
{
    using map::map;
};

If it absolutely positively has to be a member, maybe use constructor perfect forwarding (I'm not sure about the exact syntax) along the lines of:

struct MyMapLike
{
    template<typename... Initializers>
    MyMapLike(Initializers... init, decltype(new std::map<std::string, int>(...init)) = 0)
      : data(...init)
    { }

private:
    std::map<std::string, int> data;
};

回答4:

You could initialize as follows:

MyMapLike maps ( { { "One", 1 }, { "Two", 2 } } );

The outer ( ... ) are now clearly simply to surround the constructor args, and it's easier to see that the outermost { ... } defines the 'list'.

It's still a bit verbose, but it avoids the situation where { is being used to do two different things, which affects readability.

回答5:

I agree that it's ugly. I usually scrap initializer lists once I move beyond very simple cases as they're quite limited. For your case I'd go with boost::assign, which while not as terse gives better flexibility and control.

回答6:

What do you think about the former and latter form of initialization? Does it make sense to be required to have extra braces in this case?

I prefer the first form because it has a clean class interface. The constructor taking an initializer list pollutes the interface and offers little in return.

The extra braces are something we'll get used to. Just like we got used to C++'s other quirks.