## capturing perfectly-forwarded objects in lambdas

11 december 2016

Perfect forwarding and forwarding references allow developers to write generic template functions that retain the lvalueness/rvalueness of passed arguments, in order to avoid unnecessary copies or support reference semantics without having to implement multiple overloads. (This article by Eli Bendersky explains them in depth. I will assume you're familiar with these concepts for the rest of the article.)

In this article, I'll show how to correctly capture perfectly-forwarded objects into lambdas:

• We'll take a look at an example that shows how using std::forward in lambda captures can produce unexpected results.

• A perfect forward capture wrapper that prevents the aforementioned unexpected results will then be implemented.

• Macros (yuck!) will be used to reduce noise/boilerplate.

• The solution will be finally generalized to pefectly-forwarded variadic argument packs.

### The problem

Let's begin with a simple quiz. Given this definition of A...

struct A
{
int _value{0};
};

...and the following code...

auto foo = [](auto& a)
{
return [&a]{ ++a._value; };
};

A my_a;
foo(my_a)();
// What is my_a._value here?

...what is the value of my_a._value?

It should be obvious that, since a was passed by lvalue reference to foo and captured by reference from the inner lambda, the increment will be visible outside of the function call. Therefore, my_a._value is equal to 1.

(Complete example on wandbox.)

Let's now generalize foo - it will take a by forwarding reference, and should have the following semantics:

• When a is an lvalue reference, it should be captured by reference in the inner lambda, and any mutation on a should be visible outside of the function call (i.e. what happened in the example above).

• When a is an rvalue reference, it should be moved in the inner lambda. The returned lambda will therefore hold a new instance of A whose mutations will only be visible from the lambda itself.

Sounds like a job for perfect forwarding!

Let's try to use auto&&, std::forward and C++14's generalized lambda capture syntax to implement our desired semantics. The inner lambda will be marked as mutable since it will store its own instance of A when a is an rvalue reference.

auto foo = [](auto&& a)
{
return [a = std::forward<decltype(a)>(a)]() mutable
{
++a._value;

// Print to see a being mutated when moved into
// the inner lambda.
std::cout << a._value << "\n";
};
};

That was easy, wasn't it? Time to test our code.

• When a is an rvalue reference...

auto l_inner = foo(A{});
l_inner(); // Prints 1.
l_inner(); // Prints 2.
l_inner(); // Prints 3.

...the implementation works as intended.
(Complete example on wandbox.)

• When a is an lvalue reference...

A my_a;
auto l_inner = foo(my_a);
l_inner(); // Prints 1.
l_inner(); // Prints 2.
l_inner(); // Prints 3.

// Prints 0... wtf?
std::cout << my_a._value << "\n";

...reference semantics do not work as expected anymore!
(Complete example on wandbox.)

While it seems natural to use std::forward and generalized lambda captures in order to achieve the same semantics as a template function that takes forwarding references, if we stop and think for a moment our mistake will become obvious.

Let's look at foo again:

auto foo = [](auto&& a)
{
return [a = std::forward<decltype(a)>(a)]() mutable { /* ... */ };
};

What does a = std::forward<decltype(a)>(a) really mean? It becomes clear if we "play compiler" and write a struct that is equivalent to the inner lambda:

struct inner_lambda
{
A _captured_a;

template <typename T>
inner_lambda(T&& a) : _captured_a(std::forward<decltype(a)>(a))
{
}

void operator()() { /* ... */ }
};

As you can see, the type of the captured a is always a value! This is due to the fact that the initializer expression in a generalized lambda capture does not have effect on the deduced type of the capture. From cppreference/lambda/lambda_capture:

A capture with an initializer acts as if it declares and explicitly captures a variable declared with type auto, whose declarative region is the body of the lambda expression (that is, it is not in scope within its initializer), except that:

• if the capture is by-copy, the non-static data member of the closure object is another way to refer to that auto variable.

• if the capture is by-reference, the reference variable's lifetime ends when the lifetime of the closure object ends

This is used to capture move-only types with a capture such as x = std::move(x)

Even with generalized lambda captures the programmer has to be explicit when a reference is desired:

int a = 0;
int& a_ref = a;

// Captures a by copy.
[b = a_ref]{ };

// Captures a by reference.
[&b = a_ref]{ }; 

It should be clear now that our previous approach is fundamentally flawed. What we really want to do is:

• Capture by copy if a is an rvalue reference.

• Capture by reference if a is an lvalue reference.

That's impossible without adding an extra layer of indirection, as the language requires us to be explicit about our capture semantics (and does not provide any special syntax for "capture by perfect forwarding").

### The solution

One possible way of solving the aforementioned issue and achieve the desired capturing semantics is to create a wrapper class that will always be captured by copy by the inner lambda, but will internally store either an lvalue-reference or a value depending on how it is constructed.

The solution I'm going to present here was recently designed and implemented because I encountered this particular situation when experimenting with stack-allocated asynchronous chain generation. Most of the ideas behind it come from this StackOverflow question I asked back in 2014: "Capturing perfectly-forwarded variable in lambda".

The wrapper will be a template class called fwd_capture_wrapper. It will be specialized depending on the lvalue-ness/rvalue-ness of the captured object:

// Unspecialized version: stores a T instance by value.
template <typename T>
struct fwd_capture_wrapper : impl::by_value<T>
{
// "Inherit" constructors.
using impl::by_value<T>::by_value;
};

// Specialized version: stores a T reference.
template <typename T>
struct fwd_capture_wrapper<T&> : impl::by_ref<T>
{
// "Inherit" constructors.
using impl::by_ref<T>::by_ref;
};

Let's now implement impl::by_value<T>. It will store a plain T instance (initialized via a perfect forwarding constructor)* and provide .get() methods to access it.

template <typename T>
class by_value
{
private:
T _x;

public:
template <typename TFwd>
by_value(TFwd&& x) : _x{std::forward<TFwd>(x)}
{
}

auto&       get() &      { return _x; }
const auto& get() const& { return _x; }
auto        get() &&     { return std::move(_x); }
};

When designing impl::by_ref<T>, we could choose to store a T&. Unfortunately this would prevent us from easily adding copy/move operations for the wrapper itself. Thankfully std::reference_wrapper was introduced in C++11. It is TriviallyCopyable and "behaves like a pointer", but does not allow any null state and prevents accidental construction from temporary objects.

template <typename T>
class by_ref
{
private:
std::reference_wrapper<T> _x;

public:
by_ref(T& x) : _x{x}
{
}

auto&       get() &      { return _x.get(); }
const auto& get() const& { return _x.get(); }
auto        get() &&     { return std::move(_x.get()); }
};

That's it! The only missing piece is a nice fwd_capture interface function that hides the boilerplate away from the user:

template <typename T>
auto fwd_capture(T&& x)
{
return fwd_capture_wrapper<T>(std::forward<T>(x));
}

Let's revisit our flawed implementation of foo:

auto foo = [](auto&& a)
{
return [a = fwd_capture(std::forward<decltype(a)>(a))]() mutable
{
++a.get()._value;
std::cout << a.get()._value << "\n";
};
};

a is now a fwd_capture_wrapper that will store either a value or a reference as intended, depending on what a is in the outer lambda.

Our test cases now work correctly!

auto l_inner = foo(A{});
l_inner(); // Prints 1.
l_inner(); // Prints 2.
l_inner(); // Prints 3.
A my_a;
auto l_inner = foo(my_a);
l_inner(); // Prints 1.
l_inner(); // Prints 2.
l_inner(); // Prints 3.

// Prints 3, yay!
std::cout << my_a._value << "\n";

(Complete example on wandbox.)

### Reducing noise

Typing fwd_capture(std::forward<decltype(a)>(a)) is very annoying, as it's mostly noise/boilerplate. We have to perfectly-forward a into the wrapper though, so avoiding the call to std::forward is out of the question.

There's only one evil tool that can help here: macros.

Let's begin with std::forward<decltype(a)>(a): a needs to be repeated twice. We can define a FWD macro that does that for us:

#define FWD(...) ::std::forward<decltype(__VA_ARGS__)>(__VA_ARGS__)

(See "vrm_core issue #1" for discussion regarding __VA_ARGS__ in this context.)

The FWD macro can be used as follows:

template <typename T>
void foo(T&&);

template <typename T>
void bar(T&& x);
{
foo(FWD(x));
// ...equivalent to...
foo(std::forward<decltype(x)>(x));
}

Great! Now fwd_capture(std::forward<decltype(a)>(a)) can be reduced to fwd_capture(FWD(a)):

return [a = fwd_capture(FWD(a))]() mutable
{
/* ... */
};

There's still some unnecessary repetition: fwd_capture always expects the user to perfectly-forward something - there's no reasonable usage of fwd_capture without FWD. Therefore... we can hide fwd_capture in an implementation namespace and provide a macro that forwards the argument for us.

#define FWD_CAPTURE(...) impl::fwd_capture(FWD(__VA_ARGS__))

All the noise is finally hidden away:

return [a = FWD_CAPTURE(a)]() mutable
{
/* ... */
};

Remember that these macros should have some reasonably unique prefixes to avoid collisions, and that they're not mandatory - they, however, immensely increase code readability when perfectly-capturing a lot of arguments in the same lambda.

(Complete example on wandbox.)

True - let's assume that foo takes a variadic amount of arguments...

auto foo = [](auto&&... xs) { /* ... */ };

...and that we want to perfectly-capture all of them in the inner lambda. Unfortunately the following snippet doesn't work:

auto foo = [](auto&&... xs)
{
return [xs = FWD_CAPTURE(xs)...]() mutable
{

};
};

There is no way of doing a generalized variadic lambda capture. The solution (workaround?) is using std::tuple.

template <typename... Ts>
auto /*impl::*/fwd_capture_as_tuple(Ts&&... xs)
{
return std::make_tuple(FWD_CAPTURE(xs)...);
}

#define FWD_CAPTURE_PACK(...) impl::fwd_capture_as_tuple(FWD(__VA_ARGS__)...)

(std::make_tuple is going to copy/move the wrappers, not the object themselves.)

After capturing a variadic pack, we may want to "perfectly-apply" it to some function or to iterate over the captured objects. In order to do that, we can use std::experimental::apply and fold expressions:

// Expand all elements of fc into a f(...) invocation.
template <typename TF, typename TFwdCapture>
decltype(auto) apply_fwd_capture_pack(TF&& f, TFwdCapture&& fc)
{
return std::experimental::apply(
[&f](auto&&... xs) -> decltype(auto)
{
return f(FWD(xs).get()...);
}, FWD(fc));
}

// Execute f on every element of fc.
template <typename TF, typename TFwdCapture>
void for_fwd_capture_tuple(TF&& f, TFwdCapture&& fc)
{
apply_fwd_capture_pack(
[&f](auto&&... xs){ (f(FWD(xs)), ...); },
FWD(fc)
);
}

Using FWD_CAPTURE_PACK and apply_fwd_capture_pack we can now generalize our foo example even further:

auto foo = [](auto&&... xs)
{
// Perfectly-capture xs... as xs_pack.
return [xs_pack = FWD_CAPTURE_PACK(xs)]() mutable
{
// Perfectly-iterate over elements of xs_pack.
for_fwd_capture_tuple([](auto&& a)
{
++a._value;
std::cout << a._value << "\n";
}, xs_pack);
};
};

As an example, the following code snippet...

A my_a;
auto l_inner = foo(my_a, A{});
l_inner();
l_inner();

std::cout << my_a._value << "\n";

...will print:

1 # 1st invocation - my_a
1 # 1st invocation - A{}
2 # 2nd invocation - my_a
2 # 2nd invocation - A{}
2 # my_a._value

(Complete example on wandbox.)

### A simpler solution?

(This section was added on 12/12/2016.)

As shown by Nikos Athanasiou on reddit, it's possible to solve this issue in a much simpler way. All we need is std::tuple and some knowledge of template argument deduction rules.

Let's think about what fwd_capture_wrapper is doing:

• Given an lvalue reference, it stores an lvalue reference.

• Given an rvalue reference, it stores a value.

Again... this is something I asked on StackOverflow a while ago. (I find myself coming back to my own questions very often.)

Nikos suggested that it's sufficient to use std::tuple as our wrapper, using some sort of "type trait" that keeps lvalue references:

template <typename T>
using capture_t = std::conditional_t
<
std::is_lvalue_reference<T>{},

std::remove_reference_t<T>
>;

template <typename... Ts>
auto fwd_capture(Ts&&... xs)
{
return std::tuple<capture_t<Ts>...>(FWD(xs)...);
}

This works, and doesn't require all the wrapper boilerplate. std::get needs to be used in order to access the captured object, but it can always be wrapped in a nicer interface:

template <typename T>
decltype(auto) access(T&& x)
{
return std::get<0>(FWD(x));
}

(Complete example on wandbox.)

Guess what - it can be simplified even further. Turns out that capture_t is actually redundant! This definition of fwd_capture is therefore equivalent to the previous one:

template <typename... Ts>
auto fwd_capture(Ts&&... xs)
{
return std::tuple<Ts...>(FWD(xs)...);
}

Why?

Given this example function...

template <typename T>
void foo(T&&)
{
// What is T here?
std::tuple<T>{};
}

...let's look at the latest standard draft:

("P" below means "parameter".)

(from $14.8.2.1.3) If P is a cv-qualified type, the top-level cv-qualifiers of P's type are ignored for type deduction. If P is a reference type, the type referred to by P is used for type deduction. Example: template<class T> int f(const T&); int n1 = f(5); // calls f<int>(const int&) const int i = 0; int n2 = f(i); // calls f<int>(const int&) template <class T> int g(volatile T&); int n3 = g(i); // calls g<const int>(const volatile int&) The quote above means that T is deduced as T when an rvalue reference is passed to foo. (from$14.8.2.1.3)

A forwarding reference is an rvalue reference to a cv-unqualified template parameter that does not represent a template parameter of a class template (during class template argument deduction ([over.match.class.deduct])). If P is a forwarding reference and the argument is an lvalue, the type “lvalue reference to A” is used in place of A for type deduction. Example:

template <class T> int f(T&& heisenreference);
template <class T> int g(const T&&);
int i;

int n1 = f(i);     // calls f<int&>(int&)

int n2 = f(0);     // calls f<int>(int&&)

int n3 = g(i);     // error: would call g<int>(const int&&), which
// would bind an rvalue reference to an lvalue

The quote above means that T is deduced as T& when an lvalue reference is passed to foo.

Great! This means that...

template <typename... Ts>
auto fwd_capture(Ts&&... xs)
{
return std::tuple<Ts...>(FWD(xs)...);
}

...is all we need!

(Complete example on wandbox.)

#define FWD_CAPTURE(...) impl::fwd_capture(FWD(__VA_ARGS__))
}