Member functions called in lambdas

Question

This question arose when programming for Unreal Engine with TFunction<> but it is fundamentally about C++ and I believe the example is equivalent.

The questions are:

How to ensure that the callback is not called when the object doesn't exist
Is there a significant overhead from this type of member-function-from-lambda call?
Are there some other hidden dangers in this approach?

The code:

#include <iostream>
#include <functional>

// Basic class with a callback function expressed as a lambda calling a member function 
class Foo {
    private:
        int i = 0;
    public:
        Foo(int j){
            i = j;
        }

        // A member function that should be called by a manager
        void f(){
            std::cout << i << std::endl;
        }

        // Wrap the member function in a lambda
        std::function<void()> wrapper = [&](){f();}; // <---- What happens when object is destroyed before call? 
};

// Manager class which calls a specified function after some work 
class Manager {
    public:
        // A function that returns immediately and after some time calls the passed function
        void asyncWork(std::function<void()> callback){
            // Some work here...

            callback();
        } 
};

int main() {

    // Create object
    Foo foo(42);

    // Create manager
    Manager manager;

    // Pass the callback function to the manager
    manager.asyncWork(foo.wrapper);

    return 0; 
}

The program does the following:

Create a simple object which has a std::function wrapping for a member function call. This is done in Unreal Engine using the TFunction<> type.
A manager class is passed this wrapped function as a callback
After some time, the callback is triggered but the original object may not exist anymore

What behavior are you expecting? Some way to ensure that the object will always exist? If the object is destroyed, do you want to produce some error? to throw an exception? to terminate the program?... BTW, If I get it right, when a Foo object is destroyed, the closure is also but asyncWork takes it's argument by value, thus a copy of the closure is made. If it does not capture reference that could become dangling. It seems OK. — Oersted
– Oersted, Commented Mar 10 at 12:12
I want there to be no effect or errors if the object is destroyed prior to the manager trying to call the function. — dubious
– dubious, Commented Mar 10 at 12:15
I've misread the snippet: There is a dangling reference to the function pointer. Is passing a std::shared_ptr<Foo> instead of a std::function an option? — Oersted
– Oersted, Commented Mar 10 at 12:25
Note : make sure there is no async call still in progress when you destroy your Foo. A possible solution is to capture a std::shared_ptr<Foo> to your thread or let your asyncWorker return a std::future that Foo can use to synchronize with in its destructor. In any case you need to extend the lifecycle of Foo for as long as it has outstanding asynchronous calls. — Pepijn Kramer
– Pepijn Kramer, Commented Mar 10 at 12:46
Either unregister the call when object is destroyed, give the lambda the shared ownership of the object (so object is not destroyed), or provide a way to the lambda to know if object is still alive (std::weak_ptr might help in that regard). — Jarod42
– Jarod42, Commented Mar 10 at 13:35

Xist · Accepted Answer · 2025-11-03 13:27:39Z

4

To answer specific to Unreal Engine, the solution is to use a TWeakObjectPtr. This requires the class be derived from UObject.

Assuming that you change Foo to UFoo : public UObject then:

TFunction<void()> UFoo::Wrapper()
{
    TWeakObjectPtr WeakThis (this);
    return [WeakThis]()
    {
        if (ThisClass* This = WeakThis.Get())
        {
            // This is a valid object
            This->CallPrivateMethod();  // or call This->f() if you want
        }
        // else the object WeakThis points to has been destroyed
    };
}

If you are concerned about this being garbage collected while your logic is running, you can convert it to a TStrongObjectPtr before using it, like:

TFunction<void()> UFoo::Wrapper()
{
    TWeakObjectPtr WeakThis (this);
    return [WeakThis]()
    {
        if (TStrongObjectPtr<ThisClass> StrongThis = WeakThis.Pin();
            ThisClass* This = StrongThis.Get())
        {
            // This is a valid object
            This->CallPrivateMethod();  // or call This->f() if you want
        }
        // else the object WeakThis points to has been destroyed
    };
}

To generalize to C++, presumably the same method would work with a std::weak_ptr.

Thanks to Yakk for the .Pin comment, I updated this post with that info so people won't miss it.

edited Nov 3 at 13:27

answered Mar 10 at 13:51

Xist

1163 bronze badges

Sign up to request clarification or add additional context in comments.

3 Comments

dubious Mar 10 at 13:57

The weak pointer in UE has a method IsValid() which would perhaps make the code shorter

Xist Mar 10 at 14:02

You could if (WeakThis.IsValid()) { ThisClass* This = WeakThis.Get(); /*...*/ } but that effectively requires calling IsValid twice since it's baked into Get(). Calling Get() and checking for nullptr has the same effect.

Yakk - Adam Nevraumont Mar 10 at 14:10

@dubious Never use IsValid() for that purpose; you can use it to filter out invalid pointers from a container, but if you are going to use a weak pointer ideally you want to .lock() it - for UE use .Pin() so the weak pointer isn't gc'd under your feet in the function.

Jarod42 · Accepted Answer · 2025-03-11 15:48:02Z

The first rule of sane lambdas is to use [&] when your lambda lifetime (and all of its copies) will NOT leave the current scope.

Ie:

ForEachChicken( [&](auto& chicken){ chicken.LayEgg(); } );

you are using a lambda as a "block of code you can call", and injecting it into some code that wraps it.

The second rule of sane lambdas is to never use [&] if your lambda (or any copy of it) could possibly out-live the current scope.

std::function<void()> wrapper = [&](){f();}; // <---- What happens when object is destroyed before call?

this is an example of a lambda that could out-live the current scope. The use of [&] should get an instant -2 code review, regardless of if the resulting code is correct or not, because [&] puts way way way too much work on doing everything exactly correct.

So the first issue is to explicitly capture everything you need into the lambda.

std::function<void()> wrapper = [this](){f();};

the point of this is that it makes your problem clear -- this wrapper stores a copy of a this pointer. Its correct use then requires that it be the right pointer, be in lifetime, etc.

A bug you now have is that wrapper will be copied on operator= and on copy/move construction, but the this captured will be from the other instance. By storing a pointer-to-self within an object, you basically have to block copy-move assignment-construction, or rewrite them.

The second problem is the lifetime problem you noted.

Fixing the first problem, I'd make wrapper be a method itself:

std::function<void()> wrapper() const{ return [this](){f();}; }

so now it no longer takes up state, and messes with our copy/move semantics.

For the second problem (which you came here to ask about), well this is C++. You are responsible to handle object lifetimes.

I have found 2 common patterns to deal with this. The first is if the object in question is either single-threaded, or will only be invoked and destroyed from the same single thread. The second is if the object in question has a complex lifetime I want to be managed by shared pointers.

If the object in question is single-threaded (with regards to this invocation and its eventual destruction), I add a lifetime tracker to it. A lifetime tracker looks like this:

struct IsAlive {
  std::weak_ptr<void> tracking;
  explicit operator bool() const { return !tracking.expired(); }
};

struct LifetimeTracked {
  std::shared_ptr<void> track = std::make_shared<bool>(true);
  LifetimeTracked() = default;
  LifetimeTracked(LifetimeTracked const&):LifetimeTracked() {} // not default
  LifetimeTracked(LifetimeTracked&&):LifetimeTracked() {} // not default

  LifetimeTracked& operator=(LifetimeTracked const&)&{
    return *this; // not default
  }
  LifetimeTracked& operator=(LifetimeTracked &&) {
    return *this; // not default
  }
  
  IsAlive TrackLifetime() const { return {track}; }
};

you stick a LifetimeTracked into a struct or inherit from it. Then you call TrackLifetime and store the IsAlive. You can then query IsAlive to see if the object has been destroyed.

Your wrapper() then becomes:

std::function<void()> wrapper() const{ return [this, alive=TrackLifetime()](){if(alive)f();}; }

So long as deleting the object never happens in a different thread than the return of wrapper() is called in, this will work. It doesn't require the object live in the heap, it doesn't mess with the objects lifetime so you can use it as a RAII class, etc.

I've even used this outside of lambdas in menu handling code. When the menu handler could clean up the stack of submenus or not, after the handling code you don't know if this is still around (the menu that invoked the menu item). So you grab a lifetime tracker before invoking the handler, and afterwards if it says you are dead you just exit without doing cleanup (as this is now a dangling pointer).

Now, the other option is if your class is already being managed via something like a shared pointer, or if you don't mind the heap-allocation requirements and complex lifetime of making it so.

Then you store a weak_ptr or framework alternative in your callback:

std::function<void()> wrapper() const{ return
  [wp = std::weak_ptr(shared_from_this())](){
    if(auto self = wp.lock())
      self->f();
  };
}

The downsides are that the object needs to be managed by a shared pointer, and doing so makes code compile and dangling pointers less likely, but leaks infinitely more often and in ways you won't easily be able to understand. It also makes object lifetime very difficult to understand, blocks use in automatic storage (like in a std::vector of packed objects, very efficient).

The third option is ... manage your objects lifetime. Actually guarantee that the object lives longer than any callbacks that contain pointers into it by having a full understanding of the lifetime of every object in every situation. This is insanely hard in a modern application, but I leave it here for completeness.

A forth option is to not store pointers, but only store ways to find an object. In this case, objects will have identities and probably breadcrumbs describing what they are and how to find them.

std::function<void()> wrapper() const{ return
  [identity = GetIdentity()](){
    if(auto self = identity.Resolve(CurrentScope()))
      self->f();
  };
}

Now instead of storing a pointer to the object, you somehow store a description of what this object is. When invoked, you use that description to find an object that is appropriate, and then interact with it if it exists.

Imagine if your objects are mobs in a simple game. Each mob might have a guid attached to it. Your identity could then be a string saying "the mob named bob" (where bob is a guid). At invocation time you look up in the global mob map "is there a mob named bob?", and if so you provide it to be interacted with.

Very good points, thanks! The issue is that I have no control over how the object lifetime is managed and it is likely to be culled for various reasons. The game engine does all that and I do not wish to edit that part of the code. However, there is now more than enough information for a good solution.
@Yakk: Just fixed some "typos" with LifeTimeTracked as tracking was wrong buggy version. Fixed version. unsure about reference you want for operator= though.
@Jarod42 And = should be a noop; the lifetime of an object is not modifed by being assigned-to or assigned-from.

Collectives™ on Stack Overflow

Member functions called in lambdas

2 Answers 2

3 Comments

4 Comments

Your Answer

Hot Network Questions

Collectives™ on Stack Overflow

2 Answers 2

3 Comments

4 Comments

Your Answer

Sign up or log in

Post as a guest

Related