Is delete this allowed?

Question

Is it allowed to delete this; if the delete-statement is the last statement that will be executed on that instance of the class? Of course I'm sure that the object represented by the this-pointer is newly-created.

I'm thinking about something like this:

void SomeModule::doStuff()
{
    // in the controller, "this" object of SomeModule is the "current module"
    // now, if I want to switch over to a new Module, eg:

    controller->setWorkingModule(new OtherModule());

    // since the new "OtherModule" object will take the lead, 
    // I want to get rid of this "SomeModule" object:

    delete this;
}

Can I do this?

Answer 1

Yes, delete this; has defined results, as long as (as you've noted) you assure the object was allocated dynamically, and (of course) never attempt to use the object after it's destroyed. Over the years, many questions have been asked about what the standard says specifically about delete this;, as opposed to deleting some other pointer. The answer to that is fairly short and simple: it doesn't say much of anything. It just says that delete's operand must be an expression that designates a pointer to an object, or an array of objects. It goes into quite a bit of detail about things like how it figures out what (if any) deallocation function to call to release the memory, but the entire section on delete (§[expr.delete]) doesn't mention delete this; specifically at all. The section on destrucors does mention delete this in one place (§[class.dtor]/13):

At the point of definition of a virtual destructor (including an implicit definition (15.8)), the non-array deallocation function is determined as if for the expression delete this appearing in a non-virtual destructor of the destructor’s class (see 8.3.5).

That tends to support the idea that the standard considers delete this; to be valid--if it was invalid, its type wouldn't be meaningful. That's the only place the standard mentions delete this; at all, as far as I know.

Anyway, some consider delete this a nasty hack, and tell anybody who will listen that it should be avoided. One commonly cited problem is the difficulty of ensuring that objects of the class are only ever allocated dynamically. Others consider it a perfectly reasonable idiom, and use it all the time. Personally, I'm somewhere in the middle: I rarely use it, but don't hesitate to do so when it seems to be the right tool for the job.

The primary time you use this technique is with an object that has a life that's almost entirely its own. One example James Kanze has cited was a billing/tracking system he worked on for a phone company. When start to you make a phone call, something takes note of that and creates a phone_call object. From that point onward, the phone_call object handles the details of the phone call (making a connection when you dial, adding an entry to the database to say when the call started, possibly connect more people if you do a conference call, etc.) When the last people on the call hang up, the phone_call object does its final book-keeping (e.g., adds an entry to the database to say when you hung up, so they can compute how long your call was) and then destroys itself. The lifetime of the phone_call object is based on when the first person starts the call and when the last people leave the call--from the viewpoint of the rest of the system, it's basically entirely arbitrary, so you can't tie it to any lexical scope in the code, or anything on that order.

For anybody who might care about how dependable this kind of coding can be: if you make a phone call to, from, or through almost any part of Europe, there's a pretty good chance that it's being handled (at least in part) by code that does exactly this.

Answer 2

This is an old, answered, question, but @Alexandre asked "Why would anyone want to do this?", and I thought that I might provide an example usage that I am considering this afternoon.

Legacy code. Uses naked pointers Obj*obj with a delete obj at the end.

Unfortunately I need sometimes, not often, to keep the object alive longer.

I am considering making it a reference counted smart pointer. But there would be lots of code to change, if I was to use ref_cnt_ptr<Obj> everywhere. And if you mix naked Obj* and ref_cnt_ptr, you can get the object implicitly deleted when the last ref_cnt_ptr goes away, even though there are Obj* still alive.

So I am thinking about creating an explicit_delete_ref_cnt_ptr. I.e. a reference counted pointer where the delete is only done in an explicit delete routine. Using it in the one place where the existing code knows the lifetime of the object, as well as in my new code that keeps the object alive longer.

Incrementing and decrementing the reference count as explicit_delete_ref_cnt_ptr get manipulated.

But NOT freeing when the reference count is seen to be zero in the explicit_delete_ref_cnt_ptr destructor.

Only freeing when the reference count is seen to be zero in an explicit delete-like operation. E.g. in something like:

template<typename T> class explicit_delete_ref_cnt_ptr { 
 private: 
   T* ptr;
   int rc;
   ...
 public: 
   void delete_if_rc0() {
      if( this->ptr ) {
        this->rc--;
        if( this->rc == 0 ) {
           delete this->ptr;
        }
        this->ptr = 0;
      }
    }
 };

OK, something like that. It's a bit unusual to have a reference counted pointer type not automatically delete the object pointed to in the rc'ed ptr destructor. But it seems like this might make mixing naked pointers and rc'ed pointers a bit safer.

But so far no need for delete this.

But then it occurred to me: if the object pointed to, the pointee, knows that it is being reference counted, e.g. if the count is inside the object (or in some other table), then the routine delete_if_rc0 could be a method of the pointee object, not the (smart) pointer.

class Pointee { 
 private: 
   int rc;
   ...
 public: 
   void delete_if_rc0() {
        this->rc--;
        if( this->rc == 0 ) {
           delete this;
        }
      }
    }
 };

Actually, it doesn't need to be a member method at all, but could be a free function:

map<void*,int> keepalive_map;
template<typename T>
void delete_if_rc0(T*ptr) {
        void* tptr = (void*)ptr;
        if( keepalive_map[tptr] == 1 ) {
           delete ptr;
        }
};

(BTW, I know the code is not quite right - it becomes less readable if I add all the details, so I am leaving it like this.)

Answer 3

This is the core idiom for reference-counted objects.

Reference-counting is a strong form of deterministic garbage collection- it ensures objects manage their OWN lifetime instead of relying on 'smart' pointers, etc. to do it for them. The underlying object is only ever accessed via "Reference" smart pointers, designed so that the pointers increment and decrement a member integer (the reference count) in the actual object.

When the last reference drops off the stack or is deleted, the reference count will go to zero. Your object's default behavior will then be a call to "delete this" to garbage collect- the libraries I write provide a protected virtual "CountIsZero" call in the base class so that you can override this behavior for things like caching.

The key to making this safe is not allowing users access to the CONSTRUCTOR of the object in question (make it protected), but instead making them call some static member- the FACTORY- like "static Reference CreateT(...)". That way you KNOW for sure that they're always built with ordinary "new" and that no raw pointer is ever available, so "delete this" won't ever blow up.

Answer 4

If it scares you, there's a perfectly legal hack:

void myclass::delete_me()
{
    std::unique_ptr<myclass> bye_bye(this);
}

I think delete this is idiomatic C++ though, and I only present this as a curiosity.

There is a case where this construct is actually useful - you can delete the object after throwing an exception that needs member data from the object. The object remains valid until after the throw takes place.

void myclass::throw_error()
{
    std::unique_ptr<myclass> bye_bye(this);
    throw std::runtime_exception(this->error_msg);
}

Note: if you're using a compiler older than C++11 you can use std::auto_ptr instead of std::unique_ptr, it will do the same thing.