This post starts with a fairly obscure topic - how an overloaded operator delete behaves in light of polymorphism; amazingly, it then gets even more obscure - shedding light on the trickery the compiler employs to make this work, by generating more than one destructor for certain classes. If you're into such things, read on. If not, sorry about that; I heard that three new Javascript libraries were released this week for MVC JSON-based dynamic CSS layout. Everyone's switching! Hurry up to keep up with the cool guys and leave this grumpy compiler engineer to mumble to himself.

Virtual operator delete?

Consider this code sample:

#include <cstdio>

class Animal {
public:
  virtual void say() = 0;
  virtual ~Animal() {}
};

class Sheep : public Animal {
public:
  virtual void say() {
    printf("Sheep says baaaaa\n");
  }

  virtual ~Sheep() {
    printf("Sheep is dead\n");
  }

  void operator delete(void* p) {
    printf("Reclaiming Sheep storage from %p\n", p);
    ::operator delete(p);
  }
};

int main(int argc, char** argv) {
  Animal* ap = new Sheep;
  ap->say();
  delete ap;
  return 0;
}

What happens when ap is deleted? Two things:

  1. The destructor of the object pointed to by ap is called.
  2. operator delete is called on ap to reclaim heap storage.

Part 1 is fairly clear: the static type of ap is Animal, but the compiler knows that Animal has a virtual destructor. So it looks up the actual destructor to invoke in the virtual table stored in the object ap points to. Since the dynamic type of ap is Sheep, the destructor found there will be Sheep::~Sheep, which is correct.

What about that operator delete, though? Is operator delete virtual too? Is is also stored in the virtual table? Because if it isn't, how does the compiler know which operator delete to invoke?

No, operator delete is not virtual. It is not stored in the virtual table. In fact, operator delete is a static member. The C++11 standard says so explicitly in secton 12.5:

Any deallocation function for a class X is a static member (even if not explicitly declared static).

It also adds:

Since member allocation and deallocation functions are static they cannot be virtual.

And if you keep reading, it actually mandates that even though this is the case, when the base destructor is virtual operator delete will be correctly looked up in the scope of the class that is the dynamic, not the static type of the object.

Indeed, the code snippet above works correctly and prints:

Sheep says baaaaa
Sheep is dead
Reclaiming Sheep storage from 0x1ed1be0

Deleting destructor

So how does this work, if operator delete is not virtual? Then answer is in a special destructor created for by the compiler. It's called the deleting destructor and its existence is described by the Itanium C++ ABI:

deleting destructor of a class T - A function that, in addition to the actions required of a complete object destructor, calls the appropriate deallocation function (i.e,. operator delete) for T.

The ABI goes on to provide more details:

The entries for virtual destructors are actually pairs of entries. The first destructor, called the complete object destructor, performs the destruction without calling delete() on the object. The second destructor, called the deleting destructor, calls delete() after destroying the object.

So now the mechanics of this operation should be fairly clear. The compiler mimics "virtuality" of operator delete by invoking it from the destructor. Since the destructor is virtual, what ends up called eventually is the destructor for the dynamic type of the object. In our example this would be the destructor of Sheep, which can call the right operator delete since it's in the same static scope.

However, as the ABI says, such classes need two destructors. If an object is destructed but not deleted from the heap, calling operator delete is wrong. So a separate version of the destructor exists for non-delete destructions.

Examining how the compiler implements deleting destructors

That's quite a bit of theory. Let's see how this is done in practice by studying the machine code generated by gcc for our code sample. First, I'll slightly modify main to invoke another function that just creates and discards a new Sheep without involving the heap.

void foo() {
  Sheep s;
}

int main(int argc, char** argv) {
  Animal* ap = new Sheep;
  ap->say();
  delete ap;

  foo();
  return 0;
}

And compiling this with the flags [1]:

g++ -O2 -g -static -std=c++11 -fno-inline -fno-exceptions

We get the following disassembly for main. I've annotated the disassembly with comments to explain what's going on:

0000000000400cf0 <main>:
  400cf0:    push   %rbx
  400cf1:    mov    $0x8,%edi

  // Call operator new to allocate a new object of type Sheep, and call
  // the constructor of Sheep. Neither Sheep nor Animal have fields, so
  // their size is 8 bytes for the virtual table pointer.
  // The pointer to the object will live in %rbx. The vtable pointer in this
  // object (set up by the constructor of Sheep) points to the the virtual
  // table of Sheep, because this is the actual type of the object (even
  // though we hold it by a pointer to Animal here).
  400cf6:    callq  401750 <_Znwm>
  400cfb:    mov    %rax,%rbx
  400cfe:    mov    %rax,%rdi
  400d01:    callq  4011f0 <_ZN5SheepC1Ev>

  // The first 8 bytes of an Animal object is the vtable pointer. So move
  // the address of vtable into %rax, and the object pointer itself ("this")
  // into %rdi.
  // Since the vtable's first entry is the say() method, the call that
  // actually happens here is Sheep::say(ap) where ap is the object pointer
  // passed into the (implicit) "this" parameter.
  400d06:    mov    (%rbx),%rax
  400d09:    mov    %rbx,%rdi
  400d0c:    callq  *(%rax)

  // Once again, move the vtable address into %rax and the object pointer
  // into %rdi. This time, invoke the function that lives at offset 0x10 in
  // the vtable. This is the deleting destructor, as we'll soon see.
  400d0e:    mov    (%rbx),%rax
  400d11:    mov    %rbx,%rdi
  400d14:    callq  *0x10(%rax)

  // Finally call foo() and return.
  400d17:    callq  4010d0 <_Z3foov>
  400d1c:    xor    %eax,%eax
  400d1e:    pop    %rbx
  400d1f:    retq

A diagram of the memory layout of the virtual table for Sheep can be helpful here. Since neither Animal nor Sheep have any fields, the only "contents" of a Sheep object is the vtable pointer which occupies the first 8 bytes:

                          Virtual table for Sheep:
ap:
--------------            -----------------------
| vtable ptr | ---------> |     Sheep::say()    |  0x00
--------------            -----------------------
                          |   Sheep::~Sheep()   |  0x08
                          -----------------------
                          | Sheep deleting dtor |  0x10
                          -----------------------

The two destructors seen here have the roles described earlier. Let's see their annotated disassembly:

// Sheep::~Sheep
0000000000401140 <_ZN5SheepD1Ev>:
  // Call printf("Sheep is dead\n")
  401140:    push   %rbx
  401141:    mov    $0x49dc7c,%esi
  401146:    mov    %rdi,%rbx
  401149:    movq   $0x49dd50,(%rdi)
  401150:    xor    %eax,%eax
  401152:    mov    $0x1,%edi
  401157:    callq  446260 <___printf_chk>
  40115c:    mov    %rbx,%rdi
  40115f:    pop    %rbx
  // Call Animal::~Animal, destroying the base class. Note the cool tail
  // call here (using jmpq instead of a call instruction - control does not
  // return here but the return instruction from _ZN6AnimalD1Ev will return
  // straight to the caller).
  401160:    jmpq   4010f0 <_ZN6AnimalD1Ev>
  401165:    nopw   %cs:0x0(%rax,%rax,1)
  40116f:    nop

// Sheep deleting destructor. The D0 part of the mangled name for deleting
// destructors, as opposed to D1 for the regular destructor, is mandated by
// the ABI name mangling rules.
00000000004011c0 <_ZN5SheepD0Ev>:
  4011c0:    push   %rbx
  // Call Sheep::~Sheep
  4011c1:    mov    %rdi,%rbx
  4011c4:    callq  401140 <_ZN5SheepD1Ev>
  4011c9:    mov    %rbx,%rdi
  4011cc:    pop    %rbx
  // Call Sheep::operator delete
  4011cd:    jmpq   401190 <_ZN5SheepdlEPv>
  4011d2:    nopw   %cs:0x0(%rax,%rax,1)
  4011dc:    nopl   0x0(%rax)

Now, going back to the amended code sample, let's see what code is generated for foo:

00000000004010d0 <_Z3foov>:
  4010d0:    sub    $0x18,%rsp
  4010d4:    mov    %rsp,%rdi
  4010d7:    movq   $0x49dd30,(%rsp)
  4010df:    callq  401140 <_ZN5SheepD1Ev>
  4010e4:    add    $0x18,%rsp
  4010e8:    retq
  4010e9:    nopl   0x0(%rax)

foo just calls Sheep::~Sheep. It shouldn't call the deleting destructor, because it does not actually delete an object from the heap.

It is also interesting to examine how the destructor(s) of Animal look, since unlike Sheep, Animal does not define a custom operator delete:

// Animal::~Animal
00000000004010f0 <_ZN6AnimalD1Ev>:
  4010f0:    movq   $0x49dcf0,(%rdi)
  4010f7:    retq
  4010f8:    nopl   0x0(%rax,%rax,1)

// Animal deleting destructor
0000000000401100 <_ZN6AnimalD0Ev>:
  401100:    push   %rbx
  // Call Animal::~Animal
  401101:    mov    %rdi,%rbx
  401104:    callq  4010f0 <_ZN6AnimalD1Ev>
  401109:    mov    %rbx,%rdi
  40110c:    pop    %rbx
  // Call global ::operator::delete
  40110d:    jmpq   4011f0 <_ZdlPv>
  401112:    nopw   %cs:0x0(%rax,%rax,1)
  40111c:    nopl   0x0(%rax)

As expected, the destructor of Animal calls the global ::operator delete.

Classes with virtual destructors vs. regular destructors

I want to emphasize that this special treatment - generation of a deleting destructor, is done not for classes that have a custom operator delete, but for all classes with virtual destructors. This is because when we delete an object through a pointer to the base class, the compiler has no way of knowing what operator delete to invoke, so this has to be done for every class where the destructor is virtual [2]. Here's a clarifying example:

#include <cstdio>

class Regular {
public:
  ~Regular() {
    printf("Regular dtor\n");
  }
};

class Virtual {
public:
  virtual ~Virtual() {
    printf("Virtual dtor\n");
  }
};

int main(int argc, char **argv) {
  Regular* hr = new Regular;
  delete hr;

  Virtual* hv = new Virtual;
  delete hv;

  return 0;
}

The only difference between Regular and Virtual here is the destructor being virtual in the latter. Let's examine the machine code for main to see how the two delete statements are lowered:

0000000000400cf0 <main>:
  400cf0:       push   %rbx
  400cf1:       mov    $0x1,%edi
  // Allocate a new Regular object with the global ::operator new
  400cf6:       callq  4016a0 <_Znwm>

  // If hr != nullptr, call Regular::~Regular, and then call the global
  // ::operator delete on hr.
  400cfb:       test   %rax,%rax
  400cfe:       mov    %rax,%rbx
  400d01:       je     400d13 <main+0x23>
  400d03:       mov    %rax,%rdi
  400d06:       callq  401130 <_ZN7RegularD1Ev>
  400d0b:       mov    %rbx,%rdi
  400d0e:       callq  401160 <_ZdlPv>
  400d13:       mov    $0x8,%edi

  // Allocate a new Virtual object with the global ::operator new
  400d18:       callq  4016a0 <_Znwm>
  400d1d:       mov    %rax,%rbx
  400d20:       mov    %rax,%rdi

  // Call the constructor for Virtual. We didn't define a default
  // constructor, but the compiler did - to populate the vtable pointer
  // properly.
  400d23:       callq  401150 <_ZN7VirtualC1Ev>

  // If hv != nullptr, call the deleting destructor of Virtual through the
  // virtual table. Do not call operator delete for vr; this will be done by
  // the deleting destructor.
  400d28:       test   %rbx,%rbx
  400d2b:       je     400d36 <main+0x46>
  400d2d:       mov    (%rbx),%rax
  400d30:       mov    %rbx,%rdi
  400d33:       callq  *0x8(%rax)
  400d36:       xor    %eax,%eax
  400d38:       pop    %rbx
  400d39:       retq
  400d3a:       nopw   0x0(%rax,%rax,1)

The key difference here is that for deleting Regular, the compiler inserts a call to the (global) operator delete after the destructor. However, for Virtual it can't do that so it just calls the deleting destructor, which will take care of the deletion as we've seen earlier.


[1]Why this set of options? Without -O2, the code produced by the compiler is overly verbose. With -O2 it's much better but most function calls are inlined, making the special calls generated for the deleting destructor hard to follow; hence -fno-inline. I'm also disabling exceptions because these complicate the code around destructors without being relevant to the main goal of the article.
[2]

One of the derived classes may declare its own operator delete, and the compiler doesn't know that. In fact, a pointer to a derived class can come from a shared library that was built completely separately from the main program (as this sample demonstrates ).

But even if none of the derived classes defines a custom operator delete, it's important to know the dynamic type of the deleted object when the destructor is called to pass the correct address to the global operator delete. An interesting discussion of this issue can be found in this Reddit comment thread.