Does making a derived C++ class "final" change the ABI?

Final on a function declaration X::f() implies that the declaration cannot be overridden, so all calls that name that declaration can be bound early (not those calls that name a declaration in a base class): if a virtual function is final in the ABI, the produced vtables can be incompatible with the one produced almost same class without final: calls to virtual functions that name declarations marked final can be assumed to be direct: trying to use a vtable entry (that should exist in the final-less ABI) is illegal.

The compiler could use the final guarantee to cut on the size of vtables (that can sometime grow a lot) by not adding a new entry that would be usually be added and that must be according to the ABI for non final declaration.

Entries are added for a declaration overriding a function not a (inherently, always) primary base or for a non trivially covariant return type (a return type covariant on a non primary base).

Inherently primary base class: the simplest case of polymorphic inheritance

The simple case of polymorphic inheritance, a derived class inheriting non virtually from a single polymorphic base class, is the typical case of an always primary base: the polymorphic base subobject is at the beginning, the address of derived object is the same as the address of the base subobject, virtual calls can be made directly with a pointer to either, everything is simple.

These properties are true whether the derived class is a complete object (one that isn't a subobject), a most derived object, or a base class. (They are class invariants guaranteed at the ABI level for pointers of unknown origin.)

Considering the case where the return type isn't covariant; or:

Trivial covariance

An example: the case where it's covariant with the same type as *this; as in:

struct B { virtual B *f(); };
struct D : B { virtual D *f(); }; // trivial covariance

Here B is inherently, invariably the primary in D: in all D (sub)objects ever created, a B resides at the same address: the D* to B* conversion is trivial so the covariance is also trivial: it's a static typing issue.

Whenever this is the case (trivial up-cast), covariance disappears at the code generation level.

Conclusion

In these cases the type of the declaration of the overriding function is trivially different from the type of the base:

• all parameters are almost the same (with only a trivial difference on the type of this)
• the return type is almost the same (with only a possible difference on the type of a returned pointer(*) type)

(*) since returning a reference is exactly the same as returning a pointer at the ABI level, references aren't discussed specifically

So no vtable entry is added for the derived declaration.

(So making the class final wouldn't be vtable simplification.)

Never primary base

Obviously a class can only have one subobject, containing a specific scalar data member (like the vptr (*)), at offset 0. Other base classes with scalar data members will be at a non trivial offset, requiring non trivial derived to base conversions of pointers. So multiple interesting(**) inheritance will create non primary bases.

(*) The vptr isn't a normal data member at the user level; but in the generated code, it's pretty much a normal scalar data member known to the compiler. (**) The layout of non polymorphic bases isn't interesting here: for the purpose of vtable ABI, a non polymorphic base is treated like a member subobject, as it doesn't affect the vtables in any way.

The conceptually simplest interesting example of a non primary, and non trivial pointer conversion is:

struct B1 { virtual void f(); };
struct B2 { virtual void f(); };
struct D : B1, B2 { };

Each base has its own vptr scalar member, and these vptr have different purposes:

• B1::vptr points to a B1_vtable structure
• B2::vptr points to a B2_vtable structure

and these have identical layout (because the class definitions are superposable, the ABI must generate superposable layouts); and they are strictly incompatible because

1. The vtables have distinct entries:

   • B1_vtable.f_ptr points to the final overrider for B1::f()
   • B2_vtable.f_ptr points to the final overrider for B2::f()

2. B1_vtable.f_ptr must be at the same offset as B2_vtable.f_ptr (from their respective vptr data members in B1 and B2)

3. The final overriders of B1::f() and B2::f() aren't inherently (always, invariably) equivalent(*): they can have distinct final overriders that do different things.(***)

(*) Two callable runtime functions(**) are equivalent if they have same observable behavior at the ABI level. (Equivalent callable functions may not have the same declaration or C++ types.)

(**) A callable runtime function is any entry point: any address that can be called/jumped at; it can be a normal function code, a thunk/trampoline, a particular entry in a multiple entry function. Callable runtime functions often have no possible C++ declarations, like "final overrider called with a base class pointer".

(***) That they sometimes have the same final overrider in a further derived class:

struct DD : D { void f(); }

isn't useful for the purpose of defining the ABI of D.

So we see that D provably needs a non primary polymorphic base; by convention it will be D2; the first nominated polymorphic base (B1) gets to be primary.

So B2 must be at non trivial offset, and D to B2 conversion is non trivial: it requires generated code.

So the parameters of a member function of D cannot be equivalent with the parameters of a member function of B2, as the implicit this isn't trivially convertible; so:

• D must have two different vtables: a vtable corresponding with B1_vtable and one with B2_vtable (they are in practice put together in one big vtable for D but conceptually they are two distinct structures).
• the vtable entry of a virtual member of B2::g that is overridden in D needs two entries, one in the D_B2_vtable (which is just a B2_vtable layout with different values) and one in the D_B1_vtable which is an enhanced B1_vtable: a B1_vtable plus entries for new runtime features of D.

Because the D_B1_vtable is built from a B1_vtable, a pointer to D_B1_vtable is trivially a pointer to a B1_vtable, and the vptr value is the same.

Note that in theory is would be possible to omit the entry for D::g() in D_B1_vtable if the burden of making all virtual calls of D::g() via the B2 base, which as far as no non trivial covariance is used(#), is also a possibility.

(#) or if non trivial covariance occurs, "virtual covariance" (covariance in a derived to base relation involving virtual inheritance) isn't used

Not inherently primary base

Regular (non virtual) inheritance is simple like membership:

• a non virtual base subobject is a direct base of exactly one object (which implies that there always exactly one final overrider of any virtual function when virtual inheritance isn't used);
• the placement of a non virtual base is fixed;
• base subobject that don't have virtual base subobjects, just like data member, are constructed exactly like complete objects (they have exactly one runtime constructor function code for every defined C++ constructor).

A more subtle case of inheritance is virtual inheritance: a virtual base subobject can be the direct base of many base class subobjects. That implies that the layout of virtual bases is only determined at the most derived class level: the offset of a virtual base in a most derived object is well known and a compile time constant; in a arbitrary derived class object (that may or may not be a most derived object) it is a value computed at runtime.

That offset can never be known because C++ supports both unifying and duplicating inheritance:

• virtual inheritance is unifying: all virtual bases of a given type in a most derived object are one and the same subobject;

• non virtual inheritance is duplicating: all indirect non virtual bases are semantically distinct, as their virtual members don't need to have common final overriders (contrast with Java where this is impossible (AFAIK)):

  struct B { virtual void f(); }; struct D1 : B { virtual void f(); }; // final overrider struct D2 : B { virtual void f(); }; // final overrider struct DD : D1, D2 { };

Here DD has two distinct final overriders of B::f():

• DD::D1::f() is final overrider for DD::D1::B::f()
• DD::D2::f() is final overrider for DD::D2::B::f()

in two distinct vtable entries.

Duplicating inheritance, where you indirectly derive multiple times from a given class, implies multiple vptrs, vtables and possibly distinct vtable ultimate code (the ultimate aim of using a vtable entry: the high level semantic of calling a virtual function - not the entry point).

Not only C++ supports both, but the fact combinations are allowed: duplicating inheritance of a class that uses unifying inheritance:

struct VB { virtual void f(); };
struct D : virtual VB { virtual void g(); int dummy; };
struct DD1 : D { void g(); };
struct DD2 : D { void g(); };
struct DDD : DD1, DD2 { };

There is only one DDD::VB but there are two observably distinct D subobjects in DDD with different final overriders for D::g(). Whether or not a C++-like language (that supports virtual and non virtual inheritance semantic) guarantees that distinct subobjects have different addresses, the address of DDD::DD1::D cannot be at the same as the address of DDD::DD2::D.

So the offset of a VB in a D cannot be fixed (in any language that supports unification and duplication of bases).

In that particular example a real VB object (the object at runtime) has no concrete data member except the vptr, and the vptr is a special scalar member as it is a type "invariant" (not const) shared member: it is fixed on the constructor (invariant after complete construction) and its semantic is shared between bases and derived classes. Because VB has no scalar member that isn't type invariant, that in a DDD the VB subobject can be an overlay over DDD::DD1::D, as long as the vtable of D is a match for the vtable of VB.

This however cannot be the case for virtual bases that have non invariant scalar members, that is regular data members with an identity, that is members occupying a distinct range of bytes: these "real" data members cannot be overlayed on anything else. So a virtual base subobject with data members (members with with an address guaranteed to be distinct by C++ or any other the distinct C++-like language you are implementing) must be put at a distinct location: virtual bases with data members normally(##) have inherently non trivial offsets.

(##) with potentially a very narrow special case with a derived class with no data member with a virtual base with some data members

So we see that "almost empty" classes (classes with no data member but with a vptr) are special cases when used as virtual base classes: these virtual base are candidate for overlaying on derived classes, they are potential primaries but not inherent primaries:

• the offset at which they reside will only be determined in the most derived class;
• the offset might or might not be zero;
• a nul offset implies overlaying of the base, so the vtable of each directly derived class must be a match for the vtable of the base;
• a non nul offset implies non trivial conversions, so the entries in the vtables must treat conversion of the pointers to the virtual base as needing a runtime conversion (except when overlaid obviously as it wouldn't be necessary not possible).

This means that when overriding a virtual function in a virtual base, an adjustment is always assumed to be potentially needed, but in some cases no adjustment will be needed.

A morally virtual base is a base class relationship that involves a virtual inheritance (possibly plus non virtual inheritance). Performing a derived to base conversion, specifically converting a pointer d to derived D, to base B, a conversion to...

• ...a non-morally virtual base is inherently reversible in every case:

  • there is a one to one relation between the identity of a subobject B of a D and a D (which might be a subobject itself);
  • the reverse operation can be performed with a static_cast<D*>: static_cast<D*>((B*)d) is d;

• (in any C++ like language with complete support for unifying and duplicating inheritance) ...a morally virtual base is inherently non reversible in the general case (although it's reversible in common case with simple hierarchies). Note that:

  • static_cast<D*>((B*)d) is ill formed;
  • dynamic_cast<D*>((B*)d) will work for the simple cases.

So let's called virtual covariance the case where the covariance of the return type is based on morally virtual base. When overriding with virtual covariance, the calling convention cannot assume the base will be at a known offset. So a new vtable entry is inherently needed for virtual covariance, whether or not the overridden declaration is in an inherent primary:

struct VB { virtual void f(); }; // almost empty
struct D : virtual VB { }; // VB is potential primary

struct Ba { virtual VB * g(); };
struct Da : Ba { // non virtual base, so Ba is inherent primary
  D * g(); // virtually covariant: D->VB is morally virtual
};

Here VB may be at offset zero in D and no adjustment may be needed (for example for a complete object of type D), but it isn't always the case in a D subobject: when dealing with pointers to D, one cannot know whether that is the case.

When Da::g() overrides Ba::g() with virtual covariance, the general case must be assumed so a new vtable entry is strictly needed for Da::g() as there is no possible down pointer conversion from VB to D that reverses the D to VB pointer conversion in the general case.

Ba is an inherent primary in Da so the semantics of Ba::vptr are shared/enhanced:

• there are additional guarantees/invariants on that scalar member, and the vtable is extended;
• no new vptr is needed for Da.

So the Da_vtable (inherently compatible with Ba_vtable) needs two distinct entries for virtual calls to g():

• in the Ba_vtable part of the vtable: Ba::g() vtable entry: calls final overrider of Ba::g() with an implicit this parameter of Ba* and returns a VB* value.
• in the new members part of the vtable: Da::g() vtable entry: calls final overrider of Da::g() (which by is inherently the same as final overrider of Ba::g() in C++) with an implicit this parameter of Da* and returns a D* value.

Note that there is not really any ABI freedom here: the fundamentals of vptr/vtable design and their intrinsic properties imply the presence of these multiple entries for what is a unique virtual function at the high language level.

Note that making the virtual function body inline and a visible by the ABI (so that the ABI by classes with different inline function definitions could be made incompatible, allowing more information to inform memory layout) wouldn't possibly help, as inline code would only define what a call to a non overridden virtual function does: one cannot based the ABI decisions on choices that can be overridden in derived classes.

[Example of a virtual covariance that ends up being only trivially covariant as in a complete D the offset for VB is trivial and no adjustment code would have been necessary in that case:

struct Da : Ba { // non virtual base, so inherent primary
  D * g() { return new D; } // VB really is primary in complete D
                            // so conversion to VB* is trivial here
};

Note that in that code an incorrect code generation for a virtual call by a buggy compiler that would use the Ba_vtable entry to call g() would actually work because covariance ends up being trivial, as VB is primary in complete D.

The calling convention is for the general case and such code generation would fail with code that returns an object of a different class.

--end example]

But if Da::g() is final in the ABI, only virtual calls can be made via the VB * g(); declaration: covariance is made purely static, the derived to base conversion is be done at compile time as the last step of the virtual thunk, as if virtual covariance was never used.

Possible extension of final

There are two types of virtual-ness in C++: member functions (matched by function signature) and inheritance (match by class name). If final stops overriding a virtual function, could it be applied to base classes in a C++-like language?

First we need to define what is overriding a virtual base inheritance:

An "almost direct" subobject relation means that a indirect subobject is controlled almost as a direct subobject:

• an almost direct subobject can be initialized like a direct subobject;
• access control is never a really obstacle to access (inaccessible private almost direct subobjects can be made accessible at discretion).

Virtual inheritance provides almost direct access:

• constructor for each virtual bases must be called by ctor-init-list of the constructor of the most derived class;
• when a virtual base class is inaccessible because declared private in a base class, or publicly inherited in a private base class of a base class, the derived class has the discretion to declare the virtual base as a virtual base again, making it accessible.

A way to formalize virtual base overriding is to make an imaginary inheritance declaration in each derived class that overrides base class virtual inheritance declarations:

struct VB { virtual void f(); };
struct D : virtual VB { };
struct DD : D
  // , virtual VB  // imaginary overrider of D inheritance of VB
  {
  // DD () : VB() { } // implicit definition
}; 

Now C++ variants that support both forms of inheritance don't have to have C++ semantic of almost direct access in all derived classes:

struct VB { virtual void f(); };
struct D : virtual VB { };
struct DD : D, virtual final VB {
  // DD () : VB() { } // implicit definition
}; 

Here the virtual-ness of the VB base is frozen and cannot be used in further derived classes; the virtual-ness is made invisible and inaccessible to derived classes and the location of VB is fixed.

struct DDD : DD {
  DD () : 
    VB() // error: not an almost direct subobject
  { } 
}; 
struct DD2 : D, virtual final VB {
  // DD2 () : VB() { } // implicit definition
}; 
struct Diamond : DD, DD2 // error: no unique final overrider
{                        // for ": virtual VB"
}; 

The virtual-ness freeze makes it illegal to unify Diamond::DD::VB and Diamond::DD2::VB but virtual-ness of VB requires unification which makes Diamond a contradictory, illegal class definition: no class can ever derive from both DD and DD2 [analog/example: just like no useful class can directly derive from A1 and A2:

struct A1 {
  virtual int f() = 0;
};
struct A2 {
  virtual unsigned f() = 0;
};
struct UselessAbstract : A1, A2 {
  // no possible declaration of f() here
  // none of the inherited virtual functions can be overridden
  // in UselessAbstract or any derived class
};

Here UselessAbstract is abstract and no derived class are too, making that ABC (abstract base class) extremely silly, as any pointer to UselessAbstract is provably a null pointer.

-- end analog/example]

That would provide a way to freeze virtual inheritance, to provide meaningful private inheritance of classes with virtual base (without it derived classes can usurp the relationship between a class and its private base class).

Such use of final would of course freeze the location of a virtual base in a derived class and its further derived classes, avoiding additional vtable entries that are only needed because the location of virtual base isn't fixed.

I believe that adding the final keyword should not be ABI breaking, however removing it from an existing class might render some optimizations invalid. For example, consider this:

// in car.h
struct Vehicle { virtual void honk() { } };
struct Car final : Vehicle { void honk() override { } };

// in car.cpp

// Here, the compiler can assume that no derived class of Car can be passed,
// and so `honk()` can be devirtualized. However, if Car is not final
// anymore, this optimization is invalid.
void foo(Car* car) { car->honk(); }

If foo is compiled separately and e.g. shipped in a shared library, removing final (and hence making it possible for users to derive from Car) could render the optimization invalid.

I'm not 100% sure about this though, some of it is speculation.