[cfe-dev] Microsoft vtordisp field support for clang.

[John McCall]

On Feb 15, 2012, at 12:12 AM, r4start wrote:

> On 15/02/2012 05:52, John McCall wrote:
>> On Feb 14, 2012, at 3:40 AM, r4start wrote:
>>> On 14/02/2012 15:34, r4start wrote:
>>>> On 14/02/2012 05:06, John McCall wrote:
>>>>> class A { virtual void foo(); } class B : public A { }; class C : public virtual B { C(); virtual void foo(); };
>>>> For this test we will get next result:
>>>>  class C    size(12):
>>>>        +---
>>>>  0    | {vbptr}
>>>>        +---
>>>>  4    | (vtordisp for vbase B)
>>>>        +--- (virtual base B)
>>>>        | +--- (base class A)
>>>>  8    | | {vfptr}
>>>>        | +---
>>>>        +---
>>>> If somewhere in inheritance path class has declare ctor or dtor and overload virtual method from virtual base,
>>>> then before virtual base Microsoft compiler push vtordisp in layout.
>>>> 
>>> Also for this test case:
>>> 
>>> class A { virtual void foo(){} };
>>> class B : public A { };
>>> class C : public virtual B { C(){} virtual void foo(){} };
>>> class D : public C { int qqq; };
>>> 
>>>  class D    size(16):
>>>         +---
>>>         | +--- (base class C)
>>>   0    | | {vbptr}
>>>         | +---
>>>   4    | qqq
>>>         +---
>>>   8    | (vtordisp for vbase B)
>>>         +--- (virtual base B)
>>>         | +--- (base class A)
>>>  12   | | {vfptr}
>>>         | +---
>>>         +---
>>> As you can see D also has vtordisp for B.
>> Okay.  I don't think this fundamentally changes the algorithm.  I did think of
>> a case that might matter, though:  a superclass might override a method
>> from a non-virtual base that is *also* a virtual base in some derived class.
>> That is to say, is there a vtordisp in 'C' in the following code?
>> 
>> class A { virtual void foo(); };
>> class B: public A { virtual void foo(); B(); };
>> class C : public B, public virtual A { C(); };
>> 
> Class C does not have vtordisp, because C does not override 'foo' and B inherit A non-virtual.

Okay.  I was just noticing that 'C' installs a construction v-table for A, and
that v-table has to include a thunk for foo() that's really in exactly the
same situation as a thunk would be if B inherited from A virtually.  I guess
Microsoft just gave up in this case.

John.