question on the virtual base offset

[Yan Liu <yanliu@xxxxxxxxxx>]

I found that g++ 3.3, 3.4 and 3.5 have different class layout for the
following testcase:
struct A {char a;};

struct B: virtual public A {};

struct C:public B{
  int c;
  virtual void f() {};
};
struct D {
        char d;
};
struct E {int:3;};
struct F: public C, virtual E, public D {
        char f1;
        virtual void f(){}
};
int main()
{F f;}

Attached files are the class and vft  layout results with three g++
compilers using -fdump-class-hierarchy option.


(See attached file: temp4.g++35.class)(See attached file:
temp4.g++33.class)(See attached file: temp4.C.g++34_32.class)
According to the C++ ABI online document chapter 2.4
(http://www.codesourcery.com/cxx-abi/abi.html).  After mapping the virtual
base class A while mapping of the most derived class F,  the dsize(F)=11,
size(F)=11, align(F)=4, and nvsize(E)=4, nvalign(E)=4.  So, when mapping
virtual base class E, the next available bits are at offset dsize(F)=11.  E
should be put at offset(F) which is the dsize(F) =11, and the final class
size should be 12. To me, g++3.4 gives the most reasonable result, I don't
understand why g++3.5 put E at offset 12, and the final class size is 16.
Is it means that ABI has changed or just a gcc bug?

Your kind help is highly appreciated. Thanks.

Attachment: temp4.g++35.class
Description: Binary data

Attachment: temp4.g++33.class
Description: Binary data

Attachment: temp4.C.g++34_32.class
Description: Binary data