[cfe-dev] libc++ ABI versioning

[Marshall Clow]

On Jun 11, 2014, at 10:19 AM, Jorgen Tjerno <jorgen at uberent.com> wrote:
> On Wed, Jun 11, 2014 at 6:48 AM, Marshall Clow <mclow.lists at gmail.com> wrote:
>> 
>> On Jun 9, 2014, at 7:26 PM, Jorgen Tjerno <jorgen at uberent.com> wrote:
>>> On Mon, Jun 9, 2014 at 6:56 PM, Richard Smith <richard at metafoo.co.uk> wrote:
>>>> On Mon, Jun 9, 2014 at 6:32 PM, Jorgen Tjerno <jorgen at uberent.com> wrote:
>>>>> 
>>>>> I've noticed that Xcode 4.6 and Xcode 5.1 ship with headers that have
>>>>> different layouts for std::ifstream. sizeof(std::ifstream) is 448 for
>>>>> 4.6, and 576 on 5.1. You can see my test case here:
>>>>> https://gist.github.com/anonymous/bd0db5e4bcb307b70506
>>>>> 
>>>>> How does clang/libc++ ensure that the executable runs with the right
>>>>> libc++ version for headers it was built with? I can run the binaries
>>>>> produced by Xcode 4.6 on a machine that has a newer libc++, and I can
>>>>> run the binaries produced by 5.1 on a machine that has an older
>>>>> libc++. I would assume that if I built with headers that give me a
>>>>> smaller size than what is expected by libc++, that would be cause the
>>>>> newer libc++ to trash memory beyond the end of my object. Am I missing
>>>>> something?
>>>>> 
>>>>> I was unable to find any resources on how libc++ ABI versioning works
>>>>> other than the "major ABI version" -- which is 1.
>>>>> 
>>>>> The specific problem I'm seeing is that it seems that when I link with
>>>>> a static library that was built with the older libc++ version, a
>>>>> translation unit built with the new libc++ ends up using the wrong
>>>>> (old) value for the construction symbol at runtime
>>>>> 
>>>>> (__ZTCNSt3__114basic_ifstreamIcNS_11char_traitsIcEEEE0_NS_13basic_istreamIcS2_EE,
>>>>> or '
>>>>>  construction vtable for std::__1::basic_istream<char,
>>>>> std::__1::char_traits<char> >-in-std::__1::basic_ifstream<char,
>>>>> std::__1::char_traits<char> >' according to c++filt).
>>>> 
>>>> 
>>>> Looks like the difference is probably due to the std::filebuf subobject. Did
>>>> that type also change size? Did std::basic_streambuf<char>?
>>> 
>>> Yep, filebuf did, and that does seem to account for the difference in size.
>>> 
>>> themac:~ jorgen$ ./foo-modern
>>> sizeof(std::ifstream)=576, offset=0
>>> sizeof(std::istream)=168, offset=0
>>> sizeof(std::ios)=152, offset=424
>>> sizeof(std::ios_base)=136, offset=424
>>> sizeof(std::filebuf)=408, offset=0
>>> sizeof(std::basic_streambuf<char>)=64, offset=0
>>> 
>>> themac:~ jorgen$ ./foo
>>> sizeof(std::ifstream)=448, offset=0
>>> sizeof(std::istream)=168, offset=0
>>> sizeof(std::ios)=152, offset=296
>>> sizeof(std::ios_base)=136, offset=296
>>> sizeof(std::filebuf)=280, offset=0
>>> sizeof(std::basic_streambuf<char>)=64, offset=0
>> 
>> Looks like this commit was the culprit:
>> 
>> r162601 | hhinnant | 2012-08-24 14:20:56 -0700 (Fri, 24 Aug 2012) | 31 lines
>> 
>> It added a new member variable
>>        state_type __st_last_;
>> 
>> to the basic_filebuf object, and that variable is 128 bytes long.
>> 
> So why is this not an ABI-breaking change?
> 
> What causes a program built against the earlier version of
> std::ifstream to correctly execute on a machine that has the latest
> version of std::ifstream that expects a larger layout?

I think it was an ABI-breaking change.
I also believe that it was done w/o realizing that it was breaking the ABI.

— Marshall