--enable-libstdcxx-time changes ABI?

[Jonathan Wakely]

On 2 December 2012 04:36, Kenny Simpson wrote:
> Hello,
>   http://gcc.gnu.org/onlinedocs/libstdc++/manual/configure.html documents the --enable-libstdcxx-time options, but, unlike some of the other options, does not mention that it affects ABI.
>
> or maybe I'm not understanding ABI-breakage.
>
> My understanding of the affect of this is that  std::chrono::system_clock::time_point would become a different type.
>
> std::chrono::system_clock::time_point is defined as:
>       typedef chrono::time_point<system_clock, duration>        time_point;
>
> and 'duration' depends on a configuration-time setting based the enable-libstdcxx-time:
>
>     struct system_clock
>     {
> #ifdef _GLIBCXX_USE_CLOCK_REALTIME
>       typedef chrono::nanoseconds                               duration;
> #elif defined(_GLIBCXX_USE_GETTIMEOFDAY)
>       typedef chrono::microseconds                              duration;
> #else
>       typedef chrono::seconds                                   duration;
> #endif
>
> What makes this affect the ABI is system_clock::now:
>       static time_point
>       now() noexcept;
>
> Since only the return type is affected, and since the return type is not part of the mangled name, the exported symbol is not affected.
>
> However, if a program at runtime were to run against a libstdc++.so built with a differing option, the results would be times/durations that are off by large factors - silently.  The program could have compiled assuming nanoseconds, only to run against a libstdc++ that returns microseconds.
>
> As such, it seems it would never be safe to mix code that was built with differing flavors of libstdcxx-time.  If this is true, a note in the documentation would be nice - even if this isn't considered ABI-breakage since the exported symbols aren't affected.

Hi Kenny,

I agree this should be documented, I'll update the docs for
--enable-libstdcxx-time.

N.B. currently the higher resolution system_clock is only available
with --enable-libstdcxx-time=rt which causes libstdc++.so to link to
librt.so, which is a very user-visible change (although maybe not an
obvious one unless you look for it.)  At some point a future version
of GCC when used with glibc 2.17 or later will be able to use
clock_gettime and CLOCK_REALTIME without librt.so, which would
effectively make --enable-libstdcxx-time=rt equivalent to
--enable-libstdcxx-time=yes and we will need to keep in mind that it
can affect the ABI.

I wonder if we want to use the abi tag attribute to make system_clock
mangle differently depending whether it uses nanosecond, microsecond
or second resolution.