Re: [committed] libstdc++: Optimize ref-count updates in COW std::string

[Florian Weimer <fweimer@redhat.com>]

* Jonathan Wakely via Libstdc:

> diff --git a/libstdc++-v3/include/bits/cow_string.h b/libstdc++-v3/include/bits/cow_string.h
> index ced395b80b8..4fae1d02981 100644
> --- a/libstdc++-v3/include/bits/cow_string.h
> +++ b/libstdc++-v3/include/bits/cow_string.h
> @@ -105,7 +105,7 @@ _GLIBCXX_BEGIN_NAMESPACE_VERSION
>     *  destroy the empty-string _Rep object.
>     *
>     *  All but the last paragraph is considered pretty conventional
> -   *  for a C++ string implementation.
> +   *  for a Copy-On-Write C++ string implementation.
>    */
>    // 21.3  Template class basic_string
>    template<typename _CharT, typename _Traits, typename _Alloc>
> @@ -207,10 +207,10 @@ _GLIBCXX_BEGIN_NAMESPACE_VERSION
>  	  // so we need to use an atomic load. However, _M_is_leaked
>  	  // predicate does not change concurrently (i.e. the string is either
>  	  // leaked or not), so a relaxed load is enough.
> -	  return __atomic_load_n(&this->_M_refcount, __ATOMIC_RELAXED) < 0;
> -#else
> -	  return this->_M_refcount < 0;
> +	  if (!__gnu_cxx::__is_single_threaded())
> +	    return __atomic_load_n(&this->_M_refcount, __ATOMIC_RELAXED) < 0;
>  #endif
> +	  return this->_M_refcount < 0;
>  	}

Relaxed MO loads of word-size values on all current architectures only
have a compiler barrier, so I think the optimization makes things worse?
(I doubt the conditional lack of a compiler barrier leads to
optimization improvements elsewhere.)

Thanks,
Florian