On Sun, Jan 7, 2024, 5:02 PM xndcn <xndchn@gmail.com> wrote:

> Hi, I found __atomic_float constructor does not clear padding,
> while __compare_exchange assumes it as zeroed padding. So it is easy to
> reproducing a infinite loop in X86-64 with long double type like:
> ---
> -O0 -std=c++23 -mlong-double-80
> #include <stdio.h>
> #include <atomic>
>
> #define T long double
> int main() {
>     std::atomic<T> t(0.5);
>     t.fetch_add(0.5);
>     float x = t;
>     printf("%f\n", x);
> }
> ---
>
> So we should add __builtin_clear_padding in __atomic_float constructor,
> just like the generic atomic struct.
>
> regtested on x86_64-linux. Is it OK for trunk?
>
> ---
> libstdc++: atomic: Add missing clear_padding in __atomic_float constructor.
>
> libstdc++-v3/ChangeLog:
>
> * include/bits/atomic_base.h: add __builtin_clear_padding in
> __atomic_float constructor.
> ---
>  libstdc++-v3/include/bits/atomic_base.h | 7 ++++++-
>  1 file changed, 6 insertions(+), 1 deletion(-)
>
> diff --git a/libstdc++-v3/include/bits/atomic_base.h
> b/libstdc++-v3/include/bits/atomic_base.h
> index f4ce0fa53..d59c2209e 100644
> --- a/libstdc++-v3/include/bits/atomic_base.h
> +++ b/libstdc++-v3/include/bits/atomic_base.h
> @@ -1283,7 +1283,12 @@ _GLIBCXX_BEGIN_NAMESPACE_VERSION
>
>        constexpr
>        __atomic_float(_Fp __t) : _M_fp(__t)
> -      { }
> +      {
> +#if __has_builtin(__builtin_clear_padding)
> + if _GLIBCXX17_CONSTEXPR (__atomic_impl::__maybe_has_padding<_Fp>())
> +  __builtin_clear_padding(std::__addressof(_M_fp));
> +#endif
> +      }
>
>        __atomic_float(const __atomic_float&) = delete;
>        __atomic_float& operator=(const __atomic_float&) = delete;
> --
> 2.25.1
>

Can you add a testcase?

Thanks.

H.J.

>