https://gcc.gnu.org/bugzilla/show_bug.cgi?id=111077
--- Comment #8 from CVS Commits <cvs-commit at gcc dot gnu.org> --- The master branch has been updated by Jonathan Wakely <r...@gcc.gnu.org>: https://gcc.gnu.org/g:dcbec954fcba42d97760c6bd98a4c5618473ec93 commit r14-3625-gdcbec954fcba42d97760c6bd98a4c5618473ec93 Author: Jonathan Wakely <jwak...@redhat.com> Date: Wed Aug 23 12:23:37 2023 +0100 libstdc++: Use a loop in atomic_ref::compare_exchange_strong [PR111077] We need to use a loop in std::atomic_ref::compare_exchange_strong in order to properly implement the C++20 requirement that padding bits do not participate when checking the value for equality. The variable being modified by a std::atomic_ref might have an initial value with non-zero padding bits, so when the __atomic_compare_exchange built-in returns false we need to check whether that was only because of non-equal padding bits that are not part of the value representation. If the value bits differ, it's just a failed compare-exchange. If the value bits are the same, we need to retry the __atomic_compare_exchange using the value that was just read by the previous failed call. As noted in the comments, it's possible for that second try to also fail due to another thread storing the same value but with differences in padding. Because it's undefined to access a variable directly while it's held by a std::atomic_ref, and because std::atomic_ref will only ever store values with zeroed padding, we know that padding bits will never go from zero to non-zero during the lifetime of a std::atomic_ref. They can only go from an initial non-zero state to zero. This means the loop will terminate, rather than looping indefinitely as padding bits flicker on and off. In theory users could call __atomic_store etc. directly and write a value with non-zero padding bits, but we don't need to support that. Users doing that should ensure they do not write non-zero padding, to be compatibile with our std::atomic_ref's invariants. This isn't a problem for std::atomic<T>::compare_exchange_strong because the initial value (and all later stores to the variable) are performed by the library, so we ensure that stored values always have padding bits cleared. That means we can simply clear the padding bits of the 'expected' value and we will be comparing two values with equal padding bits. This means we don't need the loop for std::atomic, so update the __atomic_impl::__compare_exchange function to take a bool parameter that says whether it's being used by std::atomic_ref. If not, we can use a simpler, non-looping implementation. libstdc++-v3/ChangeLog: PR libstdc++/111077 * include/bits/atomic_base.h (__atomic_impl::__compare_exchange): Add _AtomicRef non-type template parameter and use a loop if it is true. (__atomic_impl::compare_exchange_weak): Add _AtomicRef NTTP. (__atomic_impl::compare_exchange_strong): Likewise. (atomic_ref::compare_exchange_weak): Use true for NTTP. (atomic_ref::compare_exchange_strong): Use true for NTTP. * testsuite/29_atomics/atomic_ref/compare_exchange_padding.cc: Fix test to not rely on atomic_ref::load() to return an object with padding preserved.