In a few places, the header uses identifiers with leading underscore and capital letter, which are reserved.

https://github.com/mongodb/mongo/blob/63e43f1bb47f7bddf3dc37ad03a2bbee6d2a9423/src/mongo/platform/atomic_word.h

_ATOMIC_WORD_DECLARE(AtomicUInt32, unsigned);
_ATOMIC_WORD_DECLARE(AtomicUInt64, unsigned long long);
_ATOMIC_WORD_DECLARE(AtomicInt32, int);
_ATOMIC_WORD_DECLARE(AtomicInt64, long long);
_ATOMIC_WORD_DECLARE(AtomicBool, bool);

People think AtomicInt32 wraps int32_t and it doesn't.
This is benign for the AtomicInt32 case, because it's easy to believe that and still be right because 'int32_t' is 'int'.
But for the 64-bit cases, you'll be wrong sometimes. e.g. std::int64_t is long on Linux. It's only long long on Windows.

Also we don't need the _ATOMIC_WORD_DECLARE structure.
All it's doing is providing static asserts that could be inside the AtomicWord class.