Proposed resolution:

This wording is relative to N4810.

1. Modify [atomics.ref.int] as indicated:

   integral operator op=(integral operand) const noexcept;
   

   -7- Effects: Equivalent to: return static_cast<integral>(static_cast<make_unsigned_t<integral>>(fetch_key(operand)) op static_cast<make_unsigned_t<integral>>(operand));

2. Modify [atomics.ref.memop] as indicated:

   T* operator++() const noexcept;
   

   -3- Effects: Equivalent to: return static_cast<T>(static_cast<make_unsigned_t<T>>(fetch_add(1)) + static_cast<make_unsigned_t<T>>(1));

   T* operator--(int) const noexcept;
   

   -4- Effects: Equivalent to: return static_cast<T>(static_cast<make_unsigned_t<T>>(fetch_sub(1)) - static_cast<make_unsigned_t<T>>(1));

3. Modify [atomics.types.int] as indicated:

   T operator op=(T operand) volatile noexcept;
   T operator op=(T operand) noexcept;
   

   -8- Effects: Equivalent to: return static_cast<T>(static_cast<make_unsigned_t<T>>(fetch_key(operand)) op static_cast<make_unsigned_t<T>>(operand));

   [Drafting note: atomic<integral>'s working for operator++/operator-- is shared with atomic<T*>. — end drafting note]

   [Drafting note: atomic<floating-point> seems to be correct, LWG should confirm that it is. — end drafting note]

4. Modify [atomics.types.pointer] as indicated:

   T* operator op=(ptrdiff_t operand) volatile noexcept;
   T* operator op=(ptrdiff_t operand) noexcept;
   

   -8- Effects: Equivalent to: return reinterpret_cast<T*>(reinterpret_cast<ptrdiff_t>(fetch_key(operand)) op operand);

   Remarks: The result may be an undefined address, but the operations otherwise have no undefined behavior.

5. Modify [atomics.types.memop] as indicated:

   T operator++() volatile noexcept;
   T operator++() noexcept;
   

   -3- Effects: Equivalent to: return static_cast<T>(static_cast<make_unsigned_t<T>>(fetch_add(1)) + static_cast<make_unsigned_t<T>>(1));

   T operator--() volatile noexcept;
   T operator--() noexcept;
   

   -4- Effects: Equivalent to: return static_cast<T>(static_cast<make_unsigned_t<T>>(fetch_sub(1)) - static_cast<make_unsigned_t<T>>(1));

   [Drafting note: Alternatively, LWG may want to separate the integral overload of operator++/operator-- from that of atomic<T*>. end drafting note]

[ 2019-04-15; JF Bastien comments and provides wording ]

As discussed by LWG during the San Diego 2018 meeting, Jens removed LWG 3047 from "P1236R1: Alternative Wording for P 0907R4 Signed Integers are Two's Complement".

[ 2018-01; Priority set to 3 after mailing list discussion ]

Given atomic<int> meow{INT_MAX};, meow.fetch_add(1) has well-defined behavior because [atomics.types.int] p7 says that

Remarks: For signed integer types, arithmetic is defined to use two's complement representation. There are no undefined results.

A similar issue affects the atomic<T*> partial specialization for pointers.