Why Does `std::atomic`\'s Store Use XCHG for Sequential Consistency?

Why does std::atomic's store use XCHG when sequential consistency is requested?

While a sequential-consistency store operation may appear at first as something that can be implemented using a simple store instruction followed by a memory barrier, using XCHG (with an implied lock prefix) offers a more optimal solution.

XCHG and Sequential Consistency

On x86 and x86_64 architectures, XCHG provides both a memory exchange and a full memory barrier, ensuring sequential ordering of memory operations. This makes it an effective way to implement sequential consistency stores.

Limitations of Regular Store Instructions

A regular store instruction alone, such as MOV, can not guarantee sequential consistency. It only provides release semantics, which allows it to be reordered with later operations, including acquire loads.

Performance Considerations

While both MFENCE and XCHG can be used to implement sequential-consistency stores, they exhibit different performance characteristics on various CPUs. On certain CPUs, MFENCE can stall out-of-order execution of neighboring instructions. In contrast, XCHG may be more efficient for single-threaded operations or when the cached line is hot in L1 cache.

Compiler and Kernel Practices

Various compilers and operating systems prefer different approaches for implementing sequential-consistency stores. GCC historically used MOV MFENCE, while other compilers and the Linux kernel employ XCHG.

Additional Notes

It's important to note that:

• Implicit Acquire Fences: You may have mentioned that x86 stores have an implicit acquire fence, but this is incorrect. Stores implicitly have release semantics, not acquire semantics.
• Compiler Barriers: The asm volatile("" ::: "memory") syntax you provided is a compiler-level barrier, not a runtime memory barrier.
• Fences vs. Release Operations: While fences and release operations can share similarities, they are distinct mechanisms in x86.

The above is the detailed content of Why Does `std::atomic`\'s Store Use XCHG for Sequential Consistency?. For more information, please follow other related articles on the PHP Chinese website!