When Should You Use _mm_sfence, _mm_lfence, and _mm_mfence?

Multi-threaded programming introduces concurrency-related complexities, necessitating mechanisms to maintain data integrity and synchronization. Intel's intrinsics library provides several functions, including _mm_sfence, _mm_lfence, and _mm_mfence, to control memory ordering in x86 architectures.

Memory Ordering in x86

x86 CPUs have a strongly ordered memory model, but C and C have weaker ones. Hence, additional precautions are required to ensure proper memory ordering and prevent data corruption or race conditions.

_mm_sfence

_mm_sfence is primarily used after non-temporal (NT) stores (_mm_stream_*) to prevent speculative reordering. NT stores are weakly ordered, meaning they can appear to occur out of order relative to other memory operations. _mm_sfence creates a barrier that ensures subsequent memory operations become globally visible after the NT stores are committed to memory.

_mm_lfence

_mm_lfence is rarely used as a load fence. It only has relevance when loading from Write-Combining (WC) memory regions, such as video RAM. _mm_lfence can prevent execution of subsequent instructions until it retires, which can be useful for microbenchmarking.

_mm_mfence

_mm_mfence provides sequential consistency, ensuring subsequent loads cannot read values until after preceding stores become globally visible. It can be useful if you implement your custom version of std::atomic or need to explicitly control memory ordering for operations that would otherwise be speculative.

Summary

• Use _mm_sfence after NT stores to prevent data corruption and race conditions.
• Avoid _mm_lfence for load ordering unless specifically working with WC memory regions.
• _mm_mfence offers sequential consistency but may be less efficient than locked atomic read-modify-write operations.
• Consider using C 11 std::atomic or C11 stdatomic for memory synchronization, as they provide a more convenient and optimized approach.