[libc-commits] [libc] [libc][x86] Use prefetch for write for memcpy (PR #90450)

[Guillaume Chatelet via libc-commits]

gchatelet wrote:

Reviving this PR. The current microbenchmarks available (https://github.com/google/fleetbench and `libc/benchmarks`) are not able to show the impact of this change. Although we do have supporting evidence internally that this helps in certain configurations.
Specifically, when servers with many cores are under high utilization, DRAM congestion becomes a performance bottleneck. For these servers, disabling hardware prefetchers help reclaim some of the bandwidth. This - in turn - lowers the performance of memory functions as they highly rely on the streaming effect of hardware prefetchers. This patch helps recover some of this loss. More context is available in : https://dl.acm.org/doi/10.1145/3620666.3651373.

https://github.com/llvm/llvm-project/pull/90450