AMD's Zen 2 Scheduler Model Gets Partially Fixed Up In LLVM

Landing in the LLVM compiler infrastructure code-base in January was finally an AMD Zen 2 scheduler model optimized for the latest-generation AMD processors when compiling code with Clang using the -march=znver2 targeting. However, now some important fixes to this scheduler model have landed.

It turns out there were some misunderstandings / typos in the Znver2 scheduler model that was merged in January. A commit last week to LLVM has fixed the "obvious mistakes" to the AMD Zen 2 model including for latencies but other work remains to be addressed.

The AMD Zen 2 scheduler model defects may at least partially explain some of the LLVM Clang 10 performance differences we saw last month when testing the original scheduler model addition.