AMD Zen 2 Improvements For LLVM Have Been Held Up For Months By Code Review

Back in February for LLVM Clang 9.0 was the initial AMD Zen 2 "znver2" enablement, but like the GCC support at the time it was the very basics. With time GCC picked up Zen 2 scheduler improvements and other work while sadly in the case of LLVM the improvements are still pending.

Back in August, AMD's Ganesh Gopalasubramanian sent out the znver2 scheduler model for LLVM for Zen 2 CPUs but a focus on the EPYC 7002 "Rome" processors. "There are few improvements with respect to execution units, latencies and throughput when compared with znver1. The tests that were present for znver1 for llvm-mca tool were replicated. The latencies, execution units, timeline and throughput information are updated for znver2."

That's all great with AMD Ryzen 3000 and EPYC 7002 series processors now shipping for months. But sadly that Zen 2 scheduler model hasn't been merged yet with being held up by needing code review. Hopefully that review can happen soon -- especially with AMD having contributed the code themselves -- in order to make it for the LLVM Clang 10.0 release in early 2020 but a pity that it's been held up while these new AMD processors have already been shipping for months.

For those wishing to have an AMD Zen 2 optimized compiler experience right now without resorting to patching your own LLVM/Clang setup, the AMD Optimizing C/C++ Compiler 2.0 release from August does have all of their LLVM patches in place for the znver2 target.

Moving forward, this is one area where AMD can improve upon with more punctual GCC/Clang compiler support in place. After all, Intel tends to upstream their microarchitecture enablement for the compilers usually at least one year in advance to ensure that the optimized support works its way into released versions of the compilers -- and found within shipping Linux distributions -- by the time their CPUs ship.