GNU Compiler Collection version 13 gets AMD Zen 4 CPU tuning before release

Jan Hubicka, a SUSE compiler engineer, has added additional AMD Zen 4 code, more specifically “znver4” CPU target tuning, to further prepare for the upcoming release of the GNU Compiler Collection (GCC 13). GCC 13 includes additional and improved AMD Zen 4 CPU target tuning, and numerous patches for the new version have been added in the latest 2022 compiler update patch. Support for the AMD Zen 4 architecture is expected to be better than that for Zen 3, which was mediocre at best. We should also expect new processor ISA extensions to be used for the current architecture, reports Michael Larabel, analyst and editor of Linux hardware website Phoronix.

Yesterday Hubicka added the latest updates for GCC 13, for which we have listed the patch notes below, previously posted on the GNU Compiler Collection’s Git page:

This patch adds more settings for Zen4:

• new settings for avx512 scater instructions.
  in micro benchmarks, these seem to have consistent losses compared to open coded
  disabling the use of gather for zen4
  While this is a win for micro benchmarks (based on TSVC), enabling gather is a loss for Parest. So for now, it seems safe to turn it off.
• Disable pass to avoid FMA chains for znver4, as fmadd has been optimized and does not seem to cause regressions

Larabel notes that GCC 13 and AMD’s proprietary AOCC 4.0 compiler are needed for testing to see which is better “than a downstream of LLVM/Clang.” AMD added the initial Zen 4 architecture enablement in LLVM/Clang early last month, but it hasn’t been touched since. The hope is that we will see more in the coming weeks. The Phoronix editor also mentions that “-march=znver4 can be used if you want to adapt the compiler’s instructions and optimizations to the AMD Ryzen 7000 series and AMD EPYC 9004 series processors”.

At the same time, we are waiting for more updates. As reported yesterday, AMD seems to have focused heavily on AMDGPU work for DRM-Next lately. Several targets were not initially supported, and it is encouraging to see that work is now being added to upcoming patches within the various open source projects in preparation for Q1 releases.

Source: wccftech.com