Unverified Commit 331070f8 authored by Luna Nova's avatar Luna Nova
Browse files

rocmPackages.llvm: revert loop unrolling pessimisation 

Reverting a patch that caused terrible performance for some
large langle mangle workloads.

See also rocm-systems issue 2865 and rocm llvm-project pull 1349
parent 691e4925

pkgs/development/rocm-modules/llvm/default.nix

+10 −0
Original line number Diff line number Diff line
@@ -295,6 +295,16 @@ in 
overrideLlvmPackagesRocm (s: {

  libllvm = (s.prev.libllvm.override { }).overrideAttrs (old: {

    patches = old.patches ++ [

      (fetchpatch {

        # Revert of a patch that cause perf regression by pessimising loop unrolling decisions

        # See PR and issue discussion

        # https://github.com/ROCm/rocm-systems/issues/2865 https://github.com/ROCm/llvm-project/pull/1349

        name = "rocm-llvm-revert-unrolling-regression.patch";

        url = "https://github.com/ROCm/llvm-project/commit/f58b06dce1f9c15707c5f808fd002e18c2accf7e.patch";

        hash = "sha256-pH+3C7PSDqNfOF014sA5Rvm+sc2IJMQJfysS2bvj/o0=";

        # stripLen instead of relative to avoid filterdiff mangling /dev/null on the deleted test file

        stripLen = 1;

      })

      ./perf-increase-namestring-size.patch

      # TODO: consider reapplying "Don't include aliases in RegisterClassInfo::IgnoreCSRForAllocOrder"

      # it was reverted as it's a pessimization for non-GPU archs, but this compiler