[all-commits] [llvm/llvm-project] d83cba: [AMDGPU] Reenable BackOffBarrier on GFX11/12

[Pierre van Houtryve via All-commits]

  Branch: refs/heads/users/pierre-vh/barrier-pair-fence
  Home:   https://github.com/llvm/llvm-project
  Commit: d83cba3079acb688def30e04e0071b5ef6bf885d
      https://github.com/llvm/llvm-project/commit/d83cba3079acb688def30e04e0071b5ef6bf885d
  Author: pvanhout <pierre.vanhoutryve at amd.com>
  Date:   2025-08-26 (Tue, 26 Aug 2025)

  Changed paths:
    M llvm/lib/Target/AMDGPU/AMDGPU.td
    M llvm/lib/Target/AMDGPU/SIInstrInfo.h
    M llvm/lib/Target/AMDGPU/SIMemoryLegalizer.cpp
    M llvm/test/CodeGen/AMDGPU/back-off-barrier-subtarget-feature.ll
    M llvm/test/CodeGen/AMDGPU/lds-dma-workgroup-release.ll
    M llvm/test/CodeGen/AMDGPU/llvm.amdgcn.update.dpp.ll
    A llvm/test/CodeGen/AMDGPU/memory-legalizer-barriers.ll
    M llvm/test/CodeGen/AMDGPU/waitcnt-preexisting-vscnt.mir

  Log Message:
  -----------
  [AMDGPU] Reenable BackOffBarrier on GFX11/12

Re-enable it by adding a wait on vm_vsrc before every barrier "start" instruction.
This is a less strong wait than what we do without BackOffBarrier, thus this shouldn't introduce
any new guarantees that can be abused, instead it relaxes the guarantees we have now to the bare
minimum needed to support the behavior users want (fence release + barrier works).

There is an exact memory model in the works which will be documented separately.



To unsubscribe from these emails, change your notification settings at https://github.com/llvm/llvm-project/settings/notifications