[all-commits] [llvm/llvm-project] b01fe4: [AMDGPU] Do not always add latency between LDSDMA ...

[Jeffrey Byrnes via All-commits]

  Branch: refs/heads/main
  Home:   https://github.com/llvm/llvm-project
  Commit: b01fe4e3d192a293caee9ec74ff4012958758d6b
      https://github.com/llvm/llvm-project/commit/b01fe4e3d192a293caee9ec74ff4012958758d6b
  Author: Jeffrey Byrnes <jeffrey.byrnes at amd.com>
  Date:   2026-06-08 (Mon, 08 Jun 2026)

  Changed paths:
    M llvm/lib/Target/AMDGPU/AMDGPUBarrierLatency.cpp
    M llvm/test/CodeGen/AMDGPU/ldsdmacnt_sched.mir

  Log Message:
  -----------
  [AMDGPU] Do not always add latency between LDSDMA -> S_WAIT_LDSDMA (#201942)

In loop bodies we typically see LDSDMA instructions prefetched an
iteration or more. Thus, we may have LDSDMA, followed by S_WAIT_LDSDMA
that is waiting on prior iteration LDSDMA. Currently, the scheduler
thinks there will be a long stall between this LDSDMA and S_WAIT_LDSDMA.

This adds some basic checking for LDSDMA and S_WAIT_LDSDMA in the same
region to avoid adding latency in cases where we are certain the
S_WAIT_LDSDMA does not correspond with the LDSDMA.



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