[all-commits] [llvm/llvm-project] 5eb1e6: [AMDGPU] Verify dominance when rewriting spills to...

[Austin Kerbow via All-commits]

  Branch: refs/heads/users/kerbowa/rewrite-agpr-copy-crash
  Home:   https://github.com/llvm/llvm-project
  Commit: 5eb1e66c1cb70721a9082e80a310a8211e8b036f
      https://github.com/llvm/llvm-project/commit/5eb1e66c1cb70721a9082e80a310a8211e8b036f
  Author: Austin Kerbow <Austin.Kerbow at amd.com>
  Date:   2025-11-30 (Sun, 30 Nov 2025)

  Changed paths:
    M llvm/lib/Target/AMDGPU/AMDGPURewriteAGPRCopyMFMA.cpp
    A llvm/lib/Target/AMDGPU/AMDGPURewriteAGPRCopyMFMA.h
    A llvm/unittests/Target/AMDGPU/AMDGPURewriteAGPRCopyMFMATest.cpp
    M llvm/unittests/Target/AMDGPU/CMakeLists.txt

  Log Message:
  -----------
  [AMDGPU] Verify dominance when rewriting spills to registers

Rev1: Updated condition to check for "joint domination", i.e. no reload
is reachable from entry without reaching a store to the same slot. Still
working on reduced test or unit test.

When performing spill elimination in the AGPR copy rewrite pass it was
possible to see spill reloads that were not jointly dominated by any
store. This caused invalid MIR to be generated where vreg uses were not
dominated by defs. This patch adds a joint dominance check before
rewriting spills.



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