[all-commits] [llvm/llvm-project] 5d8670: [CodeGen] Use validated wave counts for AMDGPU spi...

[Yaxun (Sam) Liu via All-commits]

  Branch: refs/heads/users/yxsamliu/pgo-wave-profiled-spill
  Home:   https://github.com/llvm/llvm-project
  Commit: 5d8670267abd25cf2e4e043014cc55b11b039131
      https://github.com/llvm/llvm-project/commit/5d8670267abd25cf2e4e043014cc55b11b039131
  Author: Yaxun (Sam) Liu <yaxun.liu at amd.com>
  Date:   2026-09-28 (Mon, 28 Sep 2026)

  Changed paths:
    M llvm/lib/CodeGen/SpillPlacement.cpp
    A llvm/test/CodeGen/AMDGPU/wave-profile-spill.mir

  Log Message:
  -----------
  [CodeGen] Use validated wave counts for AMDGPU spill costs

Lane-based block frequencies can understate the cost of a spill in
divergent GPU code: a wave still executes a block with only some lanes
active. Use measured block-wave counts to weight SpillPlacement's costs
relative to the original entry-wave count.

Opt in on AMDGPU only. Accept a measured count only when its IR block
maps uniquely to a machine block with matching predecessors and
successors. Keep the existing MBFI cost for unmeasured or rejected
blocks, and leave branch probabilities and general BFI unchanged.


  Commit: 94b07b7bb1eabc36378a198d607443210acf50b7
      https://github.com/llvm/llvm-project/commit/94b07b7bb1eabc36378a198d607443210acf50b7
  Author: Yaxun (Sam) Liu <yaxun.liu at amd.com>
  Date:   2026-09-28 (Mon, 28 Sep 2026)

  Changed paths:
    M llvm/lib/CodeGen/SpillPlacement.cpp
    A llvm/test/CodeGen/AMDGPU/wave-profile-spill-zero-cost.mir

  Log Message:
  -----------
  [CodeGen] Keep wave-profiled spill frequencies positive

SpillPlacement expects positive block weights, but a valid wave profile can
record zero executions for a CFG-reachable block. Giving such a block zero
spill cost can make the allocator choose a very different placement.

Clamp every accepted wave-derived frequency to at least one, as we already
do for nonzero counts that round down to zero. Unmeasured or rejected blocks
still use their existing MBFI frequency. Add a focused MIR test for a valid
zero-wave record.

This pattern arose in a profiled Composable Kernel convolution case. With
the separate spill correctness fixes and partial spilling enabled, the
zero-cost policy failed two CPU-reference checks; the positive floor passed
both. The test checks the cost directly; the application result was checked
separately on gfx950.


Compare: https://github.com/llvm/llvm-project/compare/7ff1282de0ce...94b07b7bb1ea

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