[all-commits] [llvm/llvm-project] 9271a7: [AMDGPU] Mark `SI_WATERFALL_LOOP` as divergent in ...

[Igor Wodiany via All-commits]

  Branch: refs/heads/main
  Home:   https://github.com/llvm/llvm-project
  Commit: 9271a73e189508743b4b2ad4393f12d5a8132fa9
      https://github.com/llvm/llvm-project/commit/9271a73e189508743b4b2ad4393f12d5a8132fa9
  Author: Igor Wodiany <igor.wodiany at amd.com>
  Date:   2026-09-14 (Mon, 14 Sep 2026)

  Changed paths:
    M llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
    M llvm/test/CodeGen/AMDGPU/machine-sink-temporal-divergence-swdev407790.mir

  Log Message:
  -----------
  [AMDGPU] Mark `SI_WATERFALL_LOOP` as divergent in `hasDivergentBranch` (#218969)

When a vector instruction takes an SGPR source operand, sinking it out
of a loop with divergent exit would cause the vector instruction to only
get the value from the last iteration in the SGPR. However, in a
divergent loop the value of the SGPR operand may vary across iterations,
so the instruction cannot be sink out of the loop as only using the last
value is incorrect. Before this change `SI_LOOP` was marked as divergent
but `SI_WATERFALL_LOOP` was not, allowing such incorrect sinking for
waterfall loops.

Assisted-by: Claude Code



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