[all-commits] [llvm/llvm-project] 807de2: [AMDGPU] Stop iDot4 chain walker at non-ADD nodes ...

[Arseniy Obolenskiy via All-commits]

  Branch: refs/heads/main
  Home:   https://github.com/llvm/llvm-project
  Commit: 807de2dc44794205b336ed8b2ca4731bc3ab865b
      https://github.com/llvm/llvm-project/commit/807de2dc44794205b336ed8b2ca4731bc3ab865b
  Author: Arseniy Obolenskiy <arseniy.obolenskiy at amd.com>
  Date:   2026-07-28 (Tue, 28 Jul 2026)

  Changed paths:
    M llvm/lib/Target/AMDGPU/SIISelLowering.cpp
    M llvm/test/CodeGen/AMDGPU/idot4-test.ll
    M llvm/test/CodeGen/AMDGPU/no-corresponding-integer-type.ll

  Log Message:
  -----------
  [AMDGPU] Stop iDot4 chain walker at non-ADD nodes (#198412)

The loop body in performAddCombine dot4 matcher unconditionally treats
TempNode operands as the next link's addends, so the chain only works
when TempNode is an `ISD::ADD`

The old getNumOperands() guard let AND/OR/XOR/etc. through, leaking
their non-addend operands (e.g. a constant mask) into the dot4
accumulator and miscompiling kernels



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