[all-commits] [llvm/llvm-project] 1dd968: [AMDGPU] Fix fmul/fma legacy sign-of-zero miscompi...

[Wooseok Lee via All-commits]

  Branch: refs/heads/main
  Home:   https://github.com/llvm/llvm-project
  Commit: 1dd9687d41e1331996ece0839e58673ab4d8e10f
      https://github.com/llvm/llvm-project/commit/1dd9687d41e1331996ece0839e58673ab4d8e10f
  Author: Wooseok Lee <wolee at amd.com>
  Date:   2026-07-30 (Thu, 30 Jul 2026)

  Changed paths:
    M llvm/lib/Target/AMDGPU/AMDGPUInstCombineIntrinsic.cpp
    M llvm/test/Transforms/InstCombine/AMDGPU/fma_legacy.ll
    M llvm/test/Transforms/InstCombine/AMDGPU/fmul_legacy.ll

  Log Message:
  -----------
  [AMDGPU] Fix fmul/fma legacy sign-of-zero miscompile (#203567)

V_MUL_LEGACY_F32 always returns +0.0 when either operand is ±0.0 or a
denormal treated as zero, while IEEE fmul XORs operand signs. This makes
results observably different when one operand is a runtime ±0.0:

  fmul.legacy(-2.0, +0.0) = +0.0  (legacy zero clause forces +0.0)
  fmul       (-2.0, +0.0) = -0.0  (IEEE sign XOR: - * + = -)

The previous code folded whenever one operand matched m_FiniteNonZero()
or both were known never Inf/NaN, without accounting for sign-of-zero
divergence on the other operand.

Restructure the guard:

- Always safe: both operands are known never zero (legacy zero clause
  cannot fire). Uses isKnownNeverLogicalZero to correctly account for
  denormals flushed to zero under the function's f32 denormal mode.

- With nsz (sign-of-zero is don't-care), two additional cases apply:
  1. One operand is not zero or infinity or NaN: zero clause cannot fire
     on that side, and 0*Inf/0*NaN (where legacy returns +0.0 but IEEE
     returns NaN) is excluded.
  2. Neither operand is infinity or NaN: 0*Inf and 0*NaN cases excluded,
     nsz covers any sign-of-zero difference from the zero clause.
  A single condition captures both cases.

- Compute Known1 lazily: bail before the second computeKnownFPClass call
  if Op0 may be zero and nsz is not set, since Op1 cannot help.

Update fmul_legacy.ll and fma_legacy.ll to reflect the corrected
behavior: cases that previously folded without nsz now require it or
require both operands to be provably non-zero.



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