[all-commits] [llvm/llvm-project] 4bd597: [AMDGPU] Work around gfx1250 VOPD3 FP64 OPY hazard...

[Joe Nash via All-commits]

  Branch: refs/heads/main
  Home:   https://github.com/llvm/llvm-project
  Commit: 4bd5979264ba739f8cb266b90c40e519eb7f3008
      https://github.com/llvm/llvm-project/commit/4bd5979264ba739f8cb266b90c40e519eb7f3008
  Author: Joe Nash <joseph.nash at amd.com>
  Date:   2026-09-24 (Thu, 24 Sep 2026)

  Changed paths:
    M llvm/lib/Target/AMDGPU/AMDGPU.td
    M llvm/lib/Target/AMDGPU/GCNVOPDUtils.cpp
    A llvm/test/CodeGen/AMDGPU/vopd3-f64-opy-last-vgpr-gfx1250.mir

  Log Message:
  -----------
  [AMDGPU] Work around gfx1250 VOPD3 FP64 OPY hazard (#224741)

In a VOPD3 instruction whose OPX is a 64-bit operation, an OPY VGPR
source operand reads back the wrong value when it is the last VGPR the
wave owns.

A wave always owns a whole number of VGPR allocation granules, 16
registers each on gfx1250, so the last VGPR it owns is always the one
just below a granule boundary. Reject a pair whose OPY reads such a
register in checkVOPDRegConstraints, leaving the two components as
separate instructions.
VOPD is only an optimization, so losing a pairing is always safe.

The workaround is gated on a new vopd3-f64-opy-src-hazard subtarget
feature, enabled for gfx1250, gfx1251, gfx1250-strict.

Assisted-by: Claude Code:claude-opus-5



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