diff --git a/llvm/lib/CodeGen/LiveRangeEdit.cpp b/llvm/lib/CodeGen/LiveRangeEdit.cpp --- a/llvm/lib/CodeGen/LiveRangeEdit.cpp +++ b/llvm/lib/CodeGen/LiveRangeEdit.cpp @@ -134,9 +134,11 @@ return false; // Check that subrange is live at UseIdx. - if (MO.getSubReg()) { + if (li.hasSubRanges()) { const TargetRegisterInfo *TRI = MRI.getTargetRegisterInfo(); - LaneBitmask LM = TRI->getSubRegIndexLaneMask(MO.getSubReg()); + unsigned SubReg = MO.getSubReg(); + LaneBitmask LM = SubReg ? TRI->getSubRegIndexLaneMask(SubReg) + : MRI.getMaxLaneMaskForVReg(MO.getReg()); for (LiveInterval::SubRange &SR : li.subranges()) { if ((SR.LaneMask & LM).none()) continue; diff --git a/llvm/test/CodeGen/AMDGPU/remat-dead-subreg.mir b/llvm/test/CodeGen/AMDGPU/remat-dead-subreg.mir --- a/llvm/test/CodeGen/AMDGPU/remat-dead-subreg.mir +++ b/llvm/test/CodeGen/AMDGPU/remat-dead-subreg.mir @@ -79,3 +79,28 @@ %6:vreg_64 = V_MOV_B64_PSEUDO %2, implicit $exec S_NOP 0, implicit %1.sub0, implicit %1.sub3 ... +--- +name: dead_subreg_whole_reg +tracksRegLiveness: true +body: | + bb.0.entry: + ; GCN-LABEL: name: dead_subreg_whole_reg + ; GCN: $m0 = IMPLICIT_DEF + ; GCN-NEXT: renamable $sgpr0_sgpr1 = S_MOV_B64 1, implicit $m0 + ; GCN-NEXT: renamable $sgpr2_sgpr3 = S_MOV_B64 renamable $sgpr0_sgpr1 + ; GCN-NEXT: SI_SPILL_S64_SAVE killed renamable $sgpr2_sgpr3, %stack.0, implicit $exec, implicit $sp_reg :: (store (s64) into %stack.0, align 4, addrspace 5) + ; GCN-NEXT: renamable $sgpr4_sgpr5 = S_MOV_B64 2, implicit $m0 + ; GCN-NEXT: renamable $sgpr2_sgpr3 = S_MOV_B64 3, implicit $m0 + ; GCN-NEXT: dead %4:vgpr_32 = V_MOV_B32_e32 $sgpr0, implicit $exec, implicit killed $sgpr4_sgpr5, implicit killed $sgpr2_sgpr3 + ; GCN-NEXT: renamable $sgpr2_sgpr3 = SI_SPILL_S64_RESTORE %stack.0, implicit $exec, implicit $sp_reg :: (load (s64) from %stack.0, align 4, addrspace 5) + ; GCN-NEXT: dead %5:vreg_64 = V_MOV_B64_PSEUDO killed $sgpr2_sgpr3, implicit $exec + ; GCN-NEXT: S_NOP 0, implicit killed renamable $sgpr0 + $m0 = IMPLICIT_DEF + %0:sreg_64_xexec = S_MOV_B64 1, implicit $m0 + %1:sreg_64 = S_MOV_B64 %0:sreg_64_xexec + %2:sreg_64 = S_MOV_B64 2, implicit $m0 + %3:sreg_64 = S_MOV_B64 3, implicit $m0 + %4:vgpr_32 = V_MOV_B32_e32 %0.sub0:sreg_64_xexec, implicit $exec, implicit %2, implicit %3 + %5:vreg_64 = V_MOV_B64_PSEUDO %1, implicit $exec + S_NOP 0, implicit %0.sub0 +...