Index: llvm/lib/CodeGen/LiveRangeEdit.cpp =================================================================== --- llvm/lib/CodeGen/LiveRangeEdit.cpp +++ llvm/lib/CodeGen/LiveRangeEdit.cpp @@ -133,6 +133,22 @@ if (OVNI != li.getVNInfoAt(UseIdx)) return false; + + // Check that subrange is live at UseIdx. + if (MO.getSubReg()) { + const TargetRegisterInfo *TRI = MRI.getTargetRegisterInfo(); + LaneBitmask LM = TRI->getSubRegIndexLaneMask(MO.getSubReg()); + for (LiveInterval::SubRange &SR : li.subranges()) { + if ((SR.LaneMask & LM).none()) + continue; + if (!SR.liveAt(UseIdx)) + return false; + // Early exit if all used lanes are checked. No need to continue. + LM &= ~SR.LaneMask; + if (LM.none()) + break; + } + } } return true; } Index: llvm/test/CodeGen/AMDGPU/remat-dead-subreg.mir =================================================================== --- /dev/null +++ llvm/test/CodeGen/AMDGPU/remat-dead-subreg.mir @@ -0,0 +1,81 @@ +# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py +# RUN: llc -march=amdgcn -start-before=greedy -stop-after=virtregrewriter -stress-regalloc=3 -verify-machineinstrs -o - %s | FileCheck -check-prefix=GCN %s + +# Check that subreg use is live at the point of materialization, not just the main range. +# Do not rematerialize if used subreg is dead at a new index. + +--- +name: dead_subreg +tracksRegLiveness: true +body: | + bb.0.entry: + ; GCN-LABEL: name: dead_subreg + ; GCN: $m0 = IMPLICIT_DEF + ; GCN-NEXT: renamable $sgpr0_sgpr1 = S_MOV_B64 1, implicit $m0 + ; GCN-NEXT: renamable $sgpr1 = S_MUL_I32 renamable $sgpr1, 3 + ; GCN-NEXT: SI_SPILL_S32_SAVE killed renamable $sgpr1, %stack.0, implicit $exec, implicit $sp_reg :: (store (s32) into %stack.0, addrspace 5) + ; GCN-NEXT: renamable $sgpr2 = S_MOV_B32 2, implicit $m0 + ; GCN-NEXT: renamable $sgpr1 = S_MOV_B32 3, implicit $m0 + ; GCN-NEXT: dead %4:vgpr_32 = V_MOV_B32_e32 $sgpr0, implicit $exec, implicit killed $sgpr2, implicit killed $sgpr1 + ; GCN-NEXT: renamable $sgpr1 = SI_SPILL_S32_RESTORE %stack.0, implicit $exec, implicit $sp_reg :: (load (s32) from %stack.0, addrspace 5) + ; GCN-NEXT: dead %5:vgpr_32 = V_MOV_B32_e32 killed $sgpr1, 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_32 = S_MUL_I32 %0.sub1:sreg_64_xexec, 3 + %2:sreg_32 = S_MOV_B32 2, implicit $m0 + %3:sreg_32 = S_MOV_B32 3, implicit $m0 + %4:vgpr_32 = V_MOV_B32_e32 %0.sub0:sreg_64_xexec, implicit $exec, implicit %2, implicit %3 + %5:vgpr_32 = V_MOV_B32_e32 %1:sreg_32, implicit $exec + S_NOP 0, implicit %0.sub0 +... +--- +name: live_subreg +tracksRegLiveness: true +body: | + bb.0.entry: + ; GCN-LABEL: name: live_subreg + ; GCN: $m0 = IMPLICIT_DEF + ; GCN-NEXT: renamable $sgpr4_sgpr5 = S_MOV_B64 1, implicit $m0 + ; GCN-NEXT: renamable $sgpr2 = S_MOV_B32 2, implicit $m0 + ; GCN-NEXT: SI_SPILL_S32_SAVE killed renamable $sgpr2, %stack.0, implicit $exec, implicit $sp_reg :: (store (s32) into %stack.0, addrspace 5) + ; GCN-NEXT: renamable $sgpr2 = S_MOV_B32 3, implicit $m0 + ; GCN-NEXT: renamable $sgpr0 = SI_SPILL_S32_RESTORE %stack.0, implicit $exec, implicit $sp_reg :: (load (s32) from %stack.0, addrspace 5) + ; GCN-NEXT: dead %4:vgpr_32 = V_MOV_B32_e32 $sgpr4, implicit $exec, implicit killed $sgpr0, implicit killed $sgpr2 + ; GCN-NEXT: renamable $sgpr0 = S_MUL_I32 renamable $sgpr5, 3 + ; GCN-NEXT: dead %5:vgpr_32 = V_MOV_B32_e32 killed $sgpr0, implicit $exec + ; GCN-NEXT: S_NOP 0, implicit killed renamable $sgpr5 + $m0 = IMPLICIT_DEF + %0:sreg_64_xexec = S_MOV_B64 1, implicit $m0 + %1:sreg_32 = S_MUL_I32 %0.sub1:sreg_64_xexec, 3 + %2:sreg_32 = S_MOV_B32 2, implicit $m0 + %3:sreg_32 = S_MOV_B32 3, implicit $m0 + %4:vgpr_32 = V_MOV_B32_e32 %0.sub0:sreg_64_xexec, implicit $exec, implicit %2, implicit %3 + %5:vgpr_32 = V_MOV_B32_e32 %1:sreg_32, implicit $exec + S_NOP 0, implicit %0.sub1 +... +--- +name: partially_live_wide_subreg +tracksRegLiveness: true +body: | + bb.0.entry: + ; GCN-LABEL: name: partially_live_wide_subreg + ; GCN: renamable $sgpr0_sgpr1 = IMPLICIT_DEF + ; GCN-NEXT: renamable $sgpr8_sgpr9_sgpr10_sgpr11 = S_LOAD_DWORDX4_IMM killed renamable $sgpr0_sgpr1, 1, 0 + ; GCN-NEXT: renamable $sgpr4_sgpr5 = S_MOV_B64 renamable $sgpr10_sgpr11 + ; GCN-NEXT: SI_SPILL_S64_SAVE killed renamable $sgpr4_sgpr5, %stack.0, implicit $exec, implicit $sp_reg :: (store (s64) into %stack.0, align 4, addrspace 5) + ; GCN-NEXT: renamable $sgpr4_sgpr5 = S_MOV_B64 3, implicit $m0 + ; GCN-NEXT: renamable $sgpr0_sgpr1 = S_MOV_B64 2, implicit $m0 + ; GCN-NEXT: dead %5:vgpr_32 = V_MOV_B32_e32 $sgpr8, implicit $exec, implicit killed $sgpr0_sgpr1, implicit killed $sgpr4_sgpr5 + ; GCN-NEXT: renamable $sgpr4_sgpr5 = SI_SPILL_S64_RESTORE %stack.0, implicit $exec, implicit $sp_reg :: (load (s64) from %stack.0, align 4, addrspace 5) + ; GCN-NEXT: dead %6:vreg_64 = V_MOV_B64_PSEUDO killed $sgpr4_sgpr5, implicit $exec + ; GCN-NEXT: S_NOP 0, implicit killed renamable $sgpr8, implicit renamable $sgpr11 + %0:sreg_64 = IMPLICIT_DEF + %1:sgpr_128 = S_LOAD_DWORDX4_IMM %0, 1, 0 + %2:sreg_64 = S_MOV_B64 %1.sub2_sub3 + %3:sreg_64 = S_MOV_B64 2, implicit $m0 + %4:sreg_64 = S_MOV_B64 3, implicit $m0 + %5:vgpr_32 = V_MOV_B32_e32 %1.sub0, implicit $exec, implicit %3, implicit %4 + %6:vreg_64 = V_MOV_B64_PSEUDO %2, implicit $exec + S_NOP 0, implicit %1.sub0, implicit %1.sub3 +...