AMDGPU/SI: Remove REGISTER_STORE/REGISTER_LOAD code which is now dead

[arsenm]

LGTM

llvm/trunk/lib/Target/AMDGPU/AMDGPUISelDAGToDAG.cpp

case ISD::ConstantFP: {
     return CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, DL,
                                   N->getValueType(0), Ops);
   }
   case ISD::LOAD:
   case ISD::STORE: {
     N = glueCopyToM0(N);
     break;
   }
-  case AMDGPUISD::REGISTER_LOAD: {
-    if (Subtarget->getGeneration() <= AMDGPUSubtarget::NORTHERN_ISLANDS)
-      break;
-    SDValue Addr, Offset;
-
-    SDLoc DL(N);
-    SelectADDRIndirect(N->getOperand(1), Addr, Offset);
-    const SDValue Ops[] = {
-      Addr,
-      Offset,
-      CurDAG->getTargetConstant(0, DL, MVT::i32),
-      N->getOperand(0),
-    };
-    return CurDAG->getMachineNode(AMDGPU::SI_RegisterLoad, DL,
-                                  CurDAG->getVTList(MVT::i32, MVT::i64,
-                                                    MVT::Other),
-                                  Ops);
-  }
-  case AMDGPUISD::REGISTER_STORE: {
-    if (Subtarget->getGeneration() <= AMDGPUSubtarget::NORTHERN_ISLANDS)
-      break;
-    SDValue Addr, Offset;
-    SelectADDRIndirect(N->getOperand(2), Addr, Offset);
-    SDLoc DL(N);
-    const SDValue Ops[] = {
-      N->getOperand(1),
-      Addr,
-      Offset,
-      CurDAG->getTargetConstant(0, DL, MVT::i32),
-      N->getOperand(0),
-    };
-    return CurDAG->getMachineNode(AMDGPU::SI_RegisterStorePseudo, DL,
-                                        CurDAG->getVTList(MVT::Other),
-                                        Ops);
-  }
 
   case AMDGPUISD::BFE_I32:
   case AMDGPUISD::BFE_U32: {
     if (Subtarget->getGeneration() < AMDGPUSubtarget::SOUTHERN_ISLANDS)
       break;
 
     // There is a scalar version available, but unlike the vector version which
     // has a separate operand for the offset and width, the scalar version packs

llvm/trunk/lib/Target/AMDGPU/SIISelLowering.cpp

if (Chains.empty())
     return Chain;
 
   return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Chains);
 }
 
 MachineBasicBlock * SITargetLowering::EmitInstrWithCustomInserter(
     MachineInstr * MI, MachineBasicBlock * BB) const {
 
-  MachineBasicBlock::iterator I = *MI;
-  const SIInstrInfo *TII =
-      static_cast<const SIInstrInfo *>(Subtarget->getInstrInfo());
-
   switch (MI->getOpcode()) {
   default:
     return AMDGPUTargetLowering::EmitInstrWithCustomInserter(MI, BB);
   case AMDGPU::BRANCH:
     return BB;
-  case AMDGPU::SI_RegisterStorePseudo: {
-    MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
-    unsigned Reg = MRI.createVirtualRegister(&AMDGPU::SReg_64RegClass);
-    MachineInstrBuilder MIB =
-        BuildMI(*BB, I, MI->getDebugLoc(), TII->get(AMDGPU::SI_RegisterStore),
-                Reg);
-    for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i)
-      MIB.addOperand(MI->getOperand(i));
-
-    MI->eraseFromParent();
-    break;
-  }
   }
   return BB;
 }
 
 bool SITargetLowering::enableAggressiveFMAFusion(EVT VT) const {
   // This currently forces unfolding various combinations of fsub into fma with
   // free fneg'd operands. As long as we have fast FMA (controlled by
   // isFMAFasterThanFMulAndFAdd), we should perform these.

llvm/trunk/lib/Target/AMDGPU/SIInstructions.td

def SI_KILL : InstSI <
   [(int_AMDGPU_kill f32:$src)]
 >;
 } // End Uses = [EXEC], Defs = [EXEC,VCC]
 
 } // end mayLoad = 1, mayStore = 1, hasSideEffects = 1
 
 let Uses = [EXEC], Defs = [EXEC,VCC,M0] in {
 
-//defm SI_ : RegisterLoadStore <VGPR_32, FRAMEri, ADDRIndirect>;
-
-let UseNamedOperandTable = 1 in {
-
-def SI_RegisterLoad : InstSI <
-  (outs VGPR_32:$dst, SReg_64:$temp),
-  (ins FRAMEri32:$addr, i32imm:$chan),
-  "", []
-> {
-  let isRegisterLoad = 1;
-  let mayLoad = 1;
-}
-
-class SIRegStore<dag outs> : InstSI <
-  outs,
-  (ins VGPR_32:$val, FRAMEri32:$addr, i32imm:$chan),
-  "", []
-> {
-  let isRegisterStore = 1;
-  let mayStore = 1;
-}
-
-let usesCustomInserter = 1 in {
-def SI_RegisterStorePseudo : SIRegStore<(outs)>;
-} // End usesCustomInserter = 1
-def SI_RegisterStore : SIRegStore<(outs SReg_64:$temp)>;
-
-
-} // End UseNamedOperandTable = 1
-
 class SI_INDIRECT_SRC<RegisterClass rc> : InstSI <
   (outs VGPR_32:$dst, SReg_64:$temp),
   (ins rc:$src, VSrc_32:$idx, i32imm:$off),
   "si_indirect_src $dst, $temp, $src, $idx, $off",
   []
 >;
 
 class SI_INDIRECT_DST<RegisterClass rc> : InstSI <