[mips] Optimize shifts for types larger than GPR size (mips2/mips3)

[abeserminji]

Summary

With this patch, shifts are lowered to optimal number of instructions
necessary to shift types larger than the general purpose register size.

[abeserminji]

Update patch for context.

[atanasyan]

LGTM with a couple of nits

lib/Target/Mips/MipsISelLowering.cpp

2443 ↗

(On Diff #180828)

clang-format the if block please

4380 ↗

(On Diff #180828)

s/instrution/instruction/

[abeserminji]

Comments resolved.

llvm/trunk/lib/Target/Mips/MipsCondMov.td

 def PseudoSELECTFP_T_D32 : SelectFP_Pseudo_T<AFGR64Opnd>, FGR_32;
 def PseudoSELECTFP_T_D64 : SelectFP_Pseudo_T<FGR64Opnd>, FGR_64;
 
 def PseudoSELECTFP_F_I : SelectFP_Pseudo_F<GPR32Opnd>;
 def PseudoSELECTFP_F_I64 : SelectFP_Pseudo_F<GPR64Opnd>;
 def PseudoSELECTFP_F_S : SelectFP_Pseudo_F<FGR32Opnd>;
 def PseudoSELECTFP_F_D32 : SelectFP_Pseudo_F<AFGR64Opnd>, FGR_32;
 def PseudoSELECTFP_F_D64 : SelectFP_Pseudo_F<FGR64Opnd>, FGR_64;
+
+let usesCustomInserter = 1 in {
+class D_SELECT_CLASS<RegisterOperand RC> :
+  PseudoSE<(outs RC:$dst1, RC:$dst2),
+           (ins GPR32Opnd:$cond, RC:$a1, RC:$a2, RC:$b1, RC:$b2), []>,
+  ISA_MIPS1_NOT_4_32;
+}
+
+def PseudoD_SELECT_I   : D_SELECT_CLASS<GPR32Opnd>;
+def PseudoD_SELECT_I64 : D_SELECT_CLASS<GPR64Opnd>;

llvm/trunk/lib/Target/Mips/MipsISelLowering.h

private:
     MachineBasicBlock *emitAtomicCmpSwap(MachineInstr &MI,
                                          MachineBasicBlock *BB) const;
     MachineBasicBlock *emitAtomicCmpSwapPartword(MachineInstr &MI,
                                                  MachineBasicBlock *BB,
                                                  unsigned Size) const;
     MachineBasicBlock *emitSEL_D(MachineInstr &MI, MachineBasicBlock *BB) const;
     MachineBasicBlock *emitPseudoSELECT(MachineInstr &MI, MachineBasicBlock *BB,
                                         bool isFPCmp, unsigned Opc) const;
+    MachineBasicBlock *emitPseudoD_SELECT(MachineInstr &MI,
+                                          MachineBasicBlock *BB) const;
   };
 
   /// Create MipsTargetLowering objects.
   const MipsTargetLowering *
   createMips16TargetLowering(const MipsTargetMachine &TM,
                              const MipsSubtarget &STI);
   const MipsTargetLowering *
   createMipsSETargetLowering(const MipsTargetMachine &TM,

llvm/trunk/lib/Target/Mips/MipsISelLowering.cpp

MipsTargetLowering::EmitInstrWithCustomInserter(MachineInstr &MI,
   case Mips::PseudoSELECTFP_F_D64:
     return emitPseudoSELECT(MI, BB, true, Mips::BC1F);
   case Mips::PseudoSELECTFP_T_I:
   case Mips::PseudoSELECTFP_T_I64:
   case Mips::PseudoSELECTFP_T_S:
   case Mips::PseudoSELECTFP_T_D32:
   case Mips::PseudoSELECTFP_T_D64:
     return emitPseudoSELECT(MI, BB, true, Mips::BC1T);
+  case Mips::PseudoD_SELECT_I:
+  case Mips::PseudoD_SELECT_I64:
+    return emitPseudoD_SELECT(MI, BB);
   }
 }
 
 // This function also handles Mips::ATOMIC_SWAP_I32 (when BinOpcode == 0), and
 // Mips::ATOMIC_LOAD_NAND_I32 (when Nand == true)
 MachineBasicBlock *
 MipsTargetLowering::emitAtomicBinary(MachineInstr &MI,
                                      MachineBasicBlock *BB) const {
SDValue MipsTargetLowering::lowerShiftRightParts(SDValue Op, SelectionDAG &DAG,
   SDValue ShiftRightLo = DAG.getNode(ISD::SRL, DL, VT, Lo, Shamt);
   SDValue Or = DAG.getNode(ISD::OR, DL, VT, ShiftLeftHi, ShiftRightLo);
   SDValue ShiftRightHi = DAG.getNode(IsSRA ? ISD::SRA : ISD::SRL,
                                      DL, VT, Hi, Shamt);
   SDValue Cond = DAG.getNode(ISD::AND, DL, MVT::i32, Shamt,
                              DAG.getConstant(VT.getSizeInBits(), DL, MVT::i32));
   SDValue Ext = DAG.getNode(ISD::SRA, DL, VT, Hi,
                             DAG.getConstant(VT.getSizeInBits() - 1, DL, VT));
+
+  if (!(Subtarget.hasMips4() || Subtarget.hasMips32())) {
+    SDVTList VTList = DAG.getVTList(VT, VT);
+    return DAG.getNode(Subtarget.isGP64bit() ? Mips::PseudoD_SELECT_I64
+                                             : Mips::PseudoD_SELECT_I,
+                       DL, VTList, Cond, ShiftRightHi,
+                       IsSRA ? Ext : DAG.getConstant(0, DL, VT), Or,
+                       ShiftRightHi);
+  }
+
   Lo = DAG.getNode(ISD::SELECT, DL, VT, Cond, ShiftRightHi, Or);
   Hi = DAG.getNode(ISD::SELECT, DL, VT, Cond,
                    IsSRA ? Ext : DAG.getConstant(0, DL, VT), ShiftRightHi);
 
   SDValue Ops[2] = {Lo, Hi};
   return DAG.getMergeValues(Ops, DL);
 }
 
BuildMI(*BB, BB->begin(), DL, TII->get(Mips::PHI), MI.getOperand(0).getReg())
       .addReg(MI.getOperand(3).getReg())
       .addMBB(copy0MBB);
 
   MI.eraseFromParent(); // The pseudo instruction is gone now.
 
   return BB;
 }
 
+MachineBasicBlock *MipsTargetLowering::emitPseudoD_SELECT(MachineInstr &MI,
+                                                          MachineBasicBlock *BB) const {
+  assert(!(Subtarget.hasMips4() || Subtarget.hasMips32()) &&
+         "Subtarget already supports SELECT nodes with the use of"
+         "conditional-move instructions.");
+
+  const TargetInstrInfo *TII = Subtarget.getInstrInfo();
+  DebugLoc DL = MI.getDebugLoc();
+
+  // D_SELECT substitutes two SELECT nodes that goes one after another and
+  // have the same condition operand. On machines which don't have
+  // conditional-move instruction, it reduces unnecessary branch instructions
+  // which are result of using two diamond patterns that are result of two
+  // SELECT pseudo instructions.
+  const BasicBlock *LLVM_BB = BB->getBasicBlock();
+  MachineFunction::iterator It = ++BB->getIterator();
+
+  //  thisMBB:
+  //  ...
+  //   TrueVal = ...
+  //   setcc r1, r2, r3
+  //   bNE   r1, r0, copy1MBB
+  //   fallthrough --> copy0MBB
+  MachineBasicBlock *thisMBB = BB;
+  MachineFunction *F = BB->getParent();
+  MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
+  MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
+  F->insert(It, copy0MBB);
+  F->insert(It, sinkMBB);
+
+  // Transfer the remainder of BB and its successor edges to sinkMBB.
+  sinkMBB->splice(sinkMBB->begin(), BB,
+                  std::next(MachineBasicBlock::iterator(MI)), BB->end());
+  sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
+
+  // Next, add the true and fallthrough blocks as its successors.
+  BB->addSuccessor(copy0MBB);
+  BB->addSuccessor(sinkMBB);
+
+  // bne rs, $0, sinkMBB
+  BuildMI(BB, DL, TII->get(Mips::BNE))
+      .addReg(MI.getOperand(2).getReg())
+      .addReg(Mips::ZERO)
+      .addMBB(sinkMBB);
+
+  //  copy0MBB:
+  //   %FalseValue = ...
+  //   # fallthrough to sinkMBB
+  BB = copy0MBB;
+
+  // Update machine-CFG edges
+  BB->addSuccessor(sinkMBB);
+
+  //  sinkMBB:
+  //   %Result = phi [ %TrueValue, thisMBB ], [ %FalseValue, copy0MBB ]
+  //  ...
+  BB = sinkMBB;
+
+  // Use two PHI nodes to select two reults
+  BuildMI(*BB, BB->begin(), DL, TII->get(Mips::PHI), MI.getOperand(0).getReg())
+      .addReg(MI.getOperand(3).getReg())
+      .addMBB(thisMBB)
+      .addReg(MI.getOperand(5).getReg())
+      .addMBB(copy0MBB);
+  BuildMI(*BB, BB->begin(), DL, TII->get(Mips::PHI), MI.getOperand(1).getReg())
+      .addReg(MI.getOperand(4).getReg())
+      .addMBB(thisMBB)
+      .addReg(MI.getOperand(6).getReg())
+      .addMBB(copy0MBB);
+
+  MI.eraseFromParent(); // The pseudo instruction is gone now.
+
+  return BB;
+}
+
 // FIXME? Maybe this could be a TableGen attribute on some registers and
 // this table could be generated automatically from RegInfo.
 unsigned MipsTargetLowering::getRegisterByName(const char* RegName, EVT VT,
                                                SelectionDAG &DAG) const {
   // Named registers is expected to be fairly rare. For now, just support $28
   // since the linux kernel uses it.
   if (Subtarget.isGP64bit()) {
     unsigned Reg = StringSwitch<unsigned>(RegName)

llvm/trunk/lib/Target/Mips/MipsSEISelDAGToDAG.cpp

bool MipsSEDAGToDAGISel::trySelect(SDNode *Node) {
 
   ///
   // Instruction Selection not handled by the auto-generated
   // tablegen selection should be handled here.
   ///
   switch(Opcode) {
   default: break;
 
+  case Mips::PseudoD_SELECT_I:
+  case Mips::PseudoD_SELECT_I64: {
+    MVT VT = Subtarget->isGP64bit() ? MVT::i64 : MVT::i32;
+    SDValue cond = Node->getOperand(0);
+    SDValue Hi1 = Node->getOperand(1);
+    SDValue Lo1 = Node->getOperand(2);
+    SDValue Hi2 = Node->getOperand(3);
+    SDValue Lo2 = Node->getOperand(4);
+
+    SDValue ops[] = {cond, Hi1, Lo1, Hi2, Lo2};
+    EVT NodeTys[] = {VT, VT};
+    ReplaceNode(Node, CurDAG->getMachineNode(Subtarget->isGP64bit()
+                                                 ? Mips::PseudoD_SELECT_I64
+                                                 : Mips::PseudoD_SELECT_I,
+                                             DL, NodeTys, ops));
+    return true;
+  }
+
   case ISD::ADDE: {
     selectAddE(Node, DL);
     return true;
   }
 
   case ISD::ConstantFP: {
     ConstantFPSDNode *CN = dyn_cast<ConstantFPSDNode>(Node);
     if (Node->getValueType(0) == MVT::f64 && CN->isExactlyValue(+0.0)) {

llvm/trunk/test/CodeGen/Mips/llvm-ir/ashr.ll

 ; MMR6-NEXT:    jrc $ra
 entry:
   %r = ashr i32 %a, %b
   ret i32 %r
 }
 
 define signext i64 @ashr_i64(i64 signext %a, i64 signext %b) {
 ; MIPS-LABEL: ashr_i64:
-; MIPS:       # %bb.0: # %entry
-; MIPS-NEXT:    srav $2, $4, $7
-; MIPS-NEXT:    andi $6, $7, 32
-; MIPS-NEXT:    beqz $6, $BB4_3
-; MIPS-NEXT:    move $3, $2
-; MIPS-NEXT:  # %bb.1: # %entry
-; MIPS-NEXT:    bnez $6, $BB4_4
-; MIPS-NEXT:    nop
-; MIPS-NEXT:  $BB4_2: # %entry
-; MIPS-NEXT:    jr $ra
-; MIPS-NEXT:    nop
-; MIPS-NEXT:  $BB4_3: # %entry
-; MIPS-NEXT:    srlv $1, $5, $7
-; MIPS-NEXT:    not $3, $7
-; MIPS-NEXT:    sll $5, $4, 1
-; MIPS-NEXT:    sllv $3, $5, $3
-; MIPS-NEXT:    beqz $6, $BB4_2
-; MIPS-NEXT:    or $3, $3, $1
-; MIPS-NEXT:  $BB4_4:
-; MIPS-NEXT:    jr $ra
+; MIPS:       # %bb.0:
+; MIPS-NEXT:    andi  $1, $7, 32
+; MIPS-NEXT:    bnez  $1, $BB4_2
+; MIPS-NEXT:    srav  $3, $4, $7
+; MIPS-NEXT:  # %bb.1:
+; MIPS-NEXT:    srlv  $1, $5, $7
+; MIPS-NEXT:    not $2, $7
+; MIPS-NEXT:    sll $4, $4, 1
+; MIPS-NEXT:    sllv  $2, $4, $2
+; MIPS-NEXT:    or  $1, $2, $1
+; MIPS-NEXT:    move  $2, $3
+; MIPS-NEXT:    jr  $ra
+; MIPS-NEXT:    move  $3, $1
+; MIPS-NEXT:  $BB4_2:
+; MIPS-NEXT:    jr  $ra
 ; MIPS-NEXT:    sra $2, $4, 31
 ;
 ; MIPS32-LABEL: ashr_i64:
 ; MIPS32:       # %bb.0: # %entry
 ; MIPS32-NEXT:    srlv $1, $5, $7
 ; MIPS32-NEXT:    not $2, $7
 ; MIPS32-NEXT:    sll $3, $4, 1
 ; MIPS32-NEXT:    sllv $2, $3, $2

 ; MMR6-NEXT:    jrc $ra
 entry:
   %r = ashr i64 %a, %b
   ret i64 %r
 }
 
 define signext i128 @ashr_i128(i128 signext %a, i128 signext %b) {
 ; MIPS-LABEL: ashr_i128:
-; MIPS:       # %bb.0: # %entry
-; MIPS-NEXT:    addiu $sp, $sp, -8
-; MIPS-NEXT:    .cfi_def_cfa_offset 8
-; MIPS-NEXT:    sw $17, 4($sp) # 4-byte Folded Spill
-; MIPS-NEXT:    sw $16, 0($sp) # 4-byte Folded Spill
-; MIPS-NEXT:    .cfi_offset 17, -4
-; MIPS-NEXT:    .cfi_offset 16, -8
-; MIPS-NEXT:    lw $25, 36($sp)
+; MIPS:       # %bb.0:
+; MIPS-NEXT:    lw  $2, 28($sp)
 ; MIPS-NEXT:    addiu $1, $zero, 64
-; MIPS-NEXT:    subu $11, $1, $25
-; MIPS-NEXT:    sllv $9, $5, $11
-; MIPS-NEXT:    andi $13, $11, 32
-; MIPS-NEXT:    addiu $2, $zero, 0
-; MIPS-NEXT:    bnez $13, $BB5_2
-; MIPS-NEXT:    addiu $3, $zero, 0
-; MIPS-NEXT:  # %bb.1: # %entry
-; MIPS-NEXT:    move $3, $9
-; MIPS-NEXT:  $BB5_2: # %entry
-; MIPS-NEXT:    not $gp, $25
-; MIPS-NEXT:    srlv $12, $6, $25
-; MIPS-NEXT:    andi $8, $25, 32
-; MIPS-NEXT:    bnez $8, $BB5_4
-; MIPS-NEXT:    move $15, $12
-; MIPS-NEXT:  # %bb.3: # %entry
-; MIPS-NEXT:    srlv $1, $7, $25
-; MIPS-NEXT:    sll $10, $6, 1
-; MIPS-NEXT:    sllv $10, $10, $gp
-; MIPS-NEXT:    or $15, $10, $1
-; MIPS-NEXT:  $BB5_4: # %entry
-; MIPS-NEXT:    addiu $10, $25, -64
-; MIPS-NEXT:    sll $17, $4, 1
-; MIPS-NEXT:    srav $14, $4, $10
-; MIPS-NEXT:    andi $24, $10, 32
-; MIPS-NEXT:    bnez $24, $BB5_6
-; MIPS-NEXT:    move $16, $14
-; MIPS-NEXT:  # %bb.5: # %entry
-; MIPS-NEXT:    srlv $1, $5, $10
-; MIPS-NEXT:    not $10, $10
-; MIPS-NEXT:    sllv $10, $17, $10
-; MIPS-NEXT:    or $16, $10, $1
-; MIPS-NEXT:  $BB5_6: # %entry
-; MIPS-NEXT:    sltiu $10, $25, 64
-; MIPS-NEXT:    beqz $10, $BB5_8
+; MIPS-NEXT:    subu  $9, $1, $2
+; MIPS-NEXT:    sllv  $10, $5, $9
+; MIPS-NEXT:    andi  $13, $9, 32
+; MIPS-NEXT:    andi  $3, $2, 32
+; MIPS-NEXT:    addiu $11, $zero, 0
+; MIPS-NEXT:    bnez  $13, $BB5_2
+; MIPS-NEXT:    addiu $12, $zero, 0
+; MIPS-NEXT:  # %bb.1:
+; MIPS-NEXT:    move  $12, $10
+; MIPS-NEXT:  $BB5_2:
+; MIPS-NEXT:    not $8, $2
+; MIPS-NEXT:    bnez  $3, $BB5_5
+; MIPS-NEXT:    srlv  $14, $6, $2
+; MIPS-NEXT:  # %bb.3:
+; MIPS-NEXT:    sll $1, $6, 1
+; MIPS-NEXT:    srlv  $11, $7, $2
+; MIPS-NEXT:    sllv  $1, $1, $8
+; MIPS-NEXT:    or  $15, $1, $11
+; MIPS-NEXT:    bnez  $13, $BB5_7
+; MIPS-NEXT:    move  $11, $14
+; MIPS-NEXT:  # %bb.4:
+; MIPS-NEXT:    b $BB5_6
+; MIPS-NEXT:    nop
+; MIPS-NEXT:  $BB5_5:
+; MIPS-NEXT:    bnez  $13, $BB5_7
+; MIPS-NEXT:    move  $15, $14
+; MIPS-NEXT:  $BB5_6:
+; MIPS-NEXT:    sllv  $1, $4, $9
+; MIPS-NEXT:    not $9, $9
+; MIPS-NEXT:    srl $10, $5, 1
+; MIPS-NEXT:    srlv  $9, $10, $9
+; MIPS-NEXT:    or  $10, $1, $9
+; MIPS-NEXT:  $BB5_7:
+; MIPS-NEXT:    addiu $24, $2, -64
+; MIPS-NEXT:    sll $13, $4, 1
+; MIPS-NEXT:    srav  $14, $4, $24
+; MIPS-NEXT:    andi  $1, $24, 32
+; MIPS-NEXT:    bnez  $1, $BB5_10
+; MIPS-NEXT:    sra $9, $4, 31
+; MIPS-NEXT:  # %bb.8:
+; MIPS-NEXT:    srlv  $1, $5, $24
+; MIPS-NEXT:    not $24, $24
+; MIPS-NEXT:    sllv  $24, $13, $24
+; MIPS-NEXT:    or  $25, $24, $1
+; MIPS-NEXT:    move  $24, $14
+; MIPS-NEXT:    sltiu $14, $2, 64
+; MIPS-NEXT:    beqz  $14, $BB5_12
 ; MIPS-NEXT:    nop
-; MIPS-NEXT:  # %bb.7:
-; MIPS-NEXT:    or $16, $15, $3
-; MIPS-NEXT:  $BB5_8: # %entry
-; MIPS-NEXT:    srav $15, $4, $25
-; MIPS-NEXT:    beqz $8, $BB5_20
-; MIPS-NEXT:    move $3, $15
-; MIPS-NEXT:  # %bb.9: # %entry
-; MIPS-NEXT:    sltiu $gp, $25, 1
-; MIPS-NEXT:    beqz $gp, $BB5_21
+; MIPS-NEXT:  # %bb.9:
+; MIPS-NEXT:    b $BB5_11
 ; MIPS-NEXT:    nop
-; MIPS-NEXT:  $BB5_10: # %entry
-; MIPS-NEXT:    beqz $10, $BB5_22
-; MIPS-NEXT:    sra $25, $4, 31
-; MIPS-NEXT:  $BB5_11: # %entry
-; MIPS-NEXT:    beqz $13, $BB5_23
+; MIPS-NEXT:  $BB5_10:
+; MIPS-NEXT:    move  $25, $14
+; MIPS-NEXT:    sltiu $14, $2, 64
+; MIPS-NEXT:    beqz  $14, $BB5_12
+; MIPS-NEXT:    move  $24, $9
+; MIPS-NEXT:  $BB5_11:
+; MIPS-NEXT:    or  $25, $15, $12
+; MIPS-NEXT:  $BB5_12:
+; MIPS-NEXT:    sltiu $12, $2, 1
+; MIPS-NEXT:    beqz  $12, $BB5_18
 ; MIPS-NEXT:    nop
-; MIPS-NEXT:  $BB5_12: # %entry
-; MIPS-NEXT:    beqz $8, $BB5_24
+; MIPS-NEXT:  # %bb.13:
+; MIPS-NEXT:    bnez  $14, $BB5_19
 ; MIPS-NEXT:    nop
-; MIPS-NEXT:  $BB5_13: # %entry
-; MIPS-NEXT:    beqz $24, $BB5_25
-; MIPS-NEXT:    move $4, $25
-; MIPS-NEXT:  $BB5_14: # %entry
-; MIPS-NEXT:    bnez $10, $BB5_26
+; MIPS-NEXT:  $BB5_14:
+; MIPS-NEXT:    beqz  $12, $BB5_20
 ; MIPS-NEXT:    nop
-; MIPS-NEXT:  $BB5_15: # %entry
-; MIPS-NEXT:    beqz $gp, $BB5_27
+; MIPS-NEXT:  $BB5_15:
+; MIPS-NEXT:    bnez  $3, $BB5_21
+; MIPS-NEXT:    srav  $4, $4, $2
+; MIPS-NEXT:  $BB5_16:
+; MIPS-NEXT:    srlv  $1, $5, $2
+; MIPS-NEXT:    sllv  $2, $13, $8
+; MIPS-NEXT:    or  $3, $2, $1
+; MIPS-NEXT:    bnez  $14, $BB5_23
+; MIPS-NEXT:    move  $2, $4
+; MIPS-NEXT:  # %bb.17:
+; MIPS-NEXT:    b $BB5_22
 ; MIPS-NEXT:    nop
-; MIPS-NEXT:  $BB5_16: # %entry
-; MIPS-NEXT:    beqz $8, $BB5_28
-; MIPS-NEXT:    move $2, $25
-; MIPS-NEXT:  $BB5_17: # %entry
-; MIPS-NEXT:    bnez $10, $BB5_19
+; MIPS-NEXT:  $BB5_18:
+; MIPS-NEXT:    beqz  $14, $BB5_14
+; MIPS-NEXT:    move  $7, $25
+; MIPS-NEXT:  $BB5_19:
+; MIPS-NEXT:    bnez  $12, $BB5_15
+; MIPS-NEXT:    or  $24, $11, $10
+; MIPS-NEXT:  $BB5_20:
+; MIPS-NEXT:    move  $6, $24
+; MIPS-NEXT:    beqz  $3, $BB5_16
+; MIPS-NEXT:    srav  $4, $4, $2
+; MIPS-NEXT:  $BB5_21:
+; MIPS-NEXT:    move  $2, $9
+; MIPS-NEXT:    bnez  $14, $BB5_23
+; MIPS-NEXT:    move  $3, $4
+; MIPS-NEXT:  $BB5_22:
+; MIPS-NEXT:    move  $2, $9
+; MIPS-NEXT:  $BB5_23:
+; MIPS-NEXT:    bnez  $14, $BB5_25
 ; MIPS-NEXT:    nop
-; MIPS-NEXT:  $BB5_18: # %entry
-; MIPS-NEXT:    move $2, $25
-; MIPS-NEXT:  $BB5_19: # %entry
-; MIPS-NEXT:    move $4, $6
-; MIPS-NEXT:    move $5, $7
-; MIPS-NEXT:    lw $16, 0($sp) # 4-byte Folded Reload
-; MIPS-NEXT:    lw $17, 4($sp) # 4-byte Folded Reload
-; MIPS-NEXT:    jr $ra
-; MIPS-NEXT:    addiu $sp, $sp, 8
-; MIPS-NEXT:  $BB5_20: # %entry
-; MIPS-NEXT:    srlv $1, $5, $25
-; MIPS-NEXT:    sllv $3, $17, $gp
-; MIPS-NEXT:    sltiu $gp, $25, 1
-; MIPS-NEXT:    bnez $gp, $BB5_10
-; MIPS-NEXT:    or $3, $3, $1
-; MIPS-NEXT:  $BB5_21: # %entry
-; MIPS-NEXT:    move $7, $16
-; MIPS-NEXT:    bnez $10, $BB5_11
-; MIPS-NEXT:    sra $25, $4, 31
-; MIPS-NEXT:  $BB5_22: # %entry
-; MIPS-NEXT:    bnez $13, $BB5_12
-; MIPS-NEXT:    move $3, $25
-; MIPS-NEXT:  $BB5_23: # %entry
-; MIPS-NEXT:    not $1, $11
-; MIPS-NEXT:    srl $5, $5, 1
-; MIPS-NEXT:    sllv $4, $4, $11
-; MIPS-NEXT:    srlv $1, $5, $1
-; MIPS-NEXT:    bnez $8, $BB5_13
-; MIPS-NEXT:    or $9, $4, $1
-; MIPS-NEXT:  $BB5_24: # %entry
-; MIPS-NEXT:    move $2, $12
-; MIPS-NEXT:    bnez $24, $BB5_14
-; MIPS-NEXT:    move $4, $25
-; MIPS-NEXT:  $BB5_25: # %entry
-; MIPS-NEXT:    beqz $10, $BB5_15
-; MIPS-NEXT:    move $4, $14
-; MIPS-NEXT:  $BB5_26:
-; MIPS-NEXT:    bnez $gp, $BB5_16
-; MIPS-NEXT:    or $4, $2, $9
-; MIPS-NEXT:  $BB5_27: # %entry
-; MIPS-NEXT:    move $6, $4
-; MIPS-NEXT:    bnez $8, $BB5_17
-; MIPS-NEXT:    move $2, $25
-; MIPS-NEXT:  $BB5_28: # %entry
-; MIPS-NEXT:    bnez $10, $BB5_19
-; MIPS-NEXT:    move $2, $15
-; MIPS-NEXT:  # %bb.29: # %entry
-; MIPS-NEXT:    b $BB5_18
-; MIPS-NEXT:    nop
+; MIPS-NEXT:  # %bb.24:
+; MIPS-NEXT:    move  $3, $9
+; MIPS-NEXT:  $BB5_25:
+; MIPS-NEXT:    move  $4, $6
+; MIPS-NEXT:    jr  $ra
+; MIPS-NEXT:    move  $5, $7
 ;
 ; MIPS32-LABEL: ashr_i128:
 ; MIPS32:       # %bb.0: # %entry
 ; MIPS32-NEXT:    lw $9, 28($sp)
 ; MIPS32-NEXT:    srlv $1, $7, $9
 ; MIPS32-NEXT:    not $2, $9
 ; MIPS32-NEXT:    sll $3, $6, 1
 ; MIPS32-NEXT:    sllv $3, $3, $2

 ; 32R6-NEXT:    or $3, $5, $3
 ; 32R6-NEXT:    selnez $3, $3, $15
 ; 32R6-NEXT:    or $3, $3, $12
 ; 32R6-NEXT:    jr $ra
 ; 32R6-NEXT:    move $5, $1
 ;
 ; MIPS3-LABEL: ashr_i128:
 ; MIPS3:       # %bb.0: # %entry
-; MIPS3-NEXT:    sll $8, $7, 0
-; MIPS3-NEXT:    dsrav $2, $4, $7
-; MIPS3-NEXT:    andi $6, $8, 64
-; MIPS3-NEXT:    beqz $6, .LBB5_3
-; MIPS3-NEXT:    move $3, $2
-; MIPS3-NEXT:  # %bb.1: # %entry
-; MIPS3-NEXT:    bnez $6, .LBB5_4
-; MIPS3-NEXT:    nop
-; MIPS3-NEXT:  .LBB5_2: # %entry
-; MIPS3-NEXT:    jr $ra
-; MIPS3-NEXT:    nop
-; MIPS3-NEXT:  .LBB5_3: # %entry
+; MIPS3-NEXT:    sll $2, $7, 0
+; MIPS3-NEXT:    andi  $1, $2, 64
+; MIPS3-NEXT:    bnez  $1, .LBB5_2
+; MIPS3-NEXT:    dsrav $3, $4, $7
+; MIPS3-NEXT:  # %bb.1:
 ; MIPS3-NEXT:    dsrlv $1, $5, $7
-; MIPS3-NEXT:    dsll $3, $4, 1
-; MIPS3-NEXT:    not $5, $8
-; MIPS3-NEXT:    dsllv $3, $3, $5
-; MIPS3-NEXT:    beqz $6, .LBB5_2
-; MIPS3-NEXT:    or $3, $3, $1
-; MIPS3-NEXT:  .LBB5_4:
-; MIPS3-NEXT:    jr $ra
-; MIPS3-NEXT:    dsra $2, $4, 63
+; MIPS3-NEXT:    dsll  $4, $4, 1
+; MIPS3-NEXT:    not $2, $2
+; MIPS3-NEXT:    dsllv $2, $4, $2
+; MIPS3-NEXT:    or  $1, $2, $1
+; MIPS3-NEXT:    move  $2, $3
+; MIPS3-NEXT:    jr  $ra
+; MIPS3-NEXT:    move  $3, $1
+; MIPS3-NEXT:  .LBB5_2:
+; MIPS3-NEXT:    jr  $ra
+; MIPS3-NEXT:    dsra  $2, $4, 63
+
 ;
 ; MIPS64-LABEL: ashr_i128:
 ; MIPS64:       # %bb.0: # %entry
 ; MIPS64-NEXT:    dsrlv $1, $5, $7
 ; MIPS64-NEXT:    dsll $2, $4, 1
 ; MIPS64-NEXT:    sll $5, $7, 0
 ; MIPS64-NEXT:    not $3, $5
 ; MIPS64-NEXT:    dsllv $2, $2, $3

llvm/trunk/test/CodeGen/Mips/llvm-ir/lshr.ll

 entry:
 
   %r = lshr i32 %a, %b
   ret i32 %r
 }
 
 define signext i64 @lshr_i64(i64 signext %a, i64 signext %b) {
 ; MIPS2-LABEL: lshr_i64:
-; MIPS2:       # %bb.0: # %entry
-; MIPS2-NEXT:    srlv $6, $4, $7
-; MIPS2-NEXT:    andi $8, $7, 32
-; MIPS2-NEXT:    beqz $8, $BB4_3
-; MIPS2-NEXT:    move $3, $6
-; MIPS2-NEXT:  # %bb.1: # %entry
-; MIPS2-NEXT:    beqz $8, $BB4_4
+; MIPS2:       # %bb.0:
+; MIPS2-NEXT:    srlv  $6, $4, $7
+; MIPS2-NEXT:    andi  $1, $7, 32
+; MIPS2-NEXT:    bnez  $1, $BB4_2
 ; MIPS2-NEXT:    addiu $2, $zero, 0
-; MIPS2-NEXT:  $BB4_2: # %entry
-; MIPS2-NEXT:    jr $ra
-; MIPS2-NEXT:    nop
-; MIPS2-NEXT:  $BB4_3: # %entry
-; MIPS2-NEXT:    srlv $1, $5, $7
+; MIPS2-NEXT:  # %bb.1:
+; MIPS2-NEXT:    srlv  $1, $5, $7
 ; MIPS2-NEXT:    not $2, $7
 ; MIPS2-NEXT:    sll $3, $4, 1
-; MIPS2-NEXT:    sllv $2, $3, $2
-; MIPS2-NEXT:    or $3, $2, $1
-; MIPS2-NEXT:    bnez $8, $BB4_2
-; MIPS2-NEXT:    addiu $2, $zero, 0
-; MIPS2-NEXT:  $BB4_4: # %entry
-; MIPS2-NEXT:    jr $ra
-; MIPS2-NEXT:    move $2, $6
+; MIPS2-NEXT:    sllv  $2, $3, $2
+; MIPS2-NEXT:    or  $3, $2, $1
+; MIPS2-NEXT:    jr  $ra
+; MIPS2-NEXT:    move  $2, $6
+; MIPS2-NEXT:  $BB4_2:
+; MIPS2-NEXT:    jr  $ra
+; MIPS2-NEXT:    move  $3, $6
 ;
 ; MIPS32-LABEL: lshr_i64:
 ; MIPS32:       # %bb.0: # %entry
 ; MIPS32-NEXT:    srlv $1, $5, $7
 ; MIPS32-NEXT:    not $2, $7
 ; MIPS32-NEXT:    sll $3, $4, 1
 ; MIPS32-NEXT:    sllv $2, $3, $2
 ; MIPS32-NEXT:    or $3, $2, $1

 entry:
 
   %r = lshr i64 %a, %b
   ret i64 %r
 }
 
 define signext i128 @lshr_i128(i128 signext %a, i128 signext %b) {
 ; MIPS2-LABEL: lshr_i128:
-; MIPS2:       # %bb.0: # %entry
-; MIPS2-NEXT:    addiu $sp, $sp, -8
-; MIPS2-NEXT:    .cfi_def_cfa_offset 8
-; MIPS2-NEXT:    sw $17, 4($sp) # 4-byte Folded Spill
-; MIPS2-NEXT:    sw $16, 0($sp) # 4-byte Folded Spill
-; MIPS2-NEXT:    .cfi_offset 17, -4
-; MIPS2-NEXT:    .cfi_offset 16, -8
-; MIPS2-NEXT:    lw $2, 36($sp)
+; MIPS2:       # %bb.0:
+; MIPS2-NEXT:    lw  $2, 28($sp)
 ; MIPS2-NEXT:    addiu $1, $zero, 64
-; MIPS2-NEXT:    subu $10, $1, $2
-; MIPS2-NEXT:    sllv $9, $5, $10
-; MIPS2-NEXT:    andi $13, $10, 32
-; MIPS2-NEXT:    addiu $8, $zero, 0
-; MIPS2-NEXT:    bnez $13, $BB5_2
-; MIPS2-NEXT:    addiu $25, $zero, 0
-; MIPS2-NEXT:  # %bb.1: # %entry
-; MIPS2-NEXT:    move $25, $9
-; MIPS2-NEXT:  $BB5_2: # %entry
-; MIPS2-NEXT:    not $3, $2
-; MIPS2-NEXT:    srlv $11, $6, $2
-; MIPS2-NEXT:    andi $12, $2, 32
-; MIPS2-NEXT:    bnez $12, $BB5_4
-; MIPS2-NEXT:    move $16, $11
-; MIPS2-NEXT:  # %bb.3: # %entry
-; MIPS2-NEXT:    srlv $1, $7, $2
-; MIPS2-NEXT:    sll $14, $6, 1
-; MIPS2-NEXT:    sllv $14, $14, $3
-; MIPS2-NEXT:    or $16, $14, $1
-; MIPS2-NEXT:  $BB5_4: # %entry
-; MIPS2-NEXT:    addiu $24, $2, -64
-; MIPS2-NEXT:    sll $17, $4, 1
-; MIPS2-NEXT:    srlv $14, $4, $24
-; MIPS2-NEXT:    andi $15, $24, 32
-; MIPS2-NEXT:    bnez $15, $BB5_6
-; MIPS2-NEXT:    move $gp, $14
-; MIPS2-NEXT:  # %bb.5: # %entry
-; MIPS2-NEXT:    srlv $1, $5, $24
-; MIPS2-NEXT:    not $24, $24
-; MIPS2-NEXT:    sllv $24, $17, $24
-; MIPS2-NEXT:    or $gp, $24, $1
-; MIPS2-NEXT:  $BB5_6: # %entry
-; MIPS2-NEXT:    sltiu $24, $2, 64
-; MIPS2-NEXT:    beqz $24, $BB5_8
-; MIPS2-NEXT:    nop
-; MIPS2-NEXT:  # %bb.7:
-; MIPS2-NEXT:    or $gp, $16, $25
-; MIPS2-NEXT:  $BB5_8: # %entry
-; MIPS2-NEXT:    srlv $25, $4, $2
-; MIPS2-NEXT:    bnez $12, $BB5_10
-; MIPS2-NEXT:    move $16, $25
-; MIPS2-NEXT:  # %bb.9: # %entry
-; MIPS2-NEXT:    srlv $1, $5, $2
-; MIPS2-NEXT:    sllv $3, $17, $3
-; MIPS2-NEXT:    or $16, $3, $1
-; MIPS2-NEXT:  $BB5_10: # %entry
-; MIPS2-NEXT:    bnez $12, $BB5_12
-; MIPS2-NEXT:    addiu $3, $zero, 0
-; MIPS2-NEXT:  # %bb.11: # %entry
-; MIPS2-NEXT:    move $3, $25
-; MIPS2-NEXT:  $BB5_12: # %entry
-; MIPS2-NEXT:    addiu $1, $zero, 63
-; MIPS2-NEXT:    sltiu $25, $2, 1
-; MIPS2-NEXT:    beqz $25, $BB5_22
-; MIPS2-NEXT:    sltu $17, $1, $2
-; MIPS2-NEXT:  # %bb.13: # %entry
-; MIPS2-NEXT:    beqz $17, $BB5_23
-; MIPS2-NEXT:    addiu $2, $zero, 0
-; MIPS2-NEXT:  $BB5_14: # %entry
-; MIPS2-NEXT:    beqz $17, $BB5_24
+; MIPS2-NEXT:    subu  $12, $1, $2
+; MIPS2-NEXT:    sllv  $10, $5, $12
+; MIPS2-NEXT:    andi  $15, $12, 32
+; MIPS2-NEXT:    andi  $8, $2, 32
 ; MIPS2-NEXT:    addiu $3, $zero, 0
-; MIPS2-NEXT:  $BB5_15: # %entry
-; MIPS2-NEXT:    beqz $13, $BB5_25
+; MIPS2-NEXT:    bnez  $15, $BB5_2
+; MIPS2-NEXT:    addiu $13, $zero, 0
+; MIPS2-NEXT:  # %bb.1:
+; MIPS2-NEXT:    move  $13, $10
+; MIPS2-NEXT:  $BB5_2:
+; MIPS2-NEXT:    not $9, $2
+; MIPS2-NEXT:    bnez  $8, $BB5_5
+; MIPS2-NEXT:    srlv  $24, $6, $2
+; MIPS2-NEXT:  # %bb.3:
+; MIPS2-NEXT:    sll $1, $6, 1
+; MIPS2-NEXT:    srlv  $11, $7, $2
+; MIPS2-NEXT:    sllv  $1, $1, $9
+; MIPS2-NEXT:    or  $14, $1, $11
+; MIPS2-NEXT:    bnez  $15, $BB5_7
+; MIPS2-NEXT:    move  $11, $24
+; MIPS2-NEXT:  # %bb.4:
+; MIPS2-NEXT:    b $BB5_6
 ; MIPS2-NEXT:    nop
-; MIPS2-NEXT:  $BB5_16: # %entry
-; MIPS2-NEXT:    beqz $12, $BB5_26
-; MIPS2-NEXT:    addiu $4, $zero, 0
-; MIPS2-NEXT:  $BB5_17: # %entry
-; MIPS2-NEXT:    beqz $15, $BB5_27
+; MIPS2-NEXT:  $BB5_5:
+; MIPS2-NEXT:    addiu $11, $zero, 0
+; MIPS2-NEXT:    bnez  $15, $BB5_7
+; MIPS2-NEXT:    move  $14, $24
+; MIPS2-NEXT:  $BB5_6:
+; MIPS2-NEXT:    sllv  $1, $4, $12
+; MIPS2-NEXT:    not $10, $12
+; MIPS2-NEXT:    srl $12, $5, 1
+; MIPS2-NEXT:    srlv  $10, $12, $10
+; MIPS2-NEXT:    or  $10, $1, $10
+; MIPS2-NEXT:  $BB5_7:
+; MIPS2-NEXT:    addiu $15, $2, -64
+; MIPS2-NEXT:    sll $12, $4, 1
+; MIPS2-NEXT:    andi  $1, $15, 32
+; MIPS2-NEXT:    bnez  $1, $BB5_10
+; MIPS2-NEXT:    srlv  $25, $4, $15
+; MIPS2-NEXT:  # %bb.8:
+; MIPS2-NEXT:    srlv  $1, $5, $15
+; MIPS2-NEXT:    not $15, $15
+; MIPS2-NEXT:    sllv  $15, $12, $15
+; MIPS2-NEXT:    or  $24, $15, $1
+; MIPS2-NEXT:    move  $15, $25
+; MIPS2-NEXT:    sltiu $25, $2, 64
+; MIPS2-NEXT:    beqz  $25, $BB5_12
 ; MIPS2-NEXT:    nop
-; MIPS2-NEXT:  $BB5_18: # %entry
-; MIPS2-NEXT:    bnez $24, $BB5_28
+; MIPS2-NEXT:  # %bb.9:
+; MIPS2-NEXT:    b $BB5_11
 ; MIPS2-NEXT:    nop
-; MIPS2-NEXT:  $BB5_19: # %entry
-; MIPS2-NEXT:    bnez $25, $BB5_21
+; MIPS2-NEXT:  $BB5_10:
+; MIPS2-NEXT:    move  $24, $25
+; MIPS2-NEXT:    sltiu $25, $2, 64
+; MIPS2-NEXT:    beqz  $25, $BB5_12
+; MIPS2-NEXT:    addiu $15, $zero, 0
+; MIPS2-NEXT:  $BB5_11:
+; MIPS2-NEXT:    or  $24, $14, $13
+; MIPS2-NEXT:  $BB5_12:
+; MIPS2-NEXT:    sltiu $13, $2, 1
+; MIPS2-NEXT:    beqz  $13, $BB5_19
 ; MIPS2-NEXT:    nop
-; MIPS2-NEXT:  $BB5_20: # %entry
-; MIPS2-NEXT:    move $6, $8
-; MIPS2-NEXT:  $BB5_21: # %entry
-; MIPS2-NEXT:    move $4, $6
-; MIPS2-NEXT:    move $5, $7
-; MIPS2-NEXT:    lw $16, 0($sp) # 4-byte Folded Reload
-; MIPS2-NEXT:    lw $17, 4($sp) # 4-byte Folded Reload
-; MIPS2-NEXT:    jr $ra
-; MIPS2-NEXT:    addiu $sp, $sp, 8
-; MIPS2-NEXT:  $BB5_22: # %entry
-; MIPS2-NEXT:    move $7, $gp
-; MIPS2-NEXT:    bnez $17, $BB5_14
+; MIPS2-NEXT:  # %bb.13:
+; MIPS2-NEXT:    bnez  $25, $BB5_20
+; MIPS2-NEXT:    nop
+; MIPS2-NEXT:  $BB5_14:
+; MIPS2-NEXT:    bnez  $13, $BB5_16
+; MIPS2-NEXT:    addiu $10, $zero, 63
+; MIPS2-NEXT:  $BB5_15:
+; MIPS2-NEXT:    move  $6, $15
+; MIPS2-NEXT:  $BB5_16:
+; MIPS2-NEXT:    sltu  $10, $10, $2
+; MIPS2-NEXT:    bnez  $8, $BB5_22
+; MIPS2-NEXT:    srlv  $11, $4, $2
+; MIPS2-NEXT:  # %bb.17:
+; MIPS2-NEXT:    srlv  $1, $5, $2
+; MIPS2-NEXT:    sllv  $2, $12, $9
+; MIPS2-NEXT:    or  $4, $2, $1
+; MIPS2-NEXT:    move  $5, $11
+; MIPS2-NEXT:    bnez  $10, $BB5_24
 ; MIPS2-NEXT:    addiu $2, $zero, 0
-; MIPS2-NEXT:  $BB5_23: # %entry
-; MIPS2-NEXT:    move $2, $3
-; MIPS2-NEXT:    bnez $17, $BB5_15
-; MIPS2-NEXT:    addiu $3, $zero, 0
-; MIPS2-NEXT:  $BB5_24: # %entry
-; MIPS2-NEXT:    bnez $13, $BB5_16
-; MIPS2-NEXT:    move $3, $16
-; MIPS2-NEXT:  $BB5_25: # %entry
-; MIPS2-NEXT:    not $1, $10
-; MIPS2-NEXT:    srl $5, $5, 1
-; MIPS2-NEXT:    sllv $4, $4, $10
-; MIPS2-NEXT:    srlv $1, $5, $1
-; MIPS2-NEXT:    or $9, $4, $1
-; MIPS2-NEXT:    bnez $12, $BB5_17
-; MIPS2-NEXT:    addiu $4, $zero, 0
-; MIPS2-NEXT:  $BB5_26: # %entry
-; MIPS2-NEXT:    bnez $15, $BB5_18
-; MIPS2-NEXT:    move $4, $11
-; MIPS2-NEXT:  $BB5_27: # %entry
-; MIPS2-NEXT:    beqz $24, $BB5_19
-; MIPS2-NEXT:    move $8, $14
-; MIPS2-NEXT:  $BB5_28:
-; MIPS2-NEXT:    bnez $25, $BB5_21
-; MIPS2-NEXT:    or $8, $4, $9
-; MIPS2-NEXT:  # %bb.29:
-; MIPS2-NEXT:    b $BB5_20
+; MIPS2-NEXT:  # %bb.18:
+; MIPS2-NEXT:    b $BB5_23
+; MIPS2-NEXT:    nop
+; MIPS2-NEXT:  $BB5_19:
+; MIPS2-NEXT:    beqz  $25, $BB5_14
+; MIPS2-NEXT:    move  $7, $24
+; MIPS2-NEXT:  $BB5_20:
+; MIPS2-NEXT:    or  $15, $11, $10
+; MIPS2-NEXT:    bnez  $13, $BB5_16
+; MIPS2-NEXT:    addiu $10, $zero, 63
+; MIPS2-NEXT:  # %bb.21:
+; MIPS2-NEXT:    b $BB5_15
+; MIPS2-NEXT:    nop
+; MIPS2-NEXT:  $BB5_22:
+; MIPS2-NEXT:    addiu $5, $zero, 0
+; MIPS2-NEXT:    move  $4, $11
+; MIPS2-NEXT:    bnez  $10, $BB5_24
+; MIPS2-NEXT:    addiu $2, $zero, 0
+; MIPS2-NEXT:  $BB5_23:
+; MIPS2-NEXT:    move  $2, $5
+; MIPS2-NEXT:  $BB5_24:
+; MIPS2-NEXT:    bnez  $10, $BB5_26
 ; MIPS2-NEXT:    nop
+; MIPS2-NEXT:  # %bb.25:
+; MIPS2-NEXT:    move  $3, $4
+; MIPS2-NEXT:  $BB5_26:
+; MIPS2-NEXT:    move  $4, $6
+; MIPS2-NEXT:    jr  $ra
+; MIPS2-NEXT:    move  $5, $7
 ;
 ; MIPS32-LABEL: lshr_i128:
 ; MIPS32:       # %bb.0: # %entry
 ; MIPS32-NEXT:    lw $9, 28($sp)
 ; MIPS32-NEXT:    addiu $1, $zero, 64
 ; MIPS32-NEXT:    subu $2, $1, $9
 ; MIPS32-NEXT:    sllv $10, $5, $2
 ; MIPS32-NEXT:    andi $11, $2, 32

 ; MIPS32R6-NEXT:    or $1, $7, $3
 ; MIPS32R6-NEXT:    selnez $1, $1, $15
 ; MIPS32R6-NEXT:    or $3, $9, $1
 ; MIPS32R6-NEXT:    lw $16, 4($sp) # 4-byte Folded Reload
 ; MIPS32R6-NEXT:    jr $ra
 ; MIPS32R6-NEXT:    addiu $sp, $sp, 8
 ;
 ; MIPS3-LABEL: lshr_i128:
-; MIPS3:       # %bb.0: # %entry
-; MIPS3-NEXT:    sll $2, $7, 0
+; MIPS3:       # %bb.0:
+; MIPS3-NEXT:    sll $3, $7, 0
 ; MIPS3-NEXT:    dsrlv $6, $4, $7
-; MIPS3-NEXT:    andi $8, $2, 64
-; MIPS3-NEXT:    beqz $8, .LBB5_3
-; MIPS3-NEXT:    move $3, $6
-; MIPS3-NEXT:  # %bb.1: # %entry
-; MIPS3-NEXT:    beqz $8, .LBB5_4
-; MIPS3-NEXT:    daddiu $2, $zero, 0
-; MIPS3-NEXT:  .LBB5_2: # %entry
-; MIPS3-NEXT:    jr $ra
-; MIPS3-NEXT:    nop
-; MIPS3-NEXT:  .LBB5_3: # %entry
+; MIPS3-NEXT:    andi  $1, $3, 64
+; MIPS3-NEXT:    bnez  $1, .LBB5_2
+; MIPS3-NEXT:    daddiu  $2, $zero, 0
+; MIPS3-NEXT:  # %bb.1:
 ; MIPS3-NEXT:    dsrlv $1, $5, $7
-; MIPS3-NEXT:    dsll $3, $4, 1
-; MIPS3-NEXT:    not $2, $2
-; MIPS3-NEXT:    dsllv $2, $3, $2
-; MIPS3-NEXT:    or $3, $2, $1
-; MIPS3-NEXT:    bnez $8, .LBB5_2
-; MIPS3-NEXT:    daddiu $2, $zero, 0
-; MIPS3-NEXT:  .LBB5_4: # %entry
-; MIPS3-NEXT:    jr $ra
-; MIPS3-NEXT:    move $2, $6
+; MIPS3-NEXT:    dsll  $2, $4, 1
+; MIPS3-NEXT:    not $3, $3
+; MIPS3-NEXT:    dsllv $2, $2, $3
+; MIPS3-NEXT:    or  $3, $2, $1
+; MIPS3-NEXT:    jr  $ra
+; MIPS3-NEXT:    move  $2, $6
+; MIPS3-NEXT:  .LBB5_2:
+; MIPS3-NEXT:    jr  $ra
+; MIPS3-NEXT:    move  $3, $6
 ;
 ; MIPS4-LABEL: lshr_i128:
 ; MIPS4:       # %bb.0: # %entry
 ; MIPS4-NEXT:    dsrlv $1, $5, $7
 ; MIPS4-NEXT:    dsll $2, $4, 1
 ; MIPS4-NEXT:    sll $5, $7, 0
 ; MIPS4-NEXT:    not $3, $5
 ; MIPS4-NEXT:    dsllv $2, $2, $3

llvm/trunk/test/CodeGen/Mips/llvm-ir/shl.ll

 entry:
 
   %r = shl i32 %a, %b
   ret i32 %r
 }
 
 define signext i64 @shl_i64(i64 signext %a, i64 signext %b) {
 ; MIPS2-LABEL: shl_i64:
-; MIPS2:       # %bb.0: # %entry
-; MIPS2-NEXT:    sllv $6, $5, $7
-; MIPS2-NEXT:    andi $8, $7, 32
-; MIPS2-NEXT:    beqz $8, $BB4_3
-; MIPS2-NEXT:    move $2, $6
-; MIPS2-NEXT:  # %bb.1: # %entry
-; MIPS2-NEXT:    beqz $8, $BB4_4
+; MIPS2:       # %bb.0:
+; MIPS2-NEXT:    sllv  $6, $5, $7
+; MIPS2-NEXT:    andi  $8, $7, 32
+; MIPS2-NEXT:    beqz  $8, $BB4_3
+; MIPS2-NEXT:    move  $2, $6
+; MIPS2-NEXT:  # %bb.1:
+; MIPS2-NEXT:    beqz  $8, $BB4_4
 ; MIPS2-NEXT:    addiu $3, $zero, 0
-; MIPS2-NEXT:  $BB4_2: # %entry
-; MIPS2-NEXT:    jr $ra
+; MIPS2-NEXT:  $BB4_2:
+; MIPS2-NEXT:    jr  $ra
 ; MIPS2-NEXT:    nop
-; MIPS2-NEXT:  $BB4_3: # %entry
-; MIPS2-NEXT:    sllv $1, $4, $7
+; MIPS2-NEXT:  $BB4_3:
+; MIPS2-NEXT:    sllv  $1, $4, $7
 ; MIPS2-NEXT:    not $2, $7
 ; MIPS2-NEXT:    srl $3, $5, 1
-; MIPS2-NEXT:    srlv $2, $3, $2
-; MIPS2-NEXT:    or $2, $1, $2
-; MIPS2-NEXT:    bnez $8, $BB4_2
+; MIPS2-NEXT:    srlv  $2, $3, $2
+; MIPS2-NEXT:    or  $2, $1, $2
+; MIPS2-NEXT:    bnez  $8, $BB4_2
 ; MIPS2-NEXT:    addiu $3, $zero, 0
-; MIPS2-NEXT:  $BB4_4: # %entry
-; MIPS2-NEXT:    jr $ra
-; MIPS2-NEXT:    move $3, $6
+; MIPS2-NEXT:  $BB4_4:
+; MIPS2-NEXT:    jr  $ra
+; MIPS2-NEXT:    move  $3, $6
 ;
 ; MIPS32-LABEL: shl_i64:
 ; MIPS32:       # %bb.0: # %entry
 ; MIPS32-NEXT:    sllv $1, $4, $7
 ; MIPS32-NEXT:    not $2, $7
 ; MIPS32-NEXT:    srl $3, $5, 1
 ; MIPS32-NEXT:    srlv $2, $3, $2
 ; MIPS32-NEXT:    or $2, $1, $2

 entry:
 
   %r = shl i64 %a, %b
   ret i64 %r
 }
 
 define signext i128 @shl_i128(i128 signext %a, i128 signext %b) {
 ; MIPS2-LABEL: shl_i128:
-; MIPS2:       # %bb.0: # %entry
+; MIPS2:       # %bb.0:
 ; MIPS2-NEXT:    addiu $sp, $sp, -8
 ; MIPS2-NEXT:    .cfi_def_cfa_offset 8
-; MIPS2-NEXT:    sw $17, 4($sp) # 4-byte Folded Spill
-; MIPS2-NEXT:    sw $16, 0($sp) # 4-byte Folded Spill
+; MIPS2-NEXT:    sw  $17, 4($sp)
+; MIPS2-NEXT:    sw  $16, 0($sp)
 ; MIPS2-NEXT:    .cfi_offset 17, -4
 ; MIPS2-NEXT:    .cfi_offset 16, -8
-; MIPS2-NEXT:    lw $8, 36($sp)
+; MIPS2-NEXT:    lw  $8, 36($sp)
 ; MIPS2-NEXT:    addiu $1, $zero, 64
-; MIPS2-NEXT:    subu $10, $1, $8
-; MIPS2-NEXT:    srlv $3, $6, $10
-; MIPS2-NEXT:    andi $13, $10, 32
+; MIPS2-NEXT:    subu  $3, $1, $8
+; MIPS2-NEXT:    srlv  $9, $6, $3
+; MIPS2-NEXT:    andi  $1, $3, 32
+; MIPS2-NEXT:    bnez  $1, $BB5_2
 ; MIPS2-NEXT:    addiu $2, $zero, 0
-; MIPS2-NEXT:    bnez $13, $BB5_2
-; MIPS2-NEXT:    addiu $25, $zero, 0
-; MIPS2-NEXT:  # %bb.1: # %entry
-; MIPS2-NEXT:    move $25, $3
-; MIPS2-NEXT:  $BB5_2: # %entry
-; MIPS2-NEXT:    not $9, $8
-; MIPS2-NEXT:    sllv $11, $5, $8
-; MIPS2-NEXT:    andi $12, $8, 32
-; MIPS2-NEXT:    bnez $12, $BB5_4
-; MIPS2-NEXT:    move $16, $11
-; MIPS2-NEXT:  # %bb.3: # %entry
-; MIPS2-NEXT:    sllv $1, $4, $8
-; MIPS2-NEXT:    srl $14, $5, 1
-; MIPS2-NEXT:    srlv $14, $14, $9
-; MIPS2-NEXT:    or $16, $1, $14
-; MIPS2-NEXT:  $BB5_4: # %entry
-; MIPS2-NEXT:    addiu $24, $8, -64
-; MIPS2-NEXT:    srl $17, $7, 1
-; MIPS2-NEXT:    sllv $14, $7, $24
-; MIPS2-NEXT:    andi $15, $24, 32
-; MIPS2-NEXT:    bnez $15, $BB5_6
-; MIPS2-NEXT:    move $gp, $14
-; MIPS2-NEXT:  # %bb.5: # %entry
-; MIPS2-NEXT:    sllv $1, $6, $24
-; MIPS2-NEXT:    not $24, $24
-; MIPS2-NEXT:    srlv $24, $17, $24
-; MIPS2-NEXT:    or $gp, $1, $24
-; MIPS2-NEXT:  $BB5_6: # %entry
-; MIPS2-NEXT:    sltiu $24, $8, 64
-; MIPS2-NEXT:    beqz $24, $BB5_8
+; MIPS2-NEXT:  # %bb.1:
+; MIPS2-NEXT:    srlv  $1, $7, $3
+; MIPS2-NEXT:    not $3, $3
+; MIPS2-NEXT:    sll $10, $6, 1
+; MIPS2-NEXT:    sllv  $3, $10, $3
+; MIPS2-NEXT:    or  $3, $3, $1
+; MIPS2-NEXT:    b $BB5_3
+; MIPS2-NEXT:    move  $15, $9
+; MIPS2-NEXT:  $BB5_2:
+; MIPS2-NEXT:    addiu $15, $zero, 0
+; MIPS2-NEXT:    move  $3, $9
+; MIPS2-NEXT:  $BB5_3:
+; MIPS2-NEXT:    not $13, $8
+; MIPS2-NEXT:    sllv  $9, $5, $8
+; MIPS2-NEXT:    andi  $10, $8, 32
+; MIPS2-NEXT:    bnez  $10, $BB5_5
+; MIPS2-NEXT:    move  $25, $9
+; MIPS2-NEXT:  # %bb.4:
+; MIPS2-NEXT:    sllv  $1, $4, $8
+; MIPS2-NEXT:    srl $11, $5, 1
+; MIPS2-NEXT:    srlv  $11, $11, $13
+; MIPS2-NEXT:    or  $25, $1, $11
+; MIPS2-NEXT:  $BB5_5:
+; MIPS2-NEXT:    addiu $14, $8, -64
+; MIPS2-NEXT:    srl $24, $7, 1
+; MIPS2-NEXT:    sllv  $11, $7, $14
+; MIPS2-NEXT:    andi  $12, $14, 32
+; MIPS2-NEXT:    bnez  $12, $BB5_7
+; MIPS2-NEXT:    move  $gp, $11
+; MIPS2-NEXT:  # %bb.6:
+; MIPS2-NEXT:    sllv  $1, $6, $14
+; MIPS2-NEXT:    not $14, $14
+; MIPS2-NEXT:    srlv  $14, $24, $14
+; MIPS2-NEXT:    or  $gp, $1, $14
+; MIPS2-NEXT:  $BB5_7:
+; MIPS2-NEXT:    sltiu $14, $8, 64
+; MIPS2-NEXT:    beqz  $14, $BB5_9
 ; MIPS2-NEXT:    nop
-; MIPS2-NEXT:  # %bb.7:
-; MIPS2-NEXT:    or $gp, $16, $25
-; MIPS2-NEXT:  $BB5_8: # %entry
-; MIPS2-NEXT:    sllv $25, $7, $8
-; MIPS2-NEXT:    bnez $12, $BB5_10
-; MIPS2-NEXT:    move $16, $25
-; MIPS2-NEXT:  # %bb.9: # %entry
-; MIPS2-NEXT:    sllv $1, $6, $8
-; MIPS2-NEXT:    srlv $9, $17, $9
-; MIPS2-NEXT:    or $16, $1, $9
-; MIPS2-NEXT:  $BB5_10: # %entry
-; MIPS2-NEXT:    bnez $12, $BB5_12
-; MIPS2-NEXT:    addiu $9, $zero, 0
-; MIPS2-NEXT:  # %bb.11: # %entry
-; MIPS2-NEXT:    move $9, $25
-; MIPS2-NEXT:  $BB5_12: # %entry
+; MIPS2-NEXT:  # %bb.8:
+; MIPS2-NEXT:    or  $gp, $25, $15
+; MIPS2-NEXT:  $BB5_9:
+; MIPS2-NEXT:    sllv  $25, $7, $8
+; MIPS2-NEXT:    bnez  $10, $BB5_11
+; MIPS2-NEXT:    addiu $17, $zero, 0
+; MIPS2-NEXT:  # %bb.10:
+; MIPS2-NEXT:    move  $17, $25
+; MIPS2-NEXT:  $BB5_11:
 ; MIPS2-NEXT:    addiu $1, $zero, 63
-; MIPS2-NEXT:    sltiu $25, $8, 1
-; MIPS2-NEXT:    beqz $25, $BB5_22
-; MIPS2-NEXT:    sltu $17, $1, $8
-; MIPS2-NEXT:  # %bb.13: # %entry
-; MIPS2-NEXT:    beqz $17, $BB5_23
-; MIPS2-NEXT:    addiu $8, $zero, 0
-; MIPS2-NEXT:  $BB5_14: # %entry
-; MIPS2-NEXT:    beqz $17, $BB5_24
-; MIPS2-NEXT:    addiu $9, $zero, 0
-; MIPS2-NEXT:  $BB5_15: # %entry
-; MIPS2-NEXT:    beqz $13, $BB5_25
+; MIPS2-NEXT:    sltiu $15, $8, 1
+; MIPS2-NEXT:    beqz  $15, $BB5_21
+; MIPS2-NEXT:    sltu  $16, $1, $8
+; MIPS2-NEXT:  # %bb.12:
+; MIPS2-NEXT:    beqz  $16, $BB5_22
+; MIPS2-NEXT:    addiu $7, $zero, 0
+; MIPS2-NEXT:  $BB5_13:
+; MIPS2-NEXT:    beqz  $10, $BB5_23
 ; MIPS2-NEXT:    nop
-; MIPS2-NEXT:  $BB5_16: # %entry
-; MIPS2-NEXT:    beqz $12, $BB5_26
+; MIPS2-NEXT:  $BB5_14:
+; MIPS2-NEXT:    beqz  $16, $BB5_24
 ; MIPS2-NEXT:    addiu $6, $zero, 0
-; MIPS2-NEXT:  $BB5_17: # %entry
-; MIPS2-NEXT:    beqz $15, $BB5_27
+; MIPS2-NEXT:  $BB5_15:
+; MIPS2-NEXT:    beqz  $10, $BB5_25
+; MIPS2-NEXT:    addiu $8, $zero, 0
+; MIPS2-NEXT:  $BB5_16:
+; MIPS2-NEXT:    beqz  $12, $BB5_26
 ; MIPS2-NEXT:    nop
-; MIPS2-NEXT:  $BB5_18: # %entry
-; MIPS2-NEXT:    bnez $24, $BB5_28
+; MIPS2-NEXT:  $BB5_17:
+; MIPS2-NEXT:    bnez  $14, $BB5_27
 ; MIPS2-NEXT:    nop
-; MIPS2-NEXT:  $BB5_19: # %entry
-; MIPS2-NEXT:    bnez $25, $BB5_21
+; MIPS2-NEXT:  $BB5_18:
+; MIPS2-NEXT:    bnez  $15, $BB5_20
 ; MIPS2-NEXT:    nop
-; MIPS2-NEXT:  $BB5_20: # %entry
-; MIPS2-NEXT:    move $5, $2
-; MIPS2-NEXT:  $BB5_21: # %entry
-; MIPS2-NEXT:    move $2, $4
-; MIPS2-NEXT:    move $3, $5
-; MIPS2-NEXT:    move $4, $9
-; MIPS2-NEXT:    move $5, $8
-; MIPS2-NEXT:    lw $16, 0($sp) # 4-byte Folded Reload
-; MIPS2-NEXT:    lw $17, 4($sp) # 4-byte Folded Reload
-; MIPS2-NEXT:    jr $ra
+; MIPS2-NEXT:  $BB5_19:
+; MIPS2-NEXT:    move  $5, $2
+; MIPS2-NEXT:  $BB5_20:
+; MIPS2-NEXT:    move  $2, $4
+; MIPS2-NEXT:    move  $3, $5
+; MIPS2-NEXT:    move  $4, $6
+; MIPS2-NEXT:    move  $5, $7
+; MIPS2-NEXT:    lw  $16, 0($sp)
+; MIPS2-NEXT:    lw  $17, 4($sp)
+; MIPS2-NEXT:    jr  $ra
 ; MIPS2-NEXT:    addiu $sp, $sp, 8
-; MIPS2-NEXT:  $BB5_22: # %entry
-; MIPS2-NEXT:    move $4, $gp
-; MIPS2-NEXT:    bnez $17, $BB5_14
+; MIPS2-NEXT:  $BB5_21:
+; MIPS2-NEXT:    move  $4, $gp
+; MIPS2-NEXT:    bnez  $16, $BB5_13
+; MIPS2-NEXT:    addiu $7, $zero, 0
+; MIPS2-NEXT:  $BB5_22:
+; MIPS2-NEXT:    bnez  $10, $BB5_14
+; MIPS2-NEXT:    move  $7, $17
+; MIPS2-NEXT:  $BB5_23:
+; MIPS2-NEXT:    sllv  $1, $6, $8
+; MIPS2-NEXT:    srlv  $6, $24, $13
+; MIPS2-NEXT:    or  $25, $1, $6
+; MIPS2-NEXT:    bnez  $16, $BB5_15
+; MIPS2-NEXT:    addiu $6, $zero, 0
+; MIPS2-NEXT:  $BB5_24:
+; MIPS2-NEXT:    move  $6, $25
+; MIPS2-NEXT:    bnez  $10, $BB5_16
 ; MIPS2-NEXT:    addiu $8, $zero, 0
-; MIPS2-NEXT:  $BB5_23: # %entry
-; MIPS2-NEXT:    move $8, $9
-; MIPS2-NEXT:    bnez $17, $BB5_15
-; MIPS2-NEXT:    addiu $9, $zero, 0
-; MIPS2-NEXT:  $BB5_24: # %entry
-; MIPS2-NEXT:    bnez $13, $BB5_16
-; MIPS2-NEXT:    move $9, $16
-; MIPS2-NEXT:  $BB5_25: # %entry
-; MIPS2-NEXT:    not $1, $10
-; MIPS2-NEXT:    sll $3, $6, 1
-; MIPS2-NEXT:    srlv $6, $7, $10
-; MIPS2-NEXT:    sllv $1, $3, $1
-; MIPS2-NEXT:    or $3, $1, $6
-; MIPS2-NEXT:    bnez $12, $BB5_17
-; MIPS2-NEXT:    addiu $6, $zero, 0
-; MIPS2-NEXT:  $BB5_26: # %entry
-; MIPS2-NEXT:    bnez $15, $BB5_18
-; MIPS2-NEXT:    move $6, $11
-; MIPS2-NEXT:  $BB5_27: # %entry
-; MIPS2-NEXT:    beqz $24, $BB5_19
-; MIPS2-NEXT:    move $2, $14
-; MIPS2-NEXT:  $BB5_28:
-; MIPS2-NEXT:    bnez $25, $BB5_21
-; MIPS2-NEXT:    or $2, $6, $3
-; MIPS2-NEXT:  # %bb.29:
-; MIPS2-NEXT:    b $BB5_20
+; MIPS2-NEXT:  $BB5_25:
+; MIPS2-NEXT:    bnez  $12, $BB5_17
+; MIPS2-NEXT:    move  $8, $9
+; MIPS2-NEXT:  $BB5_26:
+; MIPS2-NEXT:    beqz  $14, $BB5_18
+; MIPS2-NEXT:    move  $2, $11
+; MIPS2-NEXT:  $BB5_27:
+; MIPS2-NEXT:    bnez  $15, $BB5_20
+; MIPS2-NEXT:    or  $2, $8, $3
+; MIPS2-NEXT:  # %bb.28:
+; MIPS2-NEXT:    b $BB5_19
 ; MIPS2-NEXT:    nop
 ;
 ; MIPS32-LABEL: shl_i128:
 ; MIPS32:       # %bb.0: # %entry
 ; MIPS32-NEXT:    lw $8, 28($sp)
 ; MIPS32-NEXT:    addiu $1, $zero, 64
 ; MIPS32-NEXT:    subu $1, $1, $8
 ; MIPS32-NEXT:    srlv $9, $6, $1

 ; MIPS32R6-NEXT:    or $3, $5, $3
 ; MIPS32R6-NEXT:    or $4, $6, $4
 ; MIPS32R6-NEXT:    selnez $4, $4, $24
 ; MIPS32R6-NEXT:    or $4, $10, $4
 ; MIPS32R6-NEXT:    jr $ra
 ; MIPS32R6-NEXT:    move $5, $1
 ;
 ; MIPS3-LABEL: shl_i128:
-; MIPS3:       # %bb.0: # %entry
+; MIPS3:       # %bb.0:
 ; MIPS3-NEXT:    sll $3, $7, 0
 ; MIPS3-NEXT:    dsllv $6, $5, $7
-; MIPS3-NEXT:    andi $8, $3, 64
-; MIPS3-NEXT:    beqz $8, .LBB5_3
-; MIPS3-NEXT:    move $2, $6
-; MIPS3-NEXT:  # %bb.1: # %entry
-; MIPS3-NEXT:    beqz $8, .LBB5_4
-; MIPS3-NEXT:    daddiu $3, $zero, 0
-; MIPS3-NEXT:  .LBB5_2: # %entry
-; MIPS3-NEXT:    jr $ra
+; MIPS3-NEXT:    andi  $8, $3, 64
+; MIPS3-NEXT:    beqz  $8, .LBB5_3
+; MIPS3-NEXT:    move  $2, $6
+; MIPS3-NEXT:  # %bb.1:
+; MIPS3-NEXT:    beqz  $8, .LBB5_4
+; MIPS3-NEXT:    daddiu  $3, $zero, 0
+; MIPS3-NEXT:  .LBB5_2:
+; MIPS3-NEXT:    jr  $ra
 ; MIPS3-NEXT:    nop
-; MIPS3-NEXT:  .LBB5_3: # %entry
+; MIPS3-NEXT:  .LBB5_3:
 ; MIPS3-NEXT:    dsllv $1, $4, $7
-; MIPS3-NEXT:    dsrl $2, $5, 1
+; MIPS3-NEXT:    dsrl  $2, $5, 1
 ; MIPS3-NEXT:    not $3, $3
 ; MIPS3-NEXT:    dsrlv $2, $2, $3
-; MIPS3-NEXT:    or $2, $1, $2
-; MIPS3-NEXT:    bnez $8, .LBB5_2
-; MIPS3-NEXT:    daddiu $3, $zero, 0
-; MIPS3-NEXT:  .LBB5_4: # %entry
-; MIPS3-NEXT:    jr $ra
-; MIPS3-NEXT:    move $3, $6
+; MIPS3-NEXT:    or  $2, $1, $2
+; MIPS3-NEXT:    bnez  $8, .LBB5_2
+; MIPS3-NEXT:    daddiu  $3, $zero, 0
+; MIPS3-NEXT:  .LBB5_4:
+; MIPS3-NEXT:    jr  $ra
+; MIPS3-NEXT:    move  $3, $6
 ;
 ; MIPS4-LABEL: shl_i128:
 ; MIPS4:       # %bb.0: # %entry
 ; MIPS4-NEXT:    dsllv $1, $4, $7
 ; MIPS4-NEXT:    dsrl $2, $5, 1
 ; MIPS4-NEXT:    sll $4, $7, 0
 ; MIPS4-NEXT:    not $3, $4
 ; MIPS4-NEXT:    dsrlv $2, $2, $3