VMALF  X1, YDIG, ADD2H, ADD4
	VMALHF X0, YDIG, ADD1H, ADD3H // ADD1H Free->ADD3H
	VMALHF X1, YDIG, ADD2H, ADD4H // ADD2H Free->ADD4H , YDIG Free->ZER

	VZERO ZER
	VL    32(CPOOL), SEL1
	VPERM ZER, ADD1, SEL1, RED3 // [d0 0 0 d0]

	VSLDB $12, ADD2, ADD1, T0 // ADD1 Free->T0
	VSLDB $12, T2, ADD2, T1   // ADD2 Free->T1, T2 Free

	VACCQ  T0, RED1, CAR1
	VAQ    T0, RED1, T0
	VACCCQ T1, RED2, CAR1, T2

	VMULT_ADD(X0, YDIG, T0, ONE, ADD1, ADD1H)
	VMULT_ADD(X1, YDIG, T1, ONE, ADD2, ADD2H)

	VSPLTW $0, Y0, YDIG // VREPF

	//	VMALF  X0, YDIG, ADD1H, ADD3
	//	VMALF  X1, YDIG, ADD2H, ADD4
	//	VMALHF X0, YDIG, ADD1H, ADD3H // ADD1H Free->ADD3H
	//	VMALHF X1, YDIG, ADD2H, ADD4H // ADD2H Free->ADD4H , YDIG Free->ZER
	VMULT_ADD(X0, YDIG, ADD1H, ONE, ADD3, ADD3H)
	VMULT_ADD(X1, YDIG, ADD2H, ONE, ADD4, ADD4H)

  // CHECK-NEXT: %[[ADD2:.*]] = "tf.AddV2"(%[[BARG0]], %[[DELTA]])
  %add2 = "tf.AddV2"(%arg0, %constant) : (tensor<i32>, tensor<i32>) -> tensor<i32>
  // CHECK-NEXT: return %[[ADD2]], %[[ADD1]]
  %id = "tf.Identity"(%arg2) : (tensor<*x!tf_type.resource<tensor<f32>>>) -> tensor<*x!tf_type.resource<tensor<f32>>>

  // CHECK: %[[OLD_SLICE2:.*]] = "tf.Slice"(%[[READ2]],
  // CHECK: %[[RESHAPE2:.*]] = "tf.Reshape"(%[[VALUE]],
  // CHECK: %[[ADD2:.*]] = "tf.AddV2"(%[[RESHAPE2]], %[[OLD_SLICE2]]) : (tensor<1x3xf32>, tensor<1x3xf32>) -> tensor<1x3xf32>
  // CHECK: %[[UPDATE2:.*]] = "tf.XlaDynamicUpdateSlice"(%[[READ2]], %[[ADD2]],
  // CHECK: "tf.AssignVariableOp"(%[[GVAR1]], %[[UPDATE2]])

  Output add1 =
      ops::Add(root.WithOpName("add1"), sw_0.output_false, sw_2.output_false);
  Output add2 =
      ops::Add(root.WithOpName("add2"), sw_1.output_false, sw_2.output_false);
  ops::Merge m1(root.WithOpName("m1"), {add1, add2});

  Output add3 = ops::Add(root.WithOpName("add3"), add0, m1.output);

  std::unique_ptr<DeadnessAnalysis> result;
  TF_ASSERT_OK(AnalyzeDeadness(root.graph(), &result));

    %16 = "tf.AddV2"(%15, %arg10) : (tensor<1x2x2x3xi32>, tensor<i32>) -> tensor<1x2x2x3xi32>
    %17 = "tf.Div"(%cst_1, %arg9) : (tensor<f32>, tensor<f32>) -> tensor<f32>
    %18 = "tf.Round"(%17) : (tensor<f32>) -> tensor<f32>
    %19 = "tf.Cast"(%18) : (tensor<f32>) -> tensor<i32>
    %20 = "tf.AddV2"(%19, %arg10) : (tensor<i32>, tensor<i32>) -> tensor<i32>

    // CHECK: [[ACCUM_NEW:%.*]] = "tf.AddV2"([[ACCUM_MOMENTUM]], [[GRAD]])
    // CHECK: "tf.AssignVariableOp"([[ACCUM_HANDLE]], [[ACCUM_NEW]])
    // CHECK: [[GRAD_LR:%.*]] = "tf.Mul"([[GRAD]], [[LR]])
    // CHECK: [[MOMENTUM_LR:%.*]] = "tf.Mul"([[MOMENTUM]], [[LR]])
    // CHECK: [[ACCUM_NEW_MOMENTUM_LR:%.*]] = "tf.Mul"([[ACCUM_NEW]], [[MOMENTUM_LR]])
    // CHECK: [[DELTA:%.*]] = "tf.AddV2"([[GRAD_LR]], [[ACCUM_NEW_MOMENTUM_LR]])

// CHECK-NEXT: [[loop_counter:%.*]] = "tf.AddV2"(%arg2, [[cst_1]])
// CHECK-NEXT: [[weight:%.*]] = "tf.ReadVariableOp"(%arg5)
// CHECK-NEXT: [[mpy:%.*]] = "tf.MatMul"(%arg6, [[weight]])
// CHECK-NEXT: [[element_index:%.*]] = "tf.AddV2"(%arg3, [[cst_1]])
// CHECK-NEXT: [[bias:%.*]] = "tf.GatherV2"(%arg7, %arg3, [[cst_0]])
// CHECK-NEXT: [[res:%.*]] = "tf.AddV2"([[mpy]], [[bias]])

func rewriteValuedec64_OpMul64(v *Value) bool {
	v_1 := v.Args[1]
	v_0 := v.Args[0]
	b := v.Block
	typ := &b.Func.Config.Types
	// match: (Mul64 x y)
	// result: (Int64Make (Add32 <typ.UInt32> (Mul32 <typ.UInt32> (Int64Lo x) (Int64Hi y)) (Add32 <typ.UInt32> (Mul32 <typ.UInt32> (Int64Hi x) (Int64Lo y)) (Select0 <typ.UInt32> (Mul32uhilo (Int64Lo x) (Int64Lo y))))) (Select1 <typ.UInt32> (Mul32uhilo (Int64Lo x) (Int64Lo y))))
	for {
		x := v_0

  TF_Operation* feed3 = Placeholder(func_graph_, s_, "feed3");
  TF_Operation* add1 = Add(feed1, feed2, func_graph_, s_, "add1");
  TF_Operation* add2 = Add(add1, feed3, func_graph_, s_, "add2");
  Define(1, {add2}, {add1, feed3}, {add2}, {});

  // Use, run, and verify
  TF_Operation* two = ScalarConst(2, host_graph_, s_, "two");
  TF_Operation* func_feed = Placeholder(host_graph_, s_);