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)

  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));

    // 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]])

//
// is lowered to:
//
//   %sum0 = "tf.AddV2"(%0, %1)
//   %result = "tf.AddV2"(%sum0, %2)
//
// While
//
//   %result = "tf.AddN"(%0, %1, %2, %3, %4)
//
// is lowered to:
//
//   %sum0 = "tf.AddV2"(%0, %1)
//   %sum1 = "tf.AddV2"(%2, %3)
//   %sum2 = "tf.AddV2"(%sum0, %sum1)
//   %result = "tf.AddV2"(%sum2, %4)
//
class LowerAddNOp : public RewritePattern {
 public:

        then:
        from.empty
        collection.from.empty

        when:
        def add1 = from.add('a')
        def add2 = from.add('b')
        def add3 = from.add('a')

        then:
        add1
        add2
        !add3

        and:
        from as List == ['a', 'b']
        collection.from as List == ['a', 'b']

        when:

  ops::Switch switch_b(root.WithOpName("switch_b"), value, cond_b);

  Output add_a = ops::Add(root.WithOpName("add_a"), switch_a.output_true,
                          switch_b.output_true);
  Output add_b = ops::Add(root.WithOpName("add_b"), switch_a.output_true,
                          switch_b.output_true);
  Output add = ops::Add(root.WithOpName("add_c"), add_a, add_b);

  std::unique_ptr<Graph> graph(new Graph(OpRegistry::Global()));

  %0 = "tf.AddV2"(%arg0, %arg1) : (tensor<ui32>, tensor<ui32>) -> tensor<ui32>
  func.return %0 : tensor<ui32>

  // CHECK-LABEL: add_v2_uint32
  // CHECK:  %[[CAST:.*]] = "tf.Cast"(%arg0) <{Truncate = false}> : (tensor<ui32>) -> tensor<i32>
  // CHECK:  %[[CAST1:.*]] = "tf.Cast"(%arg1) <{Truncate = false}> : (tensor<ui32>) -> tensor<i32>

    %one:2 = tf_executor.island wraps "tf.Const"() { value = dense<1> : tensor<i32> } : () -> tensor<i32>
    %add2:2 = tf_executor.island wraps "tf.Add"(%arg0, %one#0) : (tensor<i32>, tensor<i32>) -> tensor<i32>
    tf_executor.fetch %add2#0, %arg1, %stack_push2#1 : tensor<i32>, tensor<f32>, !tf_executor.control
  }
  func.return %graph#0, %graph#1 : tensor<i32>, tensor<f32>
}