%0 = "tfl.add"(%arg0, %arg1) {fused_activation_function = "RELU6",  per_device_costs = {CPU = 5.0 : f32, GPU = 1.0 : f32}, tac.device = "GPU"} : (tensor<1xf32>, tensor<1xf32>) -> tensor<1xf32>
  %1 = "tfl.mul"(%0, %arg2) {fused_activation_function = "RELU6", per_device_costs = {CPU = 5.0 : f32, GPU = 1.0 : f32}, tac.device = "GPU"} : (tensor<1xf32>, tensor<1xf32>) -> tensor<1xf32>

(Mod32u x y) && buildcfg.GOPPC64 <= 8 => (SUB x (MULLW y (DIVWU x y)))

// (x + y) / 2 with x>=y => (x - y) / 2 + y
(Avg64u <t> x y) => (ADD (SRDconst <t> (SUB <t> x y) [1]) y)

(Mul64 ...) => (MULLD ...)
(Mul(32|16|8) ...) => (MULLW ...)
(Select0 (Mul64uhilo x y)) => (MULHDU x y)
(Select1 (Mul64uhilo x y)) => (MULLD x y)

(Div64 [false] x y) => (DIVD x y)
(Div64u ...) => (DIVDU ...)
(Div32 [false] x y) => (DIVW x y)

			loopHead := s.f.NewBlock(ssa.BlockPlain)
			loopBody := s.f.NewBlock(ssa.BlockPlain)

			// Pick right size ops.
			var mul, and, add, zext ssa.Op
			if s.config.PtrSize == 4 {
				mul = ssa.OpMul32
				and = ssa.OpAnd32
				add = ssa.OpAdd32
				zext = ssa.OpCopy
			} else {
				mul = ssa.OpMul64
				and = ssa.OpAnd64
				add = ssa.OpAdd64
				zext = ssa.OpZeroExt32to64
			}

        %add_out, %add_control = tf_executor.island wraps "tf.Add"(%arg2, %arg3) : (tensor<f32>, tensor<f32>) -> tensor<f32>
        %mul_out, %mul_control = tf_executor.island wraps "tf.Mul"(%arg2, %arg3) : (tensor<f32>, tensor<f32>) -> tensor<f32>
        %barrier = tf_executor.island(%assign_0_control, %assign_1_control, %add_control, %mul_control) wraps "tf.NoOp"() : () -> ()

  }];

  let arguments = (ins
    TF_FpOrI32OrI64Tensor:$input,

    I64Attr:$group_size,
    I64Attr:$group_key,
    I64Attr:$instance_key,
    TF_AnyStrAttrOf<["Min", "Max", "Mul", "Add"]>:$merge_op,
    TF_AnyStrAttrOf<["Id", "Div"]>:$final_op,
    I64ArrayAttr:$subdiv_offsets,
    DefaultValuedOptionalAttr<I64ArrayAttr, "{}">:$wait_for,

    *   Unary GPU kernels: Abs, Atanh, Acos, Acosh, Asin, Asinh, Atan, Cos, Cosh, Sin, Sinh, Tan, Tanh.
    *   Binary GPU kernels: AddV2, Sub, Div, DivNoNan, Mul, MulNoNan, FloorDiv, Equal, NotEqual, Greater, GreaterEqual, LessEqual, Less.

* `tf.lite`
    * Add experimental supports conversion of models that may be larger than 2GB before buffer deduplication

			expected: []metav1.TableRow{{Cells: []interface{}{"test5", "LoadBalancer", "10.9.8.7", "3.4.5.6,80.11.12.10", "8888/tcp", "<unknown>"}}},
		},
		{
			// Test LoadBalancer service with mul ExternalIPs
			service: api.Service{
				ObjectMeta: metav1.ObjectMeta{Name: "test6"},
				Spec: api.ServiceSpec{
					Type: api.ServiceTypeLoadBalancer,
					Ports: []api.ServicePort{
						{

    StringRef reduce_op = op.getReduceOp();
    if (reduce_op == "Add") {
      BuildReduceBody<AddOp>(element_type, &reduce_scatter.getComputation(),
                             &rewriter);
    } else if (reduce_op == "Mul") {
      BuildReduceBody<MulOp>(element_type, &reduce_scatter.getComputation(),
                             &rewriter);
    } else if (reduce_op == "Min") {