// match: (Lsh64x8 x s)
	// result: (Int64Make (Or32 <typ.UInt32> (Or32 <typ.UInt32> (Lsh32x8 <typ.UInt32> (Int64Hi x) s) (Rsh32Ux8 <typ.UInt32> (Int64Lo x) (Sub8 <typ.UInt8> (Const8 <typ.UInt8> [32]) s))) (Lsh32x8 <typ.UInt32> (Int64Lo x) (Sub8 <typ.UInt8> s (Const8 <typ.UInt8> [32])))) (Lsh32x8 <typ.UInt32> (Int64Lo x) s))
	for {
		x := v_0
		s := v_1
		v.reset(OpInt64Make)
		v0 := b.NewValue0(v.Pos, OpOr32, typ.UInt32)

(Rsh8x64 x (Const64 [c])) && uint64(c) >= 8 => (SRAconst (SLLconst <typ.UInt32> x [24]) [31])

// constants
(Const(8|16|32) [val]) => (MOVWconst [int32(val)])
(Const(32|64)F [val]) => (MOV(F|D)const [float64(val)])
(ConstNil) => (MOVWconst [0])
(ConstBool [t]) => (MOVWconst [b2i32(t)])

// truncations
// Because we ignore high parts of registers, truncates are just copies.
(Trunc16to8 ...) => (Copy ...)

(ANDQ x (MOVQconst [c])) && is32Bit(c) => (ANDQconst [int32(c)] x)
(ANDL x (MOVLconst [c])) => (ANDLconst [c] x)

(AND(L|Q)const [c] (AND(L|Q)const [d] x)) => (AND(L|Q)const [c & d] x)
(XOR(L|Q)const [c] (XOR(L|Q)const [d] x)) => (XOR(L|Q)const [c ^ d] x)
(OR(L|Q)const  [c] (OR(L|Q)const  [d] x)) => (OR(L|Q)const  [c | d] x)

(MULLconst [c] (MULLconst [d] x)) => (MULLconst [c * d] x)

  Output const0 = ops::Const(root.WithOpName("const0"), 1);
  Output const1 = ops::Const(root.WithOpName("const1"), 2);

  Output add = ops::Add(root.WithOpName("add"), const0, const1);

  root.graph()->AddControlEdge(id0.node(), const0.node());
  root.graph()->AddControlEdge(id1.node(), const1.node());

  std::unique_ptr<DeadnessAnalysis> result;

    TF_Operation* const3;
    if (grad_inputs_provided) {
      const float const3_val[] = {1.0, 1.0, 1.0, 1.0};
      const3 = FloatConst2x2(expected_graph_, s_, const3_val, "GradInputs");
    } else {
      const3 = OnesLike(expected_graph_, s_, matmul, "gradients/OnesLike");
    }

    TF_Operation* matmul1 = MatMul(expected_graph_, s_, const3, const1,

  // done in parallel.
  //
  // This graph is:
  // (Const0, Const0) -> MatMul0
  // (Const1, Const1) -> MatMul1
  // (MatMul0, MatMul1) -> MatMulCombined
  //
  // Device0: [Const0, Const0, MatMul0]
  // Device1: [Const1, Const1, MatMul1, MatMulCombined]
  //
  // Cluster0: [Const0, Const0, MatMul0]
  // Cluster1: [Const1, Const1, MatMul1]
  // Cluster2: [MatMulCombined]

  // CHECK: %[[CONST1:.*]] = "tf.Const"() <{value = dense<1> : tensor<i32>}> : () -> tensor<i32>
  %const1 = "tf.Const"() {value = dense<1> : tensor<i32>} : () -> tensor<i32>
  // CHECK: %[[SUB:.*]] = "tf.Sub"(%[[BARG1]], %[[CONST1]])
  %sub = "tf.Sub"(%arg1, %const1) : (tensor<i32>, tensor<i32>) -> tensor<i32>
  %elem = "tf._SomeOp"() : () -> tensor<3xf32>

    %cst_1 = "tf.Const"() {value = dense<-5.000000e-01> : tensor<f32>} : () -> tensor<f32>
    %cst_2 = "tf.Const"() {value = dense<-1> : tensor<i32>} : () -> tensor<i32>
    %cst_3 = "tf.Const"() {value = dense<0.0274509806> : tensor<f32>} : () -> tensor<f32>
    %cst_4 = "tf.Const"() {value = dense<-55> : tensor<i32>} : () -> tensor<i32>
    %cst_5 = "tf.Const"() {value = dense<-1.280000e+02> : tensor<f32>} : () -> tensor<f32>

  // operations specified (0).
  void DefineT(int num_opers, const std::vector<TF_Operation*>& opers,
               const std::vector<TF_Output>& inputs,
               const std::vector<TF_Output>& outputs,
               const std::vector<string>& output_names,
               bool expect_failure = false) {
    ASSERT_EQ(func_, nullptr);
    const char** output_names_ptr = ToArray(output_names);

Status MatMulGradHelper(const Scope& scope, const bool is_batch,
                        const Output& x0, const bool adj_x0, const Output& x1,
                        const bool adj_x1, const DataType x_data_type,
                        const Output& y0, const bool adj_y0, const Output& y1,
                        const bool adj_y1, const DataType y_data_type,