// CHECK-LABEL: test_batch_matmul_broadcast_to_einsum
  // CHECK: "tf.Einsum"(%arg0, %arg1) <{equation = "...mk,...kn->...mn"}> : (tensor<2x2x4xf32>, tensor<2x4x2xf32>) -> tensor<2x2x2xf32>
  %0 = "tf.BatchMatMul"(%arg0, %arg1) {adj_x = false, adj_y = false} : (tensor<2x2x4xf32>, tensor<2x4x2xf32>) -> tensor<2x2x2xf32>
  func.return %0: tensor<2x2x2xf32>
}

  func.return %emb : tensor<2x2xf32>

// CHECK-NEXT: tfl.pseudo_qconst
// CHECK-NEXT: tfl.gather
// CHECK-NEXT: tfl.dequantize
}

// CHECK-LABEL: no_pushdown_multiple_inputs
func.func @no_pushdown_multiple_inputs(%arg0: tensor<1000x1000x!quant.uniform<i8:f32, 7.812500e-03>>, %arg1: tensor<1000x1000xf32>) -> tensor<2000x1000xf32> {

func.func @lift_float_matmul(%arg0: tensor<1x12x12x512xf32>) -> (tensor<*xf32>, tensor<*xf32>) {
  %cst = "tf.Const"() {value = dense<0.000000e+00> : tensor<512x512xf32>} : () -> tensor<512x512xf32>
  %out_1 = "tf.MatMul"(%arg0, %cst) {
    device = "", transpose_a = false, transpose_b = false
  } : (tensor<1x12x12x512xf32>, tensor<512x512xf32>) -> tensor<*xf32>
  %out_2 = "tf.MatMul"(%arg0, %arg0) {
    device = "", transpose_a = false, transpose_b = true

    %0 = "tf.Div"(%arg1, %cst_0) : (tensor<2x2x2xf32>, tensor<f32>) -> tensor<2x2x2xf32>
    %1 = "tf.AddV2"(%0, %cst_1) : (tensor<2x2x2xf32>, tensor<f32>) -> tensor<2x2x2xf32>
    %2 = "tf.Floor"(%1) : (tensor<2x2x2xf32>) -> tensor<2x2x2xf32>
    %3 = "tf.ClipByValue"(%2, %cst_5, %cst_6) : (tensor<2x2x2xf32>, tensor<f32>, tensor<f32>) -> tensor<2x2x2xf32>
    %4 = "tf.Cast"(%3) {Truncate = false} : (tensor<2x2x2xf32>) -> tensor<2x2x2xi8>

  %outputs_1, %outputs_2 = "tf.IdentityN"(%outputs, %outputs_0) {device = "TPU"} : (tensor<2x2xf32>, tensor<2x2xf32>) -> (tensor<2x2xf32>, tensor<2x2xf32>)
  func.return %outputs_0, %outputs_1 : tensor<2x2xf32>, tensor<2x2xf32>
}

// CHECK-LABEL: func @keep_identity_n_test
func.func @keep_identity_n_test(%arg0: tensor<2x2xf32>, %arg1: tensor<2x2xf32>) -> (tensor<2x2xf32>, tensor<2x2xf32>) {
    // CHECK: tf.MatMul

    %0 = "tf.MatMul"(%arg0, %arg1) {attr_map = "0:transpose_a,1:transpose_b", device = "", transpose_a = false, transpose_b = false} : (tensor<2x2xf32>, tensor<2x2xf32>) -> tensor<2x2xf32>
    return %0 : tensor<2x2xf32>
  }

  // MLIR @main function corresponds to the TF function "main_original".
  OwningOpRef<ModuleOp> module_op = ParseModuleOpString(R"mlir(
    func.func private @main(%arg: tensor<1x2xf32>) -> (tensor<1x2xf32>) attributes {tf._original_func_name = "main_original"} {
      return %arg : tensor<1x2xf32>
    }
  )mlir");
  ASSERT_TRUE(module_op);

  absl::flat_hash_map<FunctionName, FunctionAlias> function_aliases;

    %1 = "tf.PartitionedCall"(%arg0, %cst_0) {_tfl_quant_trait = "fully_quantizable", config = "", config_proto = "", executor_type = "", f = @composite_matmul_fn_1} : (tensor<2x12xf32>, tensor<12x2xf32>) -> tensor<*xf32>
    func.return %1: tensor<*xf32>
  }

// CHECK: [[UNPACK:%.*]]:3 = "tf.Unpack"(%arg0) <{axis = 0 : i64}> : (tensor<3x2x2xf32>) -> (tensor<2x2xf32>, tensor<2x2xf32>, tensor<2x2xf32>)
// CHECK: [[SCALAR_ZERO:%.*]] = arith.constant dense<0> : tensor<i32>
// CHECK: [[CONCAT:%.*]] = "tf.Concat"([[SCALAR_ZERO]], [[UNPACK]]#0, [[UNPACK]]#1, [[UNPACK]]#2) : (tensor<i32>, tensor<2x2xf32>, tensor<2x2xf32>, tensor<2x2xf32>) -> tensor<?x2xf32>
// CHECK: [[LENGTHS:%.*]] = arith.constant dense<0> : tensor<0xi64>

    return %1 : tensor<1x3xf32>
  }

  func.func private @composite_dot_general_fn(%arg0: tensor<1x2xf32>, %arg1: tensor<2x3xf32>) -> tensor<1x3xf32> attributes {_from_xla_call_module} {
    %0 = stablehlo.dot_general %arg0, %arg1, contracting_dims = [1] x [0] : (tensor<1x2xf32>, tensor<2x3xf32>) -> tensor<1x3xf32>
    return %0 : tensor<1x3xf32>