} : (tensor<10xf32>, tensor<10xf32>, tensor<6xi32>) -> (tensor<10xf32>, tensor<6xi32>)
    // CHECK: return %[[RESULT_CAST]], %[[RESULT]]#1 : tensor<10xf32>, tensor<6xi32>
    func.return %0#0, %0#1 : tensor<10xf32>, tensor<6xi32>
  }

  // CHECK-LABEL: func private @main_0
  // CHECK-SAME: %[[ARG_0:.*]]: tensor<10xbf16>, %[[ARG_1:.*]]: tensor<10xbf16>, %[[ARG_2:.*]]: tensor<6xi32>
  func.func private @main_0(

  %0 = "tf.Einsum"(%arg0, %arg1) {T = "tfdtype$DT_FLOAT", equation = "ijk,j->i"}: (tensor<4x5x6xf32>, tensor<5xf32>) -> tensor<4xf32>
  func.return %0 : tensor<4xf32>
// CHECK-LABEL: einsum_no_match
// CHECK: %[[v0:.*]] = "tf.Einsum"(%arg0, %arg1) <{equation = "ijk,j->i"}> {T = "tfdtype$DT_FLOAT"} : (tensor<4x5x6xf32>, tensor<5xf32>) -> tensor<4xf32>
// CHECK: return %[[v0]]
}

        -> !tf_res
      %read = "tf.ReadVariableOp"(%id2) : (!tf_res) -> tensor<32xf32>
      %id3 = "tf.Identity"(%read) : (tensor<32xf32>) -> tensor<32xf32>
      tf_executor.yield
    }
    tf_executor.fetch %island : !tf_executor.control
  }
  func.return
}

// -----
!tf_res = tensor<*x!tf_type.resource<tensor<32xf32>>>

// Tesets that the pass can propagate through tf.If's branches.

    sym_name = "some_constant",
    type = tensor<1x64xf32>,
    value = dense<42.0> : tensor<1x64xf32>
  } : () -> ()

  // Representation for variables: mutable global tensor.
  // CHECK: tf_saved_model.global_tensor
  "tf_saved_model.global_tensor"() {
    is_mutable,
    tf_saved_model.exported_names = ["some_var", "some.other.name"],
    sym_name = "some_variable",
    type = tensor<?x64xf32>,
    value = dense<42.0> : tensor<1x64xf32>

    %3 = "tf.Mul"(%1, %2) : (tensor<10xf32>, tensor<10xf32>) -> tensor<10xf32>
    return %3 : tensor<10xf32>
  }
}

// -----

// CHECK-LABEL: module
module attributes {tf_saved_model.semantics} {
  // CHECK: ml_program.global{{.*}}mutable{{.*}}@vars.v
  "tf_saved_model.global_tensor"() { is_mutable, sym_name = "v", type = tensor<10xf32>, value = dense<[0.,10.,2.,3.,4.,5.,6.,7.,8.,9.]> : tensor<10xf32> } : () -> ()

    func.return %0 : tensor<4xf32>
  }

  // CHECK-LABEL: @ndtri
  func.func @ndtri(%input: tensor<4xf32>) -> tensor<4xf32> {
    // CHECK-NOT: tf.Ndtri
    %0 = "tf.Ndtri"(%input) : (tensor<4xf32>) -> tensor<4xf32>
    func.return %0 : tensor<4xf32>
  }

  // CHECK-LABEL: @fake_param
  func.func @fake_param() -> tensor<4xf32> {
    // CHECK-NOT: tf.FakeParam

    %assign1_control = tf_executor.island wraps "tf.AssignVariableOp"(%arg1, %read0) : (tensor<*x!tf_type.resource<tensor<32xf32>>>, tensor<32xf32>) -> ()
    %assign2_control = tf_executor.island wraps "tf.AssignVariableOp"(%arg0, %read2) : (tensor<*x!tf_type.resource<tensor<32xf32>>>, tensor<32xf32>) -> ()

  attributes {tf_saved_model.exported_names = ["assign_from_global"]} {
    // CHECK: [[name:%.*]] = "tf.VarHandleOp"
    // CHECK: "tf.AssignVariableOp"([[name]]
    %0 = "tf.Const"() {type = tensor<10xf32>, value = dense<[0.,10.,2.,3.,4.,5.,6.,7.,8.,9.]> : tensor<10xf32>} : () -> tensor<10xf32>
    "tf.AssignVariableOp"(%arg0, %0) : (tensor<!tf_type.resource<tensor<10xf32>>>, tensor<10xf32>) -> ()

  %5 = "stablehlo.concatenate"(%4, %1) {dimension = 0 : i64} : (tensor<?xf32>,  tensor<?xf32>) -> tensor<?xf32>
  %6 = "quantfork.stats"(%5) {layerStats = dense<[-1.5726943, 4.6875381]> : tensor<2xf32>} : (tensor<?xf32>) -> tensor<?xf32>
  func.return  %3, %6 : tensor<?xf32>, tensor<?xf32>
}

// -----

        tensor<2x4xf32>, tensor<2x4xf32>, tensor<2x4xf32>, tensor<2x4xf32>,
        tensor<2xf32>, tensor<2xf32>, tensor<2xf32>,
        tensor<2xf32>, tensor<2xf32>, tensor<2xf32>, tensor<2xf32>,
        tensor<4x2xf32>, tensor<4xf32>,
        tensor<1x4xf32>, tensor<1x2xf32>,
        tensor<2xf32>, tensor<2xf32>, tensor<2xf32>, tensor<2xf32>) -> tensor<*xf32>