tensor<1x1x5xf32>,
        tensor<2x5xf32>, tensor<2x5xf32>, tensor<2x5xf32>, tensor<2x5xf32>,
        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>,
        none, none, none, none) -> tensor<*xf32>

  %8 = "tfl.concatenation"(%2, %0) {axis = -1 : i32, fused_activation_function = "NONE"} : (tensor<1x1xf32>, tensor<1x1xf32>) -> tensor<1x2xf32>
  %9 = "quantfork.stats"(%8) {layerStats = dense<[-0.488159984, 0.189515018]> : tensor<2xf32>} : (tensor<1x2xf32>) -> tensor<1x2xf32>
  %10 = "tfl.concatenation"(%9, %7) {axis = -1 : i32, fused_activation_function = "NONE"} : (tensor<1x2xf32>, tensor<1x2xf32>) -> tensor<1x4xf32>

    return %2 : tensor<1x2xf32>
  }

  func.func private @composite_add_fn(%arg0: tensor<1x2xf32>, %arg1: tensor<1x2xf32>) -> tensor<1x2xf32> attributes {_from_xla_call_module} {
    %0 = stablehlo.add %arg0, %arg1 : tensor<1x2xf32>
    %1 = stablehlo.add %0, %arg1 : tensor<1x2xf32>
    return %1 : tensor<1x2xf32>
  }
}

// -----

  %5 = shape.shape_of %4 : tensor<?x10xf32> -> tensor<2xindex>
  %6 = stablehlo.dynamic_broadcast_in_dim %1, %5, dims = [1] : (tensor<10xf32>, tensor<2xindex>) -> tensor<?x10xf32>
  %7 = stablehlo.add %4, %6 : tensor<?x10xf32>
  %8 = stablehlo.clamp %2, %7, %3 : (tensor<f32>, tensor<?x10xf32>, tensor<f32>) -> tensor<?x10xf32>
  func.return %8: tensor<?x10xf32>
}

  // CHECK: %[[OLD_SLICE2:.*]] = "tf.Slice"(%[[READ2]],
  // CHECK: %[[RESHAPE2:.*]] = "tf.Reshape"(%[[VALUE]],
  // CHECK: %[[ADD2:.*]] = "tf.AddV2"(%[[RESHAPE2]], %[[OLD_SLICE2]]) : (tensor<1x3xf32>, tensor<1x3xf32>) -> tensor<1x3xf32>
  // CHECK: %[[UPDATE2:.*]] = "tf.XlaDynamicUpdateSlice"(%[[READ2]], %[[ADD2]],

  // CHECK-DAG: %[[EXP:.*]] = "tf.Exp"(%[[SHIFTED]]) : (tensor<2x3xf32>) -> tensor<2x3xf32>
  // CHECK-DAG: %[[SUM:.*]] = "tf.Sum"(%[[EXP]], %[[AXIS]]) <{keep_dims = true}> : (tensor<2x3xf32>, tensor<1xi64>) -> tensor<2x1xf32>
  // CHECK-DAG: %[[RESULT:.*]] = "tf.Div"(%[[EXP]], %[[SUM]]) : (tensor<2x3xf32>, tensor<2x1xf32>) -> tensor<2x3xf32>
  // CHECK: return %[[RESULT]]

  %4 = "tf.MatMul"(%arg0, %3) {device = "", transpose_a = false, transpose_b = false} : (tensor<2x3xf32>, tensor<3x4xf32>) -> tensor<2x4xf32>
  %5 = "tf.Identity"(%4) {device = ""} : (tensor<2x4xf32>) -> tensor<2x4xf32>
  %6 = "tf.Identity"(%5) {device = ""} : (tensor<2x4xf32>) -> tensor<2x4xf32>
  func.return %6 : tensor<2x4xf32>

    %10 = "tf.Cast"(%9) {Truncate = false, device = ""} : (tensor<1x3xi32>) -> tensor<1x3xf32>
    %11 = "tf.Mul"(%10, %cst) {device = ""} : (tensor<1x3xf32>, tensor<f32>) -> tensor<1x3xf32>
    %12 = "tf.Relu"(%11) {device = ""} : (tensor<1x3xf32>) -> tensor<1x3xf32>
    return %12 : tensor<1x3xf32>
  }
// CHECK-LABEL: func @matmul_with_relu

For example, if we have the code

```mlir
  %0 = "tf.Const"() {value = dense<[[42.0]]> : tensor<1x1xf32>} : () -> tensor<1x1xf32>
  %1 = "tf.Const"() {device = "", value = dense<[[42.0]]> : tensor<1x1xf32>} : () -> tensor<1x1xf32>
  %2 = "tf.Const"() {device = "baz", value = dense<[[42.0]]> : tensor<1x1xf32>} : () -> tensor<1x1xf32>
```

then running this pass with 'default-device=foobar', we get:

```mlir

//     -> tensor<5x2xf32>
//
// is lowered to
//
//   %shape = "tf.Const"() {value = dense<[-1, 2]> : tensor<2xi64>}
//   %inp0 = "tf.Reshape"(%arg0, %shape)
//     : (tensor<2xf32>, tensor<2xi64>) -> tensor<1x2xf32>
//   %inp1 = "tf.Reshape"(%arg1, %shape)
//     : (tensor<2x2x2xf32>, tensor<2xi64>) -> tensor<4x2xf32>
//   %items0 = "tf.Unpack"(%[[INP0]]) {axis = 0 : i64}