%4 = "stablehlo.concatenate"(%1, %3) {
      dimension = 0 : i64
    } : (tensor<3x2xf32>, tensor<1x2xf32>) -> tensor<4x2xf32>
    %5 = "quantfork.qcast"(%4) {volatile} : (tensor<4x2xf32>) -> tensor<4x2x!quant.uniform<i8:f32, 5.000000e-03:-1>>
    %6 = "quantfork.dcast"(%5) : (tensor<4x2x!quant.uniform<i8:f32, 5.000000e-03:-1>>) -> tensor<4x2xf32>

// CHECK-DAG: %[[input_19:.*]] = "quantfork.stats"({{.*}}) <{layerStats = dense<[-2.000000e+00, 4.000000e+00]> : tensor<2xf32>}> : (tensor<1x2xf32>) -> tensor<1x2xf32>

// CHECK: "tfl.unidirectional_sequence_lstm"({{.*}}, %[[input_18]], %[[input_19]], %{{[0-9]+}}, %{{[0-9]+}}, %{{[0-9]+}}, %{{[0-9]+}})

    // CHECK-NEXT:  %[[RES:.*]] = "tf.SelectV2"(%[[PRED]], %[[SCALED_GRADIENTS]], %[[SELU_GRAD_VALUE]]) : (tensor<4x8xi1>, tensor<4x8xf32>, tensor<4x8xf32>) -> tensor<4x8xf32>
    // CHECK-NEXT:  return %[[RES]] : tensor<4x8xf32>
    %2 = "tf.SeluGrad"(%gradients, %features) : (tensor<4x8xf32>, tensor<4x8xf32>) -> tensor<4x8xf32>
    func.return %2 : tensor<4x8xf32>
}

}

// CHECK-LABEL: QuantizeConcat
func.func @QuantizeConcat(tensor<1x2xf32>, tensor<1x2xf32>) -> tensor<2x2x!quant.uniform<u8:f32, 1.000000e-01:128>> {
^bb0(%arg0: tensor<1x2xf32>, %arg1: tensor<1x2xf32>):
  %0 = "tfl.concatenation"(%arg0, %arg1) {axis = 0 : i32, fused_activation_function = "NONE"} : (tensor<1x2xf32>, tensor<1x2xf32>) -> tensor<2x2xf32>

  // CHECK: %[[v0:.*]] = "tf.Reshape"(%arg1, %[[cst]]) : (tensor<5x6x2xf32>, tensor<2xi64>) -> tensor<5x12xf32>
  // CHECK: %[[v1:.*]] = "tf.BatchMatMulV2"(%arg0, %[[v0]]) <{adj_x = false, adj_y = false}> : (tensor<3x4x5xf32>, tensor<5x12xf32>) -> tensor<3x4x12xf32>
  // CHECK: %[[v2:.*]] = "tf.Reshape"(%[[v1]], %[[cst_1]]) : (tensor<3x4x12xf32>, tensor<4xi64>) -> tensor<3x4x6x2xf32>
  // CHECK: return %[[v2]] : tensor<3x4x6x2xf32>
}

  // CHECK-LABEL: @xla_svd
  func.func @xla_svd(%arg0: tensor<1x1xf32>) -> (tensor<1xf32>, tensor<1x1xf32>, tensor<1x1xf32>) {
    // CHECK-NOT: XlaSvd
    %s, %u, %v = "tf.XlaSvd"(%arg0) {max_iter = 1, epsilon = 1.0E-09 : f32, precision_config = ""} : (tensor<1x1xf32>) -> (tensor<1xf32>, tensor<1x1xf32>, tensor<1x1xf32>)
    func.return %s, %u, %v : tensor<1xf32>, tensor<1x1xf32>, tensor<1x1xf32>
  }

  func.func @identity(%arg0: f32) -> f32 {

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

      %2 = "tf.XlaCallModule"(%arg0, %1, %0) <{Sout = [#tf_type.shape<?x2>], module = "", version = 9 : i64}> {_original_entry_function = "composite_fn_1"} : (tensor<?x2xf32>, tensor<2x2xf32>, tensor<2xf32>) -> tensor<?x2xf32>
      return %2 : tensor<?x2xf32>
    }

    func.return %0 : tensor<4x128xf32>
  }
  func.func @_func(%arg0: tensor<4x128xf32>) -> tensor<4x128xf32> {
    func.return %arg0 : tensor<4x128xf32>
  }

@testLstmWithInvalidInputsRankMatch(%arg0: tensor<1x4xf32>, %arg1: tensor<4x2xf32>, %arg2: tensor<4x2xf32>, %arg3: tensor<4x2xf32>, %arg4: tensor<4x2xf32>, %arg5: tensor<4x4xf32>, %arg6: tensor<4x4xf32>, %arg7: tensor<4x4xf32>, %arg8: tensor<4x4xf32>, %arg9: tensor<4xf32>, %arg10: tensor<4xf32>, %arg11: tensor<4xf32>, %arg12: tensor<1x4xf32>, %arg13: tensor<4xf32>, %arg14: tensor<4xf32>, %arg15: tensor<4xf32>, %arg16: tensor<4x4xf32>, %arg17: tensor<4xf32>, %arg18: tensor<4xf32>, %arg19: tensor<4xf32>,...