// CHECK-LABEL: fuseTileWithBinaryOp
func.func @fuseTileWithBinaryOp(%arg0: tensor<1x1xf32>) -> tensor<1x2xf32> {
  %cst = arith.constant dense<[[1,2]]> : tensor<1x2xi32>
  %cst1 = arith.constant dense<[[3.0, 4.0]]> : tensor<1x2xf32>
  %0 = "tfl.sqrt"(%arg0) : (tensor<1x1xf32>) -> tensor<1x1xf32>
  %1 = "tfl.tile"(%0, %cst) : (tensor<1x1xf32>, tensor<1x2xi32>) -> tensor<1x2xf32>

    %17 = "tf.Mul"(%16, %cst_0) : (tensor<2x2xf32>, tensor<f32>) -> tensor<2x2xf32>
    %18 = "tf.AddV2"(%17, %cst) : (tensor<2x2xf32>, tensor<f32>) -> tensor<2x2xf32>
    %19 = "tf.Floor"(%18) : (tensor<2x2xf32>) -> tensor<2x2xf32>
    %20 = "tf.ClipByValue"(%19, %cst_5, %cst_6) : (tensor<2x2xf32>, tensor<f32>, tensor<f32>) -> tensor<2x2xf32>

  // CHECK:  %5 = "tfl.arg_max"(%arg0, %cst) : (tensor<1x32x1xf32>, tensor<1xi32>) -> tensor<1x1xi32>
  // CHECK:  return %4, %5 : tensor<1x1xf32>, tensor<1x1xi32>
}

// CHECK-LABEL:   func @convert_pytorch_argmax
func.func @convert_pytorch_argmax(%arg0: tensor<1x9xi32>) -> tensor<1xi32> {
  %0 = mhlo.constant dense<0> : tensor<i32>

    func.return %1 : tensor<1x1xi32>
  }
  func.func @predict(
    ) -> (tensor<0x!tf_type.string> {tf_saved_model.index_path = ["r"]})
      attributes {
        tf.entry_function = {control_outputs = "", inputs = "input:0", outputs = "result:0"},
        tf_saved_model.exported_names = ["predict"]
    }
  {

}

func.func @addN(%arg0: tensor<2x3xi32>, %arg1: tensor<2x3xi32>, %arg2: tensor<2x3xi32>) -> tensor<2x3xi32> {
  %0 = "tf.AddN"(%arg0, %arg1, %arg2) : (tensor<2x3xi32>, tensor<2x3xi32>, tensor<2x3xi32>) -> tensor<2x3xi32>
  func.return %0 : tensor<2x3xi32>

// CHECK-LABEL: addN
// CHECK:  "tfl.add_n"(%arg0, %arg1, %arg2) : (tensor<2x3xi32>, tensor<2x3xi32>, tensor<2x3xi32>) -> tensor<2x3xi32>
// CHECK:  return
}

  // Just to introduce extra usages for %cst.
  %3 = "tf.Mul"(%cst, %cst) {device = ""} : (tensor<1x3xf32>, tensor<1x3xf32>) -> tensor<1x3xf32>

  return %2: tensor<1x3xi32>
}
// The output of tf.Shape is considered a compile-time constant, so the
// constant leading to tf.Shape (which transitively becomes an input to the
// first arg of tf.Fill) is not duplicated.

  %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>

}

func.func @QDQsFollowedByTranspose(tensor<1x2xf32>) -> (tensor<2x1xf32>) {
^bb0(%arg0: tensor<1x2xf32>):
  %cst_0 = arith.constant dense<[1, 0]> : tensor<2xi32>
  %0 = "tfl.quantize"(%arg0){qtype = tensor<1x2x!quant.uniform<u8:f32, 1.0>>}: (tensor<1x2xf32>) -> (tensor<1x2x!quant.uniform<u8:f32, 1.0>>)
  %1 = "tfl.dequantize"(%0): (tensor<1x2x!quant.uniform<u8:f32, 1.0>>) -> (tensor<1x2xf32>)

    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>
  }
}

// -----

    %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>