// CHECK:  %4 = "tfl.arg_min"(%arg0, %cst) : (tensor<2x2x4xf32>, tensor<1xi32>) -> tensor<2x2xi32>
  // CHECK:  return %3, %4 : tensor<2x2xf32>, tensor<2x2xi32>
}

// CHECK-LABEL:   func.func @convert_argmin_bool
func.func @convert_argmin_bool(%arg0: tensor<2xi1>) -> tensor<i32> {
  %0 = "mhlo.iota"() <{iota_dimension = 0 : i64}> : () -> tensor<2xi32>

      %2 = "tf.XlaCallModule"(%arg0, %1, %0) <{Sout = [#tf_type.shape<?x2>], module = "", version = 9 : i64}> {_entry_function = @composite_dot_general_fn_1, _original_entry_function = "composite_dot_general_fn_1", _tfl_quant_trait = "fully_quantizable", _quantization_method = "weight_only_ptq { }"} : (tensor<?x2xf32>, tensor<2x2xf32>, tensor<2xf32>) -> tensor<?x2xf32>
      return %2 : tensor<?x2xf32>

    %10 = "quantfork.qcast"(%9) {volatile} : (tensor<4x5xf32>) -> tensor<4x5x!quant.uniform<i8:f32, 1.000000e-03:-3>>
    %11 = "quantfork.dcast"(%10) : (tensor<4x5x!quant.uniform<i8:f32, 1.000000e-03:-3>>) -> tensor<4x5xf32>
    return %11 : tensor<4x5xf32>
  }

  // CHECK: quantized_dot_general_fn_1

  %0 = "tf.RealDiv"(%arg0, %arg0) : (tensor<2xi32>, tensor<2xi32>) -> tensor<2xi32>
  func.return %0: tensor<2xi32>
}

// CHECK-LABEL: func @sub
func.func @sub(%arg0: tensor<2xi32>) -> tensor<2xi32> {
  // CHECK-NEXT:  %0 = mhlo.subtract %arg0, %arg0 : tensor<2xi32>
  // CHECK-NEXT:  return %0 : tensor<2xi32>
  %0 = "tf.Sub"(%arg0, %arg0) : (tensor<2xi32>, tensor<2xi32>) -> tensor<2xi32>
  func.return %0: tensor<2xi32>
}

  %0 = "tf.Div"(%arg0, %cst) : (tensor<2x2xf32>, tensor<2x2xf32>) -> tensor<2x2xf32>
  func.return %0 : tensor<2x2xf32>

  // CHECK-LABEL: RemoveTrivialDiv
  // CHECK-NEXT: return %arg0 : tensor<2x2xf32>
}

func.func @RemoveTrivialRealDiv(%arg0: tensor<2x2xf32>, %arg1: tensor<2x2xf32>) -> tensor<2x2xf32> {
  %cst = arith.constant dense<1.0> : tensor<2x2xf32>

      return %1 : tensor<1x3xf32>
    }

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

}

func.func @main2(%arg0: tensor<2x4xf32>, %arg1: tensor<2x4xf32>) -> tensor<2x4xf32> {
  %0 = "tfl.quantize"(%arg0) {qtype = tensor<2x4x!quant.uniform<u8:f32, 0.49803921568627452>>} : (tensor<2x4xf32>) -> tensor<2x4x!quant.uniform<u8:f32, 0.49803921568627452>>
  %1 = "tfl.quantize"(%arg1) {qtype = tensor<2x4x!quant.uniform<u8:f32, 0.49803921568627452>>} : (tensor<2x4xf32>) -> tensor<2x4x!quant.uniform<u8:f32, 0.49803921568627452>>

// CHECK-LABEL: test_gather
func.func @test_gather(%arg0 : tensor<1x2x20xf32>, %arg1 : tensor<3x5xi32>) -> (tensor<1x3x5x20xf32>){
    // expected-remark@+1 {{Cost: 1}}
    %0 = "tf.Const"() { value = dense<[1]> : tensor<1xi32> } : () -> tensor<1xi32>
    // expected-remark@+1 {{Cost: 300}}
    %1 = "tf.GatherV2"(%arg0, %arg1, %0) : (tensor<1x2x20xf32>, tensor<3x5xi32>, tensor<1xi32>) -> tensor<1x3x5x20xf32>
    // expected-remark@+1 {{Cost: 40}}

        }) : () -> tensor<f32>
        "tf_device.cluster"() ({
          %group = "tf.Const"() <{value = dense<[[0, 1]]> : tensor<1x2xi32>}> : () -> tensor<1x2xi32>
          %arg1_reduced = "tf.XlaAllReduce"(%arg1_id, %group) <{mode = "CrossReplica", reduce_op = "Add"}> : (tensor<f32>, tensor<1x2xi32>) -> tensor<f32>
          "tf.OpA"(%arg1_reduced) : (tensor<f32>) -> ()
          tf_device.return
        }) : () -> ()

//
// Examples:
//   * If `xla_gather_op_output_type` == tensor<*xf32>, then it returns:
//     tensor<*xf32>.
//   * If `xla_gather_op_output_type` == tensor<3x5xi32> and `collapsed_dims` ==
//     {0}, then it returns: tensor<1x3x5xi32>.
//   * If `xla_gather_op_output_type` == tensor<3x5xf32> and `collapsed_dims` ==
//     {1, 3}, then it returns: tensor<3x1x5x1xf32>.
Type GetSliceOpOutputType(Type xla_gather_op_output_type,