%split_dim_2 = arith.constant dense<1> : tensor<1xi32>
  %4, %5 = "tfl.split"(%split_dim_2, %arg0) {num_splits = 2 : i32} : (tensor<1xi32>, tensor<16x4xf32>) -> (tensor<16x2xf32>, tensor<16x2xf32>)
  %6:2 = "tfl.split"(%split_dim_2, %arg0) {num_splits = 2 : i32} : (tensor<1xi32>, tensor<16x4xf32>) -> (tensor<16x2xf32>, tensor<16x?xf32>)

// dimensions into a single dimension. For example,
//
//   %shape = arith.constant dense<[1, 128, 64]> : tensor<3xi32>
//   %reshape = tfl.reshape(%input, %shape) // %input: tensor<128x64xf32>
//   %fc = tfl.fully_connected(%reshape, %filter, %bias)
//           {keep_num_dims = false, weights_format = "DEFAULT"}
//
// can be canonicalized to
//

// -----

// Verify that the batch_dims can be equal to the rank of the indices.
func.func @testGatherV2(%arg0: tensor<16x4xf32>, %arg1: tensor<16xi32>) -> tensor<16xf32> {
  %0 = "tf.Const"() { value = dense<[1]> : tensor<1xi32> } : () -> tensor<1xi32>
  %1 = "tf.GatherV2"(%arg0, %arg1, %0) {batch_dims = 1 : i64} : (tensor<16x4xf32>, tensor<16xi32>, tensor<1xi32>) -> tensor<16xf32>
  func.return %1 : tensor<16xf32>
}

// -----