// CHECK-DAG-SAME: strides = dense<1> : tensor<1xi64>} : (tensor<2xi64>) -> tensor<1xi64>
  // CHECK: %[[RESHAPED_START2:.*]] = mhlo.reshape %[[SLICED_START2]] : (tensor<1xi64>) -> tensor<i64>
  // CHECK: %[[RESULT:.*]] = "mhlo.dynamic_slice"(%arg0, %[[RESHAPED_START1]], %[[RESHAPED_START2]]) <{slice_sizes = dense<[1, 4]> : tensor<2xi64>}> : (tensor<3x4xi32>, tensor<i64>, tensor<i64>) -> tensor<1x4xi32>
  // CHECK: return %[[RESULT]] : tensor<1x4xi32>

      (field_name == "broadcast_dimensions" ||
       field_name == "known_expanding_dimensions" ||
       field_name == "known_nonexpanding_dimensions"))
    return true;
  if ((op_name == "stablehlo.dynamic_slice" || op_name == "stablehlo.gather") &&
      field_name == "slice_sizes")
    return true;
  if (op_name == "stablehlo.fft" && field_name == "fft_length") return true;

      // Dynamic start indices should be clamped with upper limit of
      // `shape(operand) - slice_sizes)` as per semantics of
      // `stablehlo.dynamic_slice`.
      // (https://github.com/openxla/stablehlo/blob/main/docs/spec.md#dynamic_slice)
      start_indices[i] =
          rewriter.create<TFL::MinimumOp>(op->getLoc(), cast, upper_limit_cst);
    }

    Value concatenated = start_indices[0];

  // CHECK:   mhlo.return [[CMP]]
  // CHECK: } do {
  // CHECK:   [[SRC_IDX:%.*]] = "mhlo.dynamic_slice"([[ITER_ARG2]], [[ITER_ARG0]]) <{slice_sizes = dense<1> : tensor<1xi64>}> : (tensor<4xi32>, tensor<i32>) -> tensor<1xi32>
  // CHECK:   [[SWP_IDX:%.*]] = "mhlo.dynamic_slice"([[ITER_ARG1]], [[ITER_ARG0]]) <{slice_sizes = dense<1> : tensor<1xi64>}> : (tensor<4xi32>, tensor<i32>) -> tensor<1xi32>

// CHECK-NOT: tfl.gather

// -----

// Tests that a quantized `stablehlo.dynamic_slice` is converted to `tfl.slice`.

// CHECK-LABEL: func @dynamic_slice
// CHECK-SAME: %[[ARG0:.+]]: tensor<4x4x!quant.uniform<i8:f32, 3.000000e-01:-5>>, %[[ARG1:.+]]: tensor<i64>, %[[ARG2:.+]]: tensor<i64>
func.func @dynamic_slice(
    %arg0: tensor<4x4x!quant.uniform<i8:f32, 3.000000e-01:-5>>,
    %arg1: tensor<i64>,

      testing.parameter_combinations([{
          'same_scale_op': (
              'reshape',  # This corresponds to stablehlo.dynamic_reshape
              'slice',  # This corresponds to stablehlo.dynamic_slice.
              # TODO: b/326242075 - Support other same-scale ops.
          ),
          'dim_sizes': (([None, 1024], [1024, 3]),),
      }])
  )
  @test_util.run_in_graph_and_eager_modes

  %ta0:2 = "tf.TensorArrayV3"(%size) {dtype = f32, element_shape = #tf_type.shape<3>, dynamic_size = false, clear_after_read = true, identical_element_shapes = true, tensor_array_name = "ta"} : (tensor<i32>) -> (tensor<!tf_type.resource>, tensor<f32>)

// CHECK:           return %[[VAL_14]] : tensor<4x2xf32>
// CHECK:         }
func.func @convert_dynamic_slice(%arg0: tensor<7x3xf32>, %arg1: tensor<i32>, %arg2: tensor<i32>) -> tensor<4x2xf32> {
  %0 = "mhlo.dynamic_slice"(%arg0, %arg1, %arg2) <{slice_sizes = dense<[4, 2]> : tensor<2xi64>}> : (tensor<7x3xf32>, tensor<i32>, tensor<i32>) -> tensor<4x2xf32>
  func.return %0 : tensor<4x2xf32>
}

  %handle_0, %flow_0 = "tf.TensorArrayV3"(%size) {clear_after_read = true, device = "/job:localhost/replica:0/task:0/device:CPU:0", dtype = f32, dynamic_size = false, element_shape = #tf_type.shape<?x100>, identical_element_shapes = true, tensor_array_name = "processed_embeddings/bidirectional_rnn/bw/bw/dynamic_rnn/input_0"} : (tensor<i32>) -> (tensor<2x!tf_type.resource<tensor<?x100xf32>>>, tensor<f32>)...

  // CHECK: tf.TensorArrayV3
  // CHECK-SAME: element_shape = #corert.shape<*>
  %ta:2 = "tf.TensorArrayV3"(%size) {device = "/device:CPU:0", dtype = f32, element_shape = #tf_type.shape<*>, dynamic_size = false, clear_after_read = true, identical_element_shapes = true, tensor_array_name = "ta"} : (tensor<i32>) -> (tensor<!tf_type.resource>, tensor<f32>)
  func.return
}

// CHECK-LABEL: func @resource