%cst = arith.constant dense<1> : tensor<1xi32>
  %0 = "tfl.exp"(%arg0) : (tensor<8x128xf32>) -> tensor<8x128xf32>
  %1 = "tfl.sum"(%0, %cst) {keep_dims = true} : (tensor<8x128xf32>, tensor<1xi32>) -> tensor<8x1xf32>
  %2 = "tfl.div"(%0, %1) {fused_activation_function = "NONE"} : (tensor<8x128xf32>, tensor<8x1xf32>) -> tensor<8x128xf32>
  func.return %2 : tensor<8x128xf32>

  func.return %0 : tensor<8xf32>
}

// -----

// CHECK-LABEL: strided_slice_grad_ellipsis_mask
// CHECK-SAME: [[GRAD:%.*]]: tensor<2x4x8xf32>
func.func @strided_slice_grad_ellipsis_mask(%grad: tensor<2x4x8xf32>) -> tensor<4x4x8xf32> {
  // Input to StridedSlice was of shape 4x4x8xf32
  // Strided slice gets input[2:4, ...]

// CHECK-SAME:                         %[[VAL_1:.*]]: tensor<8x8x8x64xf32>) -> tensor<?x3x3x64xf32> {
// CHECK:           %[[VAL_2:.*]] = "tf.Conv2D"(%[[VAL_0]], %[[VAL_1]]) <{data_format = "NHWC", dilations = [1, 1, 1, 1], explicit_paddings = [0, 0, 1, 1, 1, 1, 0, 0], padding = "EXPLICIT", strides = [1, 1, 1, 1], use_cudnn_on_gpu = true}> : (tensor<?x8x8x8xf32>, tensor<8x8x8x64xf32>) -> tensor<?x3x3x64xf32>