auto input = block->getArgument(0);
    auto size = block->getArgument(3);

    // Subtract `input_rank` by 1 to get the item's rank, which is used as
    // `partial_position_shape`.
    auto input_rank = rewriter->create<TF::RankOp>(
        loc, tensorflow::GetTypeFromTFTensorShape({}, shape_dtype), input);
    auto partial_position_shape = rewriter->create<TF::SubOp>(

type byEvictionPriority []evictionapi.Threshold

func (a byEvictionPriority) Len() int      { return len(a) }
func (a byEvictionPriority) Swap(i, j int) { a[i], a[j] = a[j], a[i] }

// Less ranks memory before all other resources, and ranks thresholds with no resource to reclaim last
func (a byEvictionPriority) Less(i, j int) bool {
	_, jSignalHasResource := signalToResource[a[j].Signal]

    "ExtractSingleElementAsFloat($_self.cast<ElementsAttr>())">;

// Checks if the value has rank at most 'n'.
class HasRankAtMost<int n> : Constraint<
    CPred<"$0.getType().cast<ShapedType>().hasRank() && "
          "$0.getType().cast<ShapedType>().getRank() <= " # n>>;

// Checks if the value has rank 'n'.
class HasRank<int n> : Constraint<
    CPred<"$0.getType().cast<ShapedType>().hasRank() && "

  // Reduce along all dimensions except the feature dimension.
  int64_t feature_dim = GetTensorFeatureDimIndex(rank, format);
  llvm::SmallVector<int64_t, 4> dims_to_reduce(rank - 1);
  std::iota(dims_to_reduce.begin(), dims_to_reduce.begin() + feature_dim, 0);
  std::iota(dims_to_reduce.begin() + feature_dim, dims_to_reduce.end(),
            feature_dim + 1);
  return GetI64ElementsAttr(dims_to_reduce, builder);
}

  Scope cpu_scope = scope.WithDevice("/cpu:0");

  // [3]
  auto rank = Rank(cpu_scope, op.input(0));

  // Normalize any negative indices in the reduction_axes to positive values.
  auto reduction_indices_pos =
      Mod(cpu_scope, Add(cpu_scope, reduction_indices, rank), rank);

  // [1]
  auto reduced = Cast(cpu_scope, reduction_indices_pos, DataType::DT_INT32);

  // [0, 1, 2]

```
  }];

  let arguments = (ins
    TF_BoolTensor:$condition,
    Arg<TF_Tensor, [{= A `Tensor` which may have the same shape as `condition`.
If `condition` is rank 1, `x` may have higher rank,
but its first dimension must match the size of `condition`.}]>:$then_value,
    Arg<TF_Tensor, [{= A `Tensor` with the same type and shape as `x`.}]>:$else_value
  );

  let results = (outs

        !filter_type.hasStaticShape())
      return failure();

    Value input = tf_op.getInput();
    RankedTensorType input_type =
        mlir::dyn_cast<RankedTensorType>(input.getType());
    // Only rank size four input will be only available by the tf.Conv2D
    // operator verification.
    if (!input_type || input_type.isDynamicDim(3)) {
      return failure();
    }

                                              const char* attr_name,
                                              const char* value, size_t length);

// Set `num_dims` to -1 to represent "unknown rank".  Otherwise,
// `dims` points to an array of length `num_dims`.  `dims[i]` must be
// >= -1, with -1 meaning "unknown dimension".
TF_CAPI_EXPORT extern void TF_SetAttrShape(TF_OperationDescription* desc,

  ASSERT_EQ(TF_OK, TF_GetCode(s)) << TF_Message(s);
  EXPECT_EQ(dims[0], returned_dims[0]);
  EXPECT_EQ(dims[1], returned_dims[1]);

  // Try to set 'unknown' with unknown rank on the shape and see that
  // it doesn't change.
  TF_GraphSetTensorShape(graph, feed_out_0, /*dims=*/nullptr, -1, s);
  EXPECT_EQ(TF_OK, TF_GetCode(s)) << TF_Message(s);

  %5 = "quantfork.stats"(%4) {layerStats = dense<[-56.2916565, 122.922478]> : tensor<2xf32>} : (tensor<1x4xf32>) -> tensor<1x4xf32>
  %6 = "tfl.svdf"(%0, %1, %2, %3, %5) {fused_activation_function = "RELU", rank = 1 : i32} : (tensor<1x3xf32>, tensor<2x3xf32>, tensor<2x1xf32>, tensor<2xf32>, tensor<1x4xf32>) -> tensor<1x2xf32>