// taking the most specialized. This combines `10x?x?` and `?x?x8` into
  // `10x?x8`.
  if (!lhs_shape_type.hasRank()) {
    if (rhs_shape_type.hasRank()) {
      shape.append(rhs_shape_type.getShape().begin(),
                   rhs_shape_type.getShape().end());
      refined_shape = true;
    }
  } else if (rhs_shape_type.hasRank()) {
    for (auto shape_elts : llvm::enumerate(

// 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() && "
          "$0.getType().cast<ShapedType>().getRank() == " # n>>;

    // Scalar identity is broadcastable to any operand shape, we only need to
    // check that operand has the same shape as a result.
    bool scalar_identity = identity_ty.hasRank() && identity_ty.getRank() == 0;
    if (scalar_identity) return operand_ty == result_ty;

    // If identity is not a scalar, we must verify that identity shape is

    first_operand_type =
        mlir::cast<TF::ResourceType>(element_type).getSubtypes().front();
  }

  auto tensor_type = mlir::dyn_cast_or_null<TensorType>(first_operand_type);
  if (!(tensor_type && tensor_type.hasRank())) {
    return op->emitError()
           << "cannot convert op with unranked or non-tensor input type "
           << tensor_type << ".";
  }

  int rank = tensor_type.getRank();
  if (rank <= partition_dim) {

    // Matches fail when lhs|rhs|cond is unranked tensor.
    // TODO(b/176202543): Support unranked tensor.
    if (!mlir::cast<ShapedType>(lhs.getType()).hasRank() ||
        !mlir::cast<ShapedType>(rhs.getType()).hasRank() ||
        !mlir::cast<ShapedType>(cond.getType()).hasRank()) {
      return failure();
    }

    // Calculates symbolic broadcast shape that is only used in types.

  if ((input_ty && input_ty.hasRank() && input_ty.getRank() != 4) ||
      (filter_ty && filter_ty.hasRank() && filter_ty.getRank() != 4)) {
    return emitOptionalError(location, "Invalid ranks");
  }

  // If either input or filter is unranked, we will just return unranked output
  // shape.
  if (!input_ty || !filter_ty || !input_ty.hasRank() || !filter_ty.hasRank()) {
    Type result_type;

  // If the updated type doesn't have a rank, then it can't be a more refined
  // type.
  if (!updated.hasRank()) return false;

  // If the original type doesn't have a rank, then refine as the updated type
  // has a rank.
  if (!original.hasRank()) return true;

  // Both types must have the same rank.
  if (original.getRank() != updated.getRank()) return false;

          "requires compatible element types for all operands and results");
    }
  }
  return success();
}

inline ShapedType MergeType(ShapedType a, ShapedType b) {
  if (!a.hasRank()) {
    return b;
  }
  if (!b.hasRank()) {
    return a;
  }
  int64_t rank = a.getRank();
  SmallVector<int64_t, 4> dims;
  dims.resize(rank);
  for (int i = 0, e = rank; i != e; i++) {

def ZeroIntAttr
  : AttrConstraint<CPred<"$_self.cast<::mlir::IntegerAttr>().getInt() == 0">>;

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

// Checks value is not produced by a TFL_Quant or
// from TFL_Quant Op with same quant type.
def NotFromQuantOpOrSameQuantType : Constraint<

ProcessTensorSpec(mlir::TensorType type) {
  tensorflow::DataType dtype;
  TF_RETURN_IF_ERROR(
      ConvertScalarTypeToDataType(type.getElementType(), &dtype));

  if (!type.hasRank())
    return std::make_pair(dtype, tensorflow::PartialTensorShape());

  auto shape = type.getShape();
  llvm::SmallVector<int64_t, 4> dims;
  dims.assign(shape.begin(), shape.end());