}
    rewriter.replaceOp(op, new_results);
    return success();
  }
};

class TFConstOpQuantToIntPattern : public OpConversionPattern<TF::ConstOp> {
 public:
  using OpConversionPattern::OpConversionPattern;

  LogicalResult matchAndRewrite(
      TF::ConstOp op, TF::ConstOpAdaptor adaptor,
      ConversionPatternRewriter &rewriter) const override {
    if (!IsIllegalType(op.getOutput().getType())) return failure();

} >&2

# OS specific support (must be 'true' or 'false').
cygwin=false
msys=false
darwin=false
nonstop=false
case "\$( uname )" in                #(
  CYGWIN* )         cygwin=true  ;; #(
  Darwin* )         darwin=true  ;; #(
  MSYS* | MINGW* )  msys=true    ;; #(
  NONSTOP* )        nonstop=true ;;
esac

CLASSPATH=$classpath
<% if ( mainClassName.startsWith('--module ') ) { %>
MODULE_PATH=$modulePath

    if (isa<TF::VarHandleOp>(resource_operand.getDefiningOp()) &&
        isa<TF::ConstOp>(assigned_value_operand.getDefiningOp())) {
      return success();
    } else {
      return failure();
    }
  }

  void rewrite(TF::AssignVariableOp assign_op,
               PatternRewriter& rewriter) const override {
    // `TF::ConstOp` and `TF::VarHandleOp` are not manually erased.

    // CreateConstOpWithSingleValue. Also, make sure it works with QConst.
    return failure();
  }

  auto zero_bias = rewriter.create<TFL::ConstOp>(op->getLoc(), bias_attr);
  op->setOperand(bias_idx, zero_bias);

  return success();
}

TF::ConstOp PadConstValues(Operation* input_op, int value_to_pad,
                           int pad_dimensions, Location loc,
                           OpBuilder* builder) {

                 .getOperation();
           });

  // mlir::TF::ConstOp
  py::class_<mlir::TF::ConstOp>(m, "Tf_ConstOp")
      .def("create",
           [](mlir::OpBuilder& opb, mlir::Location loc,
              mlir::Attribute value) -> mlir::Operation* {
             return opb.create<mlir::TF::ConstOp>(loc, value).getOperation();
           });

  // mlir::TF::EqualOp

    auto begin = rewriter.create<TF::ConstOp>(slice_op.getLoc(),
                                              slice_op.getStartIndices());
    auto end = rewriter.create<TF::ConstOp>(slice_op.getLoc(),
                                            slice_op.getLimitIndices());
    auto strides =
        rewriter.create<TF::ConstOp>(slice_op.getLoc(), slice_op.getStrides());

        }
        operand = 0;
      }

      ShapedType type;
      float ratio_threshold = kBlockOverRandomSparsityRatio;
      if (isa<ConstOp>(inst)) {
        supported_block_size = sparse_op.GetFloatBlockSize();
        type = mlir::cast<ShapedType>(dyn_cast<ConstOp>(inst).getType());
      } else if (isa<QConstOp>(inst)) {
        supported_block_size = sparse_op.GetQuantizedBlockSize();

                   .Input(FakeInput(DT_INT32))
                   .Attr("T", DT_INT32)
                   .Device("/job:localhost/replica:0/task:0/device:XLA_CPU:0")
                   .Finalize(node_def()));
  TF_EXPECT_OK(InitOp());

  // Add inputs.
  Tensor* a = CreateDeviceTensor<int32>(TensorShape({1, 3}), {1, 2, 3});
  Tensor* b = CreateDeviceTensor<int32>(TensorShape({1, 3}), {4, 5, 6});
  inputs_.push_back({nullptr, a});

  }];

  let arguments = (ins
    StrAttr:$node_def,
    I32Attr:$op_key
  );

  let assemblyFormat = "attr-dict";
}

def ConstOp: TensorflowMlrt_Op<"constop", []> {
  let summary = "The tf_mlrt ConstOp";

  let description = [{
    The ConstOp creates a constant tensorflow::Tensor from serialized proto.
  }];

  let arguments = (ins
    StrAttr:$tensor_proto
  );

  let results = (outs

    return "";
  }

  auto assigned_value_operand = assign_var_op.getOperand(1);
  auto const_op =
      llvm::dyn_cast<mlir::TF::ConstOp>(assigned_value_operand.getDefiningOp());
  if (!const_op) {
    assign_var_op->emitRemark(
        "Operand idx 1 is not a tf.ConstOp. The initializing tensor is not "
        "saved to checkpoint.");
    return "";
  }

  Tensor const_tensor{};