mean_op.getAxis(), mean_op.getKeepDims());

  // Insert a requant op.
  rewriter.replaceOpWithNewOp<TFL::QuantizeOp>(
      mean_op, output_type, new_mean_op, mlir::TypeAttr::get(output_type));
  return success();
}

}  // namespace tac
}  // namespace TFL

//			reg = reg (REG_*)
//			index = index (REG_*)
//			scale = scale (1, 2, 4, 8)
//
//	$<mem>
//		Effective address of memory reference <mem>, defined above.
//		Encoding: same as memory reference, but type = TYPE_ADDR.
//
//	$<±integer value>
//		This is a special case of $<mem>, in which only ±offset is present.
//		It has a separate type for easy recognition.
//		Encoding:
//			type = TYPE_CONST
//			offset = ±integer value

		// Make sure 1 prints as 1.0
		if !strings.ContainsAny(str, ".e") {
			str += ".0"
		}
		fmt.Fprintf(w, "$(%s)", str)

	case TYPE_SCONST:
		fmt.Fprintf(w, "$%q", a.Val.(string))

	case TYPE_ADDR:
		io.WriteString(w, "$")
		a.writeNameTo(w, abiDetail)

	case TYPE_SHIFT:
		v := int(a.Offset)
		ops := "<<>>->@>"
		switch buildcfg.GOARCH {
		case "arm":
			op := ops[((v>>5)&3)<<1:]

  int64_t out = 0;
  for (int64_t outer = 0; outer < outer_size; ++outer) {
    for (auto op : operands) {
      auto typed_attr = op.cast<TypedAttr>();
      const int64_t dim_size =
          typed_attr.getType().cast<RankedTensorType>().getDimSize(axis);
      const int64_t inner_size = dim_size * base_inner_size;

  return builder.getI32IntegerAttr(value);
}

static mlir::Attribute BuildTypeAttr(tflite::TensorType value,
                                     mlir::Builder builder) {
  return mlir::TypeAttr::get(ConvertElementType(value, builder));
}

static mlir::Attribute BuildTFL_AFAttr(tflite::ActivationFunctionType value,
                                       mlir::Builder builder) {

    // types. But we need to wrap it up with ElementsAttr to construct
    // valid TensorFlow constants.
    auto typed_attr = mlir::cast<TypedAttr>(value);
    type = tensorflow::GetTypeFromTFTensorShape(/*shape=*/{},
                                                typed_attr.getType());
    return ConstOp::build(builder, result, DenseElementsAttr::get(type, value));
  }

// requires relocation.
func markRelocs(p *obj.Prog) {
	switch p.As {
	case AMOV, AMOVB, AMOVH, AMOVW, AMOVBU, AMOVHU, AMOVWU, AMOVF, AMOVD:
		switch {
		case p.From.Type == obj.TYPE_ADDR && p.To.Type == obj.TYPE_REG:
			switch p.From.Name {
			case obj.NAME_EXTERN, obj.NAME_STATIC:
				p.Mark |= NEED_PCREL_ITYPE_RELOC
			}
		case p.From.Type == obj.TYPE_MEM && p.To.Type == obj.TYPE_REG:

    // Const op can have a result of dynamic shaped type (e.g. due to constant
    // folding), but we can still derive the shape of a constant tensor for
    // its attribute type.
    auto tensor_attr = mlir::cast<mlir::TypedAttr>(inst->getAttr("value"));
    llvm::ArrayRef<int64_t> shape_ref =
        mlir::cast<TensorType>(tensor_attr.getType()).getShape();
    if (mlir::failed(check_shape(shape_ref))) return std::nullopt;

    TypeAttr:$key_dtype,
    TypeAttr:$value_dtype
  );

  let results = (outs
    Res<TF_ResourceTensor, [{Handle to a table.}], [TF_LookupTableAlloc]>:$table_handle
  );

  let builders = [
    OpBuilder<(ins "StringAttr":$container, "StringAttr":$shared_name,
      "BoolAttr":$use_node_name_sharing, "TypeAttr":$key_dtype, "TypeAttr":$value_dtype),
    [{

				return C_SOREG
			}
			return C_LOREG
		}

		return C_GOK

	case obj.TYPE_TEXTSIZE:
		return C_TEXTSIZE

	case obj.TYPE_CONST,
		obj.TYPE_ADDR:
		switch a.Name {
		case obj.NAME_NONE:
			c.instoffset = a.Offset
			if a.Reg != obj.REG_NONE {
				if -BIG <= c.instoffset && c.instoffset <= BIG {
					return C_SACON
				}