// Check complex types.
  if (auto complex_type = element_type.dyn_cast<mlir::ComplexType>()) {
    auto complex_element_type = complex_type.getElementType();
    if (complex_element_type.isF32() || complex_element_type.isF64())
      return true;
  }
  // Check quantized types.
  if (auto quant_type = element_type.dyn_cast<mlir::quant::QuantizedType>()) {
    // TFLite supports QI16, QI32, QI8, and QUI8

  } else if (element_type.isF32()) {
    return DenseElementsAttr::get<float>(shaped_type,
                                         static_cast<float>(value));
  } else if (auto complex_type =
                 mlir::dyn_cast<mlir::ComplexType>(element_type)) {
    auto etype = complex_type.getElementType();
    if (etype.isF32()) {
      tensorflow::TensorProto repr;

  if (!image_type || !image_type.getElementType().isF32() ||
      image_type.getRank() != 4) {
    return func_.emitWarning() << "Image should be a 4D float tensor";
  }

  auto flow_type = mlir::dyn_cast_or_null<RankedTensorType>(
      func_.getFunctionType().getInput(1));
  if (!flow_type || !flow_type.getElementType().isF32() ||
      flow_type.getRank() != 4) {

    if (HasVariantInputOrOutput(op)) {
      std::optional<mlir::Type> element_type =
          GetSingularVariantBaseType(op->getOperand(0));
      if (element_type.has_value()) {
        return element_type->isF32() || element_type->isInteger(32);
      }
    }
  }

  // Op is vacuously "supported" if it is not a tensorlist op.
  StringRef op_name = op->getName().getStringRef();

      CreateI8F32UniformQuantizedType(UnknownLoc::get(&ctx_), ctx_,
                                      /*scale=*/1.0, /*zero_point=*/0);

  EXPECT_TRUE(quantized_type.getExpressedType().isF32());
}

TEST_F(CreateI8F32UniformQuantizedTypeTest, SignedQuantizedTypeSucceeds) {
  const UniformQuantizedType quantized_type =
      CreateI8F32UniformQuantizedType(UnknownLoc::get(&ctx_), ctx_,

  }
}

Status ConvertScalarTypeToDataType(Type type, DataType* dtype) {
  if (type.isF16()) {
    *dtype = DT_HALF;
    return absl::OkStatus();
  } else if (type.isF32()) {
    *dtype = DT_FLOAT;
    return absl::OkStatus();
  } else if (type.isF64()) {
    *dtype = DT_DOUBLE;
    return absl::OkStatus();
  } else if (type.isBF16()) {
    *dtype = DT_BFLOAT16;

    return tflite::TensorType_BFLOAT16;
  } else if (type.isF32()) {
    return tflite::TensorType_FLOAT32;
  } else if (type.isF64()) {
    return tflite::TensorType_FLOAT64;
  } else if (mlir::isa<mlir::TF::StringType>(type)) {
    return tflite::TensorType_STRING;
  } else if (auto complex_type = mlir::dyn_cast<mlir::ComplexType>(type)) {
    if (complex_type.getElementType().isF32()) {
      return tflite::TensorType_COMPLEX64;

def QuantizeByQuantizedType : NativeCodeCall<"quant::Quantize($0, $1.getValue())">;
def F32ElementsAttr : ElementsAttrBase<
  CPred<"$_self.cast<ElementsAttr>().getShapedType().getElementType().isF32()">, "float constant tensor">;

// Squash tfl.dequantize and tfl.quantize pairs.
// TODO(fengliuai): Compare the scale of input and output. This can also be
// squashed to a requantize op if the scales are different.

    return DenseElementsAttr::get(scalar_ty, attr);
  } else if (auto complex_ty = mlir::dyn_cast<ComplexType>(ty)) {
    Type complex_element_ty = complex_ty.getElementType();
    if (complex_element_ty.isF32()) {
      return DenseElementsAttr::get(
          scalar_ty, static_cast<std::complex<float>>(raw_value));
    } else if (complex_element_ty.isF64()) {
      return DenseElementsAttr::get(

        if (auto dq_op =
                dyn_cast_or_null<DequantizeOpT>(operand.getDefiningOp())) {
          inputs.push_back(dq_op.getOperand());
        } else if (!ele_type.isF32()) {
          // If the operand is an integer tensor, then it doesn't require the
          // DequantizeOp in the pattern.
          inputs.push_back(operand);
        } else if (weight_only_quantizable) {