mlir::isa<tf_type::TensorFlowRefType>(output_tensor_ty.getElementType());
  for (Type operand_type : merge.getOperandTypes()) {
    if (mlir::isa<ControlType>(operand_type)) break;

    // TODO(hinsu): Update ControlOperandsAfterAllData trait to verify this
    // constraint.
    TensorType operand_tensor_ty = mlir::dyn_cast<TensorType>(operand_type);
    if (!operand_tensor_ty)
      return merge.emitOpError()

template <typename OpT, typename std::enable_if<llvm::is_one_of<
                            OpT, ConcatOp, ConcatV2Op>::value>::type* = nullptr>
static LogicalResult Verify(OpT op) {
  // TODO(hinsu): Convert variadic length attributes to derived attributes.
  Operation::operand_range values = op.getValues();

  int axis_idx = std::is_same<OpT, ConcatOp>() ? 0 : 1;
  Value axis = *op.getODSOperands(axis_idx).begin();

}

// Broadcasts `input` to the shape of `broadcast_to` value following
// TF::BroadcastTo semantics.
//
// Requires that input is a ranked tensor.
//
// TODO(hinsu): Utilize TF::ShapeOp followed by TF::BroadcastTo once ShapeOp
// supports unranked inputs in the lowering.
static Value BroadcastToShapeOf(Location loc, Value input, Value broadcast_to,

  return builder->create<TF::CastOp>(loc, type, x, truncate);
}
}  // namespace

//===----------------------------------------------------------------------===//
// The actual PrepareTF Pass.
//
// TODO(hinsu): Add and use TensorFlow dialect ops for the ops created in this
// pass.
namespace {
#define GEN_PASS_DEF_PREPARETFPASS
#include "tensorflow/compiler/mlir/lite/transforms/passes.h.inc"

// PackOp
//===----------------------------------------------------------------------===//

LogicalResult PackOp::verify() {
  PackOp op = *this;
  // TODO(hinsu): Convert variadic length attributes to derived attributes.
  Operation::operand_range values = op.getValues();

  if (failed(VerifyTypesCompatibility(values,
                                      /*mask_one_dim=*/false,

                  $filter_width, $filter_height, TFL_AF_None)),
    (TFL_MaxPool2DOp $input, $padding, $stride_w, $stride_h,
        $filter_width, $filter_height, ActFnAttr),
    [(HasOneUse $pool_out)]>;
}

// TODO(hinsu): Also fuse ops corresponding to SIGN_BIT fused
// activation functions.
// Currently we're not fusing tanh, sigmoid, hard_swish and other activations
// those cannot be simply translated into clamping.

                      TF::TensorListPushBackOp, TF::TensorListReserveOp,
                      TF::TensorListStackOp, TF::TensorListResizeOp,
                      TF::TensorListConcatV2Op>();
  // TODO(hinsu): Use TFLite constant op for constants.
  target.addLegalOp<arith::ConstantOp>();
  target.addLegalOp<func::FuncOp>();
  target.addDynamicallyLegalOp<func::ReturnOp>(is_legal);

  LogicalResult matchAndRewrite(Operation *op,
                                PatternRewriter &rewriter) const override {
    auto addn_op = cast<AddNOp>(op);

    // TODO(hinsu): Support variant with TensorList type. tf.AddV2 doesn't
    // support variant type so variant types require special handling.
    if (mlir::isa<VariantType>(getElementTypeOrSelf(addn_op.getType())))
      return failure();

  // Here, custom folder doesn't handle complex dtypes and it is folded through
  // the constant folding hook.
  // TODO(hinsu): Handle complex dtypes in the custom folder for FillOp.
  // CHECK-DAG: "tf.Const"() <{value = dense<(0.000000e+00,1.000000e+00)> : tensor<3x2x1xcomplex<f32>>}> : () -> tensor<*xcomplex<f32>>

                       element_type, status.status().message())),
           false;
  }
  return true;
}

// Returns true if the module holds all the invariants expected by the
// Translator class.
// TODO(hinsu): Now that translation is done by making a single pass over the
// MLIR module, consider inlining these validation checks at the place where
// these invariants are assumed instead of checking upfront.