"GetOpQuantSpec(mlir::Operation *op, bool "
            "disable_per_channel_for_dense_layers = false) {\n";
  // TODO(b/176258587): Move to OpTrait if this should be generalized.
  // Add special handling for LSTM.
  OUT(2) << "if (auto lstm_op = llvm::dyn_cast<TFL::LSTMOp>(op)) {\n";
  OUT(4) << "return GetLstmOpQuantSpec<TFL::LSTMOp>(lstm_op);\n";
  OUT(2) << "} else if (auto lstm_op = "

  let summary = "Bidirectional sequence lstm operator";

  let description = [{
    Bidirectional lstm is essentially two lstms, one running forward & the
    other running backward. And the output is the concatenation of the two
    lstms.
  }];

  let arguments = (
    ins TFL_TensorOf<[F32, I8]>:$input,

    // Forward LSTM Weights

  // expected-error @+1 {{'tfl.lstm' op failed to verify that either projection weight must be specified or both projection weight and projection bias must not be specified}}

        used inside the composition should have support for these platforms.
        These ops can also be composite ops.

*   (Performance) User defines a custom kernel for a regular structure (i.e.
    LSTM), but it is hard to add the logic to fuse the individual ops to target
    this kernel in the inference graph.

    *   *Solution*: The user should define a new TF op, which corresponds to the

            3);

  auto it = fused_lstm_func_.getBody().back().rbegin();
  EXPECT_EQ(it->getName().getStringRef(),
            mlir::func::ReturnOp::getOperationName());
  it++;  // tensor_cast
  it++;  // lstm
  EXPECT_EQ(it->getName().getStringRef(),
            mlir::TFL::LSTMOp::getOperationName());
  EXPECT_EQ(it->getNumOperands(), 24);
  EXPECT_EQ(it->getNumResults(), 1);
  // cifg = false, so input2input is not None.

  // The conversion pipeline has to follow the following orders:
  // 1) Saved model related optimization like decompose resource ops
  // 2) Convert composite functions like lstm/rnns, along with proper function
  // inlining & dce.
  // 3) Lower static tensor list pass.

  // This decomposes resource ops like ResourceGather into read-variable op

  %arg20: tensor<1x2048xf32>
) -> tensor<1x640xf32> {
  %cst = "tfl.no_value"() {value = unit} : () -> none
  %zero = "tfl.pseudo_const"() {value = dense<0.0> : tensor<640xf32>} : () -> tensor<640xf32>
  %0 = "tfl.lstm"(%arg0, %arg1, %arg2, %arg3, %arg4, %arg5, %arg6, %arg7, %arg8, %cst, %cst, %cst, %arg9, %arg10, %arg11, %arg12, %arg13, %zero, %arg19, %arg20, %arg15, %arg16, %arg17, %arg18) ({}) {

// CHECK-DAG: %[[input_15:.*]] = "tfl.dequantize"({{.*}}) : (tensor<1x3x!quant.uniform<i16:f32, 3.0517578125E-5>>) -> tensor<1x3xf32>
// CHECK: %[[lstm:.*]] = "tfl.unidirectional_sequence_lstm"(
// CHECK-SAME: %[[input_0]],
// CHECK-SAME: %[[input_1]], %[[input_2]], %[[input_3]], %[[input_4]],
// CHECK-SAME: %[[input_5]], %[[input_6]], %[[input_7]], %[[input_8]],

    rewriter.replaceOpWithNewOp<QConstOp>(op, TypeAttr::get(result_type),
                                          reshaped_elements);
    return success();
  }
};

// Removes operations with side effect (i.e. LSTM, SVDF) that have dangling
// output.
template <typename OpTy>
struct PruneUnusedOpsWithSideEffect : public OpRewritePattern<OpTy> {
 public:
  explicit PruneUnusedOpsWithSideEffect(

// LSTMOpQuantized-DAG: %[[dq3:.*]] = "tfl.dequantize"({{.*}}) : (tensor<20x!quant.uniform<i8<-127:127>:f32, 0.0078740157480314959>>) -> tensor<20xf32>
// LSTMOpQuantized: %[[lstm:.*]] = "tfl.unidirectional_sequence_lstm"(%arg0, %[[dq1]], %[[dq1]], %[[dq1]], %[[dq1]], %[[dq1]], %[[dq1]], %[[dq1]], %[[dq1]], %[[dq3]], %[[dq3]], %[[dq3]], %cst_0, %cst_0, %cst_0, %cst_0, %[[dq1]], %0, %cst_1, %cst_1, %0, %0, %0, %0)