%arg1, %arg2, %arg3, %arg4, %arg5, %arg6, %arg7, %arg8, %arg9, %arg10, %arg11, %arg12, %arg13, %arg14, %arg15, %arg16, %arg17, %cst0, %cst1, %arg18, %arg19, %arg20, %arg21) ({}) {cell_clip = 0.000000e+00 : f32, fused_activation_function = "NONE", kernel_type = #tfl<lstm_kernel_type_attr FULL>, proj_clip = 0.000000e+00 : f32} : (tensor<1x4xf32>, tensor<4x4xf32>, tensor<4x4xf32>, tensor<4x4xf32>, tensor<4x4xf32>, tensor<4x4xf32>, tensor<4x4xf32>, tensor<4x4xf32>, tensor<4x4xf32>, tensor<4xf32>, tensor<4xf32>,...

// CHECK-LABEL: @main

  mlir::TFL::LSTMKernelType kernel_type;
  switch (value) {
    case tflite::LSTMKernelType_FULL:
      kernel_type = mlir::TFL::LSTMKernelType::FULL;
      break;
    case tflite::LSTMKernelType_BASIC:
      kernel_type = mlir::TFL::LSTMKernelType::BASIC;
      break;
  }
  return mlir::TFL::LSTMKernelTypeAttr::get(builder.getContext(), kernel_type);
}

  %0 = "tfl.lstm"(%arg0, %arg1, %arg2, %arg3, %arg4, %arg5, %arg6, %arg7, %arg8, %arg9, %arg10, %arg11, %arg12, %arg13, %arg14, %arg15, %arg16, %arg17, %arg18, %arg19, %arg20, %arg21, %arg22, %arg23) ({}) {cell_clip = 1.000000e+00 : f32, fused_activation_function = "NONE", kernel_type = #tfl<lstm_kernel_type_attr BASIC>} : (tensor<?xf32>,...

     cell_clip = 1.000000e+01 : f32, fused_activation_function = "TANH", kernel_type = #tfl<lstm_kernel_type_attr FULL>, proj_clip = 0.01 : f32

  cell_clip: float; // Optional, 0.0 means no clipping
  proj_clip: float; // Optional, 0.0 means no clipping

  // Parameters for LSTM version 2 or above.
  // Basic kernel is only supported in version 2 or above.
  kernel_type: LSTMKernelType = FULL;

  // Parameters for LSTM version 4 or above.
  asymmetric_quantize_inputs: bool;
}

// An implementation of TensorFlow dynamic_rnn with LSTMCell.

  cell_clip: float; // Optional, 0.0 means no clipping
  proj_clip: float; // Optional, 0.0 means no clipping

  // Parameters for LSTM version 2 or above.
  // Basic kernel is only supported in version 2 or above.
  kernel_type: LSTMKernelType = FULL;

  // Parameters for LSTM version 4 or above.
  asymmetric_quantize_inputs: bool;
}

// An implementation of TensorFlow dynamic_rnn with LSTMCell.

    // Reshape kernel to add a new spatial dimension.
    auto kernel_type = mlir::cast<ShapedType>(conv_op.getRhs().getType());
    SmallVector<int64_t, 4> kernel_2d_shape;
    for (int64_t dim : kernel_type.getShape()) {
      kernel_2d_shape.push_back(dim);
    }
    kernel_2d_shape.push_back(1);
    auto kernel_2d_type =
        RankedTensorType::get(kernel_2d_shape, kernel_type.getElementType());
    auto kernel_2d_op =

    // precise than the input types, are used for this op.
    Type kernel_type = mlir::cast<TensorType>(scale.getType()).getElementType();
    grad = rewriter.create<ConvertOp>(loc, grad, kernel_type);
    act = rewriter.create<ConvertOp>(loc, act, kernel_type);

    tensorflow::TensorFormat data_format;
    if (!FormatFromString(op.getDataFormat().str(), &data_format))

// CHECK-SAME:      <{cell_clip = 1.000000e+01 : f32, fused_activation_function = "TANH", kernel_type = #tfl<lstm_kernel_type_attr FULL>, proj_clip = 0.000000e+00 : f32}> ({