return func_.emitWarning()
             << "Invalid number of arguments in the embedding "
                "matmul composite function";
    }
    if (func_.getFunctionType().getNumResults() != 1) {
      return func_.emitWarning() << "Invalid number of results in the "
                                    "embedding matmul composite function";
    }
    return success();
  }

 private:
  func::FuncOp func_;
};

}

func.func @matmul(%arg0: tensor<40x37xf32>, %arg1: tensor<37x40xf32>) -> tensor<40x40xf32> {
  %0 = "tf.MatMul"(%arg0, %arg1) {T = "tfdtype$DT_FLOAT", device = "/device:CPU:0", name = "MatMul", transpose_a = false, transpose_b = false} :
(tensor<40x37xf32>, tensor<37x40xf32>) -> tensor<40x40xf32>
  func.return %0 : tensor<40x40xf32>
// CHECK-LABEL: matmul

      %outputs_16, %control_17 = tf_executor.island wraps "tf.Reshape"(%outputs_14, %outputs_6) {device = ""} : (tensor<16x16x16x?xf32>, tensor<2xi32>) -> tensor<?x16384xf32>
      %outputs_18, %control_19 = tf_executor.island wraps "tf.MatMul"(%outputs_16, %outputs_4) {device = "", transpose_a = false, transpose_b = false} : (tensor<?x16384xf32>, tensor<*xf32>) -> tensor<?x?xf32>

    Value input_rhs = bmm_op.getY();

    Value output_lhs =
        bmm_op.getAdjX() ? create_z_x_transpose_op(input_lhs) : input_lhs;

    // The rhs need to be transposed if adj_y == false AND this matmul will be
    // legalized to tfl.fully_connected
    Value output_rhs =
        !bmm_op.getAdjY() ? create_z_x_transpose_op(input_rhs) : input_rhs;

    Type output_type = bmm_op.getResult().getType();

#include "xla/xla_data.pb.h"

namespace mlir::quant {
namespace {

constexpr StringRef kTfQuantCreatedEinsum = "__tf_quant_created_einsum";

// Replaces mixed-type Conv and Matmul cast hacks with TF XLA ops.
// TODO(b/228403741): Support conversion for dynamic-shaped TF ops.
class ReplaceCastHacksWithTFXLAOpsPass
    : public PassWrapper<ReplaceCastHacksWithTFXLAOpsPass,

      return %0 : tensor<2x2xf32>
    }
    func.func private @composite_fn_1(%arg0: tensor<2x2xf32>, %arg1: tensor<2x2xf32>) -> tensor<2x2xf32> attributes {tf_quant.composite_function} {

        loc, value, Create1DConstValue(builder, loc, new_shape));
  }
  return ConstantFoldOpIfPossible(value.getDefiningOp()).front();
}

// Matches convolution op with "NHWC" data format or matmul op with false adj_y.
// The list of supported ops in this function is:
// - Conv2DOp
// - Conv3DOp
// - DepthwiseConv2dNativeOp
// - MatMulOp
// - BatchMatMulV2Op

             "bool", "false",
             "Disable folding mul and fully connected ops during optimization pass.">,
  ];
}

def OptimizeBatchMatmulPass : Pass<"tfl-optimize-batch-matmul", "mlir::func::FuncOp"> {
  let summary = "Optimize FC with BatchMatmul within the TensorFlow Lite dialect";
  let constructor = "CreateOptimizeBatchMatmulPass()";
  let dependentDialects = ["TFL::TensorFlowLiteDialect"];
}

std::unique_ptr<OperationPass<mlir::func::FuncOp>> CreateOptimizePass();

// Creates an instance of the ReplaceCastHacksWithTFXLAOpsPass, which will
// replace mixed-type convolution and matmul cast hacks by XLA Conv2DOp and
// MatmulOp.
std::unique_ptr<OperationPass<func::FuncOp>>
CreateReplaceCastHacksWithTFXLAOpsPass();

// Creates a pass that moves & merges initializer function's ops into the @main

  %7 = "tf.Reshape"(%6, %cst_2) {device = "/job:localhost/replica:0/task:0/device:CPU:0"} : (tensor<9xf32>, tensor<2xi32>) -> tensor<3x3xf32>
  %8 = "tf.MatMul"(%2, %7) {device = "/job:localhost/replica:0/task:0/device:CPU:0", transpose_a = false, transpose_b = false} : (tensor<3x3xf32>, tensor<3x3xf32>) -> tensor<3x3xf32>