%2 = func.call @func_2_CPU_FLOAT(%0, %1) {tac.device = "CPU", tac.inference_type = "FLOAT", tac.interface_name = "func_2"} : (tensor<100xf32>, tensor<100xf32>) -> tensor<2x100xf32>
    func.return %2 : tensor<2x100xf32>
  }

// -----

func.func @pack_CPU(%arg0: tensor<100xf32>, %arg1: tensor<100xf32>) -> tensor<2x100xf32> attributes {tac.device = "CPU", tac.interface_name = "func_2"} {
  // CHECK: tac.cost = 1.000000e+02
  %0 = "tfl.pack"(%arg0, %arg1) {axis = 0 : i32, tac.device = "CPU", values_count = 2 : i32} : (tensor<100xf32>, tensor<100xf32>) -> tensor<2x100xf32>
  func.return %0 : tensor<2x100xf32>
}

func.func @func_0_GPU(%arg0: tensor<256x32x32x3xf32>, %arg1: tensor<256x32x32x3xf32>) -> tensor<256x32x32x3xf32> attributes {tac.device = "GPU", tac.interface_name = "func_0"} {
  %0 = "tfl.add"(%arg0, %arg1) {fused_activation_function = "RELU", tac.device = "GPU"} : (tensor<256x32x32x3xf32>, tensor<256x32x32x3xf32>) -> tensor<256x32x32x3xf32>
  func.return %0 : tensor<256x32x32x3xf32>
}

// -----

// CHECK: tac.cost = 1.000000e+03

  func.return %0 : tensor<i32>
}
}

// CHECK:     func.func @simpleWhile(%arg0: tensor<i32>) -> tensor<i32> {
// CHECK:       %0 = call @func_0_CPU_FLOAT(%arg0) {tac.device = "CPU", tac.inference_type = "FLOAT", tac.interface_name = "func_0"} : (tensor<i32>) -> tensor<i32>
// CHECK:       return %0 : tensor<i32>
// CHECK:     }

module {
  func.func @simpleTest(%arg0: tensor<1xf32>, %arg1: tensor<1xf32>, %arg2: tensor<1xf32>, %arg3: tensor<1xf32>) -> tensor<2x1xf32> {
    %0 = func.call @func_0_GPU_FLOAT(%arg0, %arg1, %arg2) {tac.interface_name = "func_0"} : (tensor<1xf32>, tensor<1xf32>, tensor<1xf32>) -> tensor<1xf32>
    %1 = func.call @func_1_GPU_FLOAT(%arg0, %arg3) {tac.interface_name = "func_1"} : (tensor<1xf32>, tensor<1xf32>) -> tensor<1xf32>

    %2 = call @func_2_CPU_FLOAT(%0, %1) {tac.device = "CPU", tac.inference_type = "FLOAT", tac.interface_name = "func_2"} : (tensor<1xf32>, tensor<1xf32>) -> tensor<2x1xf32>
    return %2 : tensor<2x1xf32>
  }
```

Why we need to raise those ops into `FuncOps`? Please see the following section.

			b := 3
			return func(y int) int { // ERROR "can inline main.func27.1" "can inline main.main.func27.func34"
				c := 5
				return func(z int) int { // ERROR "can inline main.func27.1.1" "can inline main.main.func27.func34.1" "can inline main.func27.main.func27.1.2" "can inline main.main.func27.main.main.func27.func34.func36"
					return a*x + b*y + c*z
				}(10) // ERROR "inlining call to main.func27.1.1"

  if (GetName(funcs[0]) == GetName(func0)) {
    AssertEqual(func0, funcs[0]);
    AssertEqual(func1, funcs[1]);
  } else {
    AssertEqual(func0, funcs[1]);
    AssertEqual(func1, funcs[0]);
  }

  TF_DeleteFunction(funcs[0]);
  TF_DeleteFunction(funcs[1]);

  TF_DeleteFunction(func0);
  TF_DeleteFunction(func1);
}

}  // namespace

// This pass is used to fold tfl.const ops to each subgraph (func::FuncOp):
// See the example below:
//
// In main:
// %0 = tfl.const...
// %1 = tfl.const...
// %2 = call func_1(..., %0,...)
// %3 = call func_2(..., %0, ..., %1...)
// ...
//
// Then those consts will be copied into each function and replace their usage.
// func_1:
//   %0 = tfl.const...
// func_2:
//   %0 = tfl.const...
//   %1 = tfl.const...

// Returns true iff the function's symbol is public.
bool IsPublicFuncOp(func::FuncOp func_op) {
  return SymbolTable::getSymbolVisibility(&*func_op) ==
         SymbolTable::Visibility::Public;
}

}  // namespace

func::FuncOp FindMainFuncOp(ModuleOp module_op) {
  if (const auto main_func_op = module_op.lookupSymbol<func::FuncOp>(
          kImportModelDefaultGraphFuncName);