return func(x int) int { // ERROR "can inline main.func12"
				return x + 1
			}, 42
		}() // ERROR "func literal does not escape" "inlining call to main.func12"
		if y(40) != 41 {
			ppanic("y(40) != 41")
		}
	}

	{
		func() { // ERROR "func literal does not escape"
			y := func(x int) int { // ERROR "func literal does not escape" "can inline main.func13.1"
				return x + 2
			}

func.func @func_0_CPU(%arg0: tensor<256x32x32x3xf32>, %arg1: tensor<256x32x32x3xf32>) -> tensor<256x32x32x3xf32> attributes {tac.device = "CPU", tac.interface_name = "func_0"} {
  // CHECK: tac.cost = 7.864320e+05
  %0 = "tfl.add"(%arg0, %arg1) {fused_activation_function = "RELU", tac.device = "CPU"} : (tensor<256x32x32x3xf32>, tensor<256x32x32x3xf32>) -> tensor<256x32x32x3xf32>

	}
	go func() { // func10
		c, err := ln.Accept()
		if err != nil {
			log.Printf("failed to accept: %v", err)
			return
		}
		c.Close()
	}()
	rp, wp, err := os.Pipe()
	if err != nil {
		log.Fatalf("failed to create a pipe: %v", err)
	}
	defer rp.Close()
	defer wp.Close()
	pipeReadDone := make(chan bool)
	go func() { // func11
		var data [1]byte
		rp.Read(data[:])

// 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...
class FoldConstantsToSubgraphPass

Robert Griesemer <******@****.***> 1607048150 -0800

// CHECK: tac.cost = 7.864320e+05
func.func @func_0_CPU(%arg0: tensor<256x32x32x3xf32>, %arg1: tensor<256x32x32x3xf32>) -> tensor<256x32x32x3xf32> attributes {tac.device = "CPU", tac.interface_name = "func_0"} {
  %0 = "tfl.add"(%arg0, %arg1) {fused_activation_function = "RELU", tac.device = "CPU"} : (tensor<256x32x32x3xf32>, tensor<256x32x32x3xf32>) -> tensor<256x32x32x3xf32>
  func.return %0 : tensor<256x32x32x3xf32>
}

// Take the following example:
//
// call @func_1_CPU {tac.interface_name = "func_1"}
//
// "func_1" is the interface name where "func_1_cpu" is the real implementation
// we can have multiple FuncOps like "func_1_cpu" and "func_1_gpu" and they
// both implement "func_1".
//
// The attribute on the FuncOp means what it actually implements while the

// exist.
// TODO(b/284222084): Move executor dialect utilities to a new library.
GraphOp GetGraphOpFromFuncOp(func::FuncOp func_op) {
  if (func_op->getNumRegions() == 0 || func_op.getBody().empty()) return {};

  auto graph_op_range = func_op.front().without_terminator();
  if (llvm::hasSingleElement(graph_op_range)) {
    // The pass runs on a valid tf_executor dialect, so the op should be the
    // GraphOp.

static bool FunctionContainsOnlyNoSideEffectOpOrTensorArrayOp(
    mlir::func::FuncOp func_op) {
  for (mlir::Operation& op : func_op.front()) {
    if (!mlir::isMemoryEffectFree(&op) && !IsTensorArrayOp(&op)) return false;
  }

  return true;
}

TensorArraySideEffectAnalysis::TensorArraySideEffectAnalysis(
    mlir::ModuleOp module) {
  for (auto func_op : module.getOps<mlir::func::FuncOp>()) {

    SymbolTable symbol_table(call_op->getParentOfType<ModuleOp>());
    auto func_op =
        dyn_cast_or_null<func::FuncOp>(symbol_table.lookup(func_name));
    if (!func_op) return failure();
    // The quantized fusion should have requantize and return ops at the end.
    auto return_op = dyn_cast_or_null<func::ReturnOp>(
        func_op.getRegion().getBlocks().front().getTerminator());
    if (!return_op) return failure();