$mlir_module is a serialized MLIR module with a `main` function that contains target computation.
    $metadata is a serialized TPUCompileMetadataProto describing the shapes and types of the inputs to the computation, as well as a mapping onto the TPU pod topology.

    XlaComputation& computation) {
  xla::DebugOptions debug_options;
  TF_ASSIGN_OR_RETURN(auto hlo_module_config,
                      xla::HloModule::CreateModuleConfigFromProto(
                          computation.proto(), debug_options));
  TF_ASSIGN_OR_RETURN(
      std::unique_ptr<xla::HloModule> hlo_module,
      xla::HloModule::CreateFromProto(computation.proto(), hlo_module_config));

    cache.invalidateAll();
    // let the computation proceed
    computingLatch.countDown();
    // don't check cache.size() until we know the get("b") call is complete
    computationComplete.await();

    // At this point, the listener should be holding the seed value (a -> a), and the map should
    // contain the computed value (b -> b), since the clear() happened before the computation
    // completed.

// run

// Copyright 2013 The Go Authors. All rights reserved.
// Use of this source code is governed by a BSD-style
// license that can be found in the LICENSE file.

// Issue 5244: the init order computation uses the wrong
// order for top-level blank identifier assignments.
// The example used to panic because it tries calling a
// nil function instead of assigning to f before.

package main

var f = func() int { return 1 }
var _ = f() + g()

    %add_result = "tf.Add"(%computation, %computation) : (tensor<i32>, tensor<i32>) -> tensor<i32>
    // expected-error@+1 {{Input of TPUPartitionedOutputV2 must be in tpu computation.}}
    %partitioned_output:2 = "tf.TPUPartitionedOutputV2"(%add_result) {N = 2 : i64, partition_dims = []} : (tensor<i32>) -> (tensor<i32>, tensor<i32>)

// Issue 15747: If an ODCL is dropped, for example when inlining,
// then it's easy to end up not initializing the '&x' pseudo-variable
// to point to an actual allocation. The liveness analysis will detect
// this and abort the computation, so this test just checks that the
// compilation succeeds.

package p

type R [100]byte

func (x R) New() *R {
	return &x

  xla::ExecutableBuildOptions build_options = GetExecutableBuildOptions(
      options, result, client_->default_device_ordinal());
  TF_ASSIGN_OR_RETURN(
      auto executables,
      client_->Compile(*result.computation, argument_layouts, build_options));
  TF_RET_CHECK(executables.size() == 1);
  return std::move(executables[0]);
}

absl::StatusOr<std::string> XlaDeviceCompilerClient::SerializeExecutable(

namespace tf2xla {
namespace v2 {

// Legalizes the given mlir::Module into XLA HLO. If successful, returns the
// compiled XLA HLO. V1 of the tf2xla uses MLIR whereas V0 does not use MLIR.
//
// Inputs:
//  computation - The MLIR module op. It currently takes in
//  tpu::FunctionToHloArgs but this is deprecated. arg_shapes - The shapes of
//  the arguments in module_op. device_type - The device type to compile for.

	for len(p) >= chunk {
		// Can interlace the computation of w with the
		// rounds below if needed for speed.
		for i := 0; i < 16; i++ {
			j := i * 4
			w[i] = uint32(p[j])<<24 | uint32(p[j+1])<<16 | uint32(p[j+2])<<8 | uint32(p[j+3])
		}

		a, b, c, d, e := h0, h1, h2, h3, h4

		// Each of the four 20-iteration rounds
		// differs only in the computation of f and
		// the choice of K (_K0, _K1, etc).
		i := 0

  func.return %0#0 : tensor<*xi32>
}

// CHECK: func private @compute_0