"      input: 'var'"
      "      device: '/job:localhost/replica:0/task:0/device:CPU:0'"
      "      attr {"
      "        key: 'dtype'"
      "        value {"
      "          type: DT_FLOAT"
      "        }"
      "      }"
      "    }"
      "    node_def {"
      "      name: 'read1'"
      "      op: 'ReadVariableOp'"
      "      input: 'var'"
      "      device: '/job:localhost/replica:0/task:1/device:CPU:0'"

  assigned_id->setAttr(kICIWeightDistributionMlirBridgeMarker,
                       before_cluster_builder.getBoolAttr(true));
  std::string device = tensorflow::GetDeviceAliasForHostOfLogicalCore(0);
  LaunchOp launch = tensorflow::WrapOpInLaunch(
      &before_cluster_builder, val_bcast.getLoc(), assigned_id, device);

  Value all_reduce =
      CreateAllReduce(replicate, inner_builder, launch.getResult(0));

#define _sa (20*8)
#define _se (21*8)
#define _si (22*8)
#define _so (23*8)
#define _su (24*8)

// Temporary registers
#define rT1  AX

// Round vars
#define rpState DI
#define rpStack SP

#define rDa BX
#define rDe CX
#define rDi DX
#define rDo R8
#define rDu R9

#define rBa R10
#define rBe R11
#define rBi R12
#define rBo R13
#define rBu R14

#define rCa SI
#define rCe BP

// CHECK: [[const:%.*]] = "tf.Const"() <{value = dense<0> : tensor<i32>}> {device = "/CPU:0"} : () -> tensor<i32>
// CHECK: [[x:%.*]] = "tf.AddV2"([[const]], [[const]]) {device = "/CPU:0"} : (tensor<i32>, tensor<i32>) -> tensor<i32>
// CHECK: "tf._TfrtSetResource"([[x]]) <{index = 0 : i64}> {device = "/CPU:0"} : (tensor<i32>) -> ()
// CHECK: [[const_1:%.*]] = "tf.Const"() <{value = dense<1> : tensor<i32>}> {device = "/CPU:0"} : () -> tensor<i32>

// constant XlaLiteral is included in the HLO graph, and subsequently, in
// the actual executable, which is copied to the device before being
// executed. Thus, when this executable runs, the constant is available in
// device memory.
tensorflow::MemoryTypeVector GetInputMemoryTypes(
    const tensorflow::FunctionBody* fbody,
    absl::Span<int const> constant_arg_indices,

#define PRECALC_33(REG_SUB_28,REG) \
	VPXOR REG_SUB_28, REG, REG

#define PRECALC_34(REG_SUB_16) \
	VPXOR REG_SUB_16, Y0, Y0

#define PRECALC_35(REG) \
	VPXOR Y0, REG, REG

#define PRECALC_36(REG) \
	VPSLLD $2, REG, Y0

#define PRECALC_37(REG) \
	VPSRLD $30, REG, REG \
	VPOR REG, Y0, REG

#define PRECALC_39(REG,K_OFFSET,OFFSET) \
	VPADDD K_OFFSET(R8), REG, Y0 \

        sourceFile << """
            #define HEADER1_NAME1 "hello1.h"
            #define HEADER1_NAME2 "hello2.h"
            #define HEADER2_NAME1 "hello3.h"
            #define HEADER2_NAME2 "hello4.h"
            #if 0
            #define HEADER HEADER1_
            #define NAME NAME1
            #else
            #define HEADER HEADER2_
            #define NAME NAME2
            #endif

#include "xla/mlir_hlo/mhlo/transforms/rewriters.h"
#include "xla/mlir_hlo/mhlo/utils/type_conversion.h"
#include "tensorflow/core/lib/monitoring/counter.h"

namespace mlir {
namespace mhlo {
namespace {

#define GEN_PASS_DEF_LEGALIZETF
#include "tensorflow/compiler/mlir/tf2xla/transforms/xla_legalize_tf_passes.h.inc"

auto *mlir_legalization_count = tensorflow::monitoring::Counter<1>::New(

    // CHECK-SAME: <{device = "tpu0"}>
    // CHECK: %[[B_OUTPUT:[0-9]*]] = "tf.B"(%[[A_OUTPUT]]) : (tensor<?xi32>) -> tensor<?xi32>
    %3 = "tf.B"(%2) {device = "tpu0"} : (tensor<?xi32>) -> tensor<?xi32>

    // CHECK: %[[C_OUTPUT:[0-9]*]] = "tf.C"(%[[A_OUTPUT]], %[[B_OUTPUT]]) : (tensor<?xi32>, tensor<?xi32>) -> tensor<?xi32>
    %4 = "tf.C"(%2, %3) {device = "tpu0"} : (tensor<?xi32>, tensor<?xi32>) -> tensor<?xi32>

      %control_31 = tf_executor.island wraps "tf.NoOp"() {device = "/device:CPU:0"} : () -> ()
      %outputs_32, %control_33 = tf_executor.island wraps "tf.Const"() {device = "/device:CPU:0", value = dense<4> : tensor<i32>} : () -> tensor<i32>
      %control_34 = tf_executor.island wraps "tf.NoOp"() {device = "/device:CPU:0"} : () -> ()