%0 = "tf.AddV2"(%arg0, %cst) {device = ""} : (tensor<i32>, tensor<i32>) -> tensor<i32>
    %1 = "tf.Identity"(%0) {device = ""} : (tensor<i32>) -> tensor<i32>
    %2 = "tf.Identity"(%arg1) {device = ""} : (tensor<i32>) -> tensor<i32>
    %4 = "tf.AddV2"(%arg2, %cst_0) {device = ""} : (tensor<1x5x5x1024xf32>, tensor<1x5x5x1024xf32>) -> tensor<1x5x5x1024xf32>

	// TCP has a rarely used mechanism called a 'simultaneous connection' in
	// which Dial("tcp", addr1, addr2) run on the machine at addr1 can
	// connect to a simultaneous Dial("tcp", addr2, addr1) run on the machine
	// at addr2, without either machine executing Listen. If laddr == nil,
	// it means we want the kernel to pick an appropriate originating local

  // CHECK: %[[SUM0:.*]] = "tf.AddV2"(%arg0, %arg1)
  // CHECK: %[[SUM1:.*]] = "tf.AddV2"(%arg2, %arg3)
  // CHECK: %[[SUM2:.*]] = "tf.AddV2"(%[[SUM0]], %[[SUM1]])
  // CHECK: %[[SUM3:.*]] = "tf.AddV2"(%[[SUM2]], %arg4)
  // return %[[SUM3]]

		if h2r0.hi != 0 {
			panic("poly1305: unexpected overflow")
		}
		if h2r1.hi != 0 {
			panic("poly1305: unexpected overflow")
		}

		m0 := h0r0
		m1 := add128(h1r0, h0r1) // These two additions don't overflow thanks again
		m2 := add128(h2r0, h1r1) // to the 4 masked bits at the top of r0 and r1.
		m3 := h2r1

		t0 := m0.lo
		t1, c := bits.Add64(m1.lo, m0.hi, 0)
		t2, c := bits.Add64(m2.lo, m1.hi, c)

    %30 = "tf.CrossReplicaSum"(%29, %3) : (tensor<f32>, tensor<1x2xi32>) -> tensor<f32>
    %31 = "tf.AddV2"(%arg5, %23) : (tensor<f32>, tensor<f32>) -> tensor<f32>
    %32 = "tf.CrossReplicaSum"(%1, %3) : (tensor<f32>, tensor<1x2xi32>) -> tensor<f32>
    %33 = "tf.AddV2"(%arg6, %32) : (tensor<f32>, tensor<f32>) -> tensor<f32>
    %34 = "tf.AddV2"(%arg7, %30) : (tensor<f32>, tensor<f32>) -> tensor<f32>

	ORL	BX, (AX)
	RET

// func And(addr *uint32, v uint32)
TEXT ·And(SB), NOSPLIT, $0-8
	MOVL	ptr+0(FP), AX
	MOVL	val+4(FP), BX
	LOCK
	ANDL	BX, (AX)
	RET

// func And32(addr *uint32, v uint32) old uint32
TEXT ·And32(SB), NOSPLIT, $0-12
	MOVL	ptr+0(FP), BX
	MOVL	val+4(FP), CX
casloop:
	MOVL 	CX, DX
	MOVL	(BX), AX
	ANDL	AX, DX
	LOCK
	CMPXCHGL	DX, (BX)
	JNZ casloop
	MOVL 	AX, ret+8(FP)

  %0 = "tf.Mul"(%arg0, %arg1) : (tensor<i32>, tensor<i32>) -> tensor<i32>
  // CHECK: %[[RESULT:.*]]:4 = "tf_device.cluster"
  %1:4 = "tf_device.cluster"() ({
    // CHECK: %[[ADD:.*]] = "tf.AddV2"
    %2 = "tf.AddV2"(%arg0, %arg1) : (tensor<i32>, tensor<i32>) -> tensor<i32>
    // CHECK: %[[SUB:.*]] = "tf.Sub"
    %3 = "tf.Sub"(%arg0, %arg1) : (tensor<i32>, tensor<i32>) -> tensor<i32>

// CHECK-NOT: tf.BatchFunction
// CHECK: %[[ADD0:.*]] = "tf.AddV2"(%[[ARG0]], %[[ARG1]])
// CHECK: return %[[ADD0]] : tensor<1xf32>

func.func private @batched_func(%arg0: tensor<1xf32>, %arg1: tensor<1xf32>) -> tensor<1xf32> {
  %0 = "tf.Identity"(%arg0) : (tensor<1xf32>) -> tensor<1xf32>
  %1 = "tf.Identity"(%arg1) : (tensor<1xf32>) -> tensor<1xf32>
  %2 = "tf.AddV2"(%0, %1) : (tensor<1xf32>, tensor<1xf32>) -> tensor<1xf32>

func.func private @sum_reducer(%arg0: tensor<*xf32>, %arg1: tensor<*xf32>) -> tensor<*xf32> {
  %0 = "tf.AddV2"(%arg0, %arg1) {device = ""} : (tensor<*xf32>, tensor<*xf32>) -> tensor<*xf32>
  func.return %0 : tensor<*xf32>
}

func.func @xla_func_0(%arg0: tensor<7xf32>, %arg1: tensor<f32>) -> tensor<10xf32> {
  %cst = "tf.Const"() {value = dense<0> : tensor<1x2xi32>} : () -> tensor<1x2xi32>

}

func.func @no_side_effect_body(%arg: tensor<i32>) -> tensor<i32> {
  %0 = "tf.AddV2"(%arg, %arg) : (tensor<i32>, tensor<i32>) -> tensor<i32>
  func.return %0 : tensor<i32>
}

func.func @no_side_effect_body2(%arg: tensor<i32>) -> tensor<i32> {
  %0 = "tf.Const"() {value = dense<1> : tensor<i32> } : () -> tensor<i32>
  %1 = "tf.AddV2"(%arg, %0) : (tensor<i32>, tensor<i32>) -> tensor<i32>
  func.return %1 : tensor<i32>
}