test32()
	test64()
}

func test32() {
	var a uint32 = 0x11111111
	var want, got uint32
	if want, got = a+c32a, add32a(a); got != want {
		panic(fmt.Sprintf("add32a(%x) = %x, want %x", a, got, want))
	}
	if want, got = a+c32s, add32s(a); got != want {
		panic(fmt.Sprintf("add32s(%x) = %x, want %x", a, got, want))
	}
	if want, got = a-c32a, sub32a(a); got != want {

// CHECK: %[[relu:.*]] = "tf.Relu"(%[[dequant1]]
// CHECK: %[[clamped:.*]] = "tf.Minimum"(%[[relu]]

// CHECK: %[[rescale1:.*]] = "tf.Mul"(%[[cast]]
// CHECK: %[[add2:.*]] = "tf.AddV2"(%[[rescale1]]
// CHECK: %[[maximum2:.*]] = "tf.Maximum"(%[[add2]]
// CHECK: %[[minimum2:.*]] = "tf.Minimum"(%[[maximum2]]
// CHECK: %[[round2:.*]] = "tf.Round"(%[[minimum2]]

  Output in = ops::Placeholder(on_gpu.WithOpName("in"), DT_FLOAT);
  Output c = ops::Const(on_gpu.WithOpName("const"), 1.0f, {});
  Output add1 = ops::AddV2(on_gpu.WithOpName("add1"), in, c);
  Output add2 = ops::AddV2(on_gpu.WithOpName("add2"), add1, c);

  std::unique_ptr<Graph> result;
  TF_ASSERT_OK(CloneConstantsForBetterClustering(root, &result));

  OutputTensor add1_operand;
  TF_ASSERT_OK(

func ExampleAdd32() {
	// First number is 33<<32 + 12
	n1 := []uint32{33, 12}
	// Second number is 21<<32 + 23
	n2 := []uint32{21, 23}
	// Add them together without producing carry.
	d1, carry := bits.Add32(n1[1], n2[1], 0)
	d0, _ := bits.Add32(n1[0], n2[0], carry)
	nsum := []uint32{d0, d1}
	fmt.Printf("%v + %v = %v (carry bit was %v)\n", n1, n2, nsum, carry)

	// First number is 1<<32 + 2147483648
	n1 = []uint32{1, 0x80000000}

	inf := 1.0 / zero
	negZero := -1 / inf
	if 1/neg64(negZero) != inf {
		panic("-negZero != posZero (64 bit)")
	}
}

//go:noinline
func add32(x float32) float32 {
	return x + 0
}

func testAdd32() {
	var zero float32
	inf := 1.0 / zero
	negZero := -1 / inf
	if 1/add32(negZero) != inf {
		panic("negZero+0 != posZero (32 bit)")
	}
}

//go:noinline
func sub32(x float32) float32 {
	return x - 0
}

  %2 = tfrt_fallback_async.executeop key(1) cost(512) device("/job:localhost/replica:0/task:0/device:CPU:0") "tf.AddV2"(%0, %0) {T = f32} : 1
  %3 = tfrt_fallback_async.executeop key(1) cost(512) device("/job:localhost/replica:0/task:0/device:CPU:0") "tf.AddV2"(%0, %0) {T = f32} : 1
  %4 = tfrt_fallback_async.executeop key(1) cost(512) device("/job:localhost/replica:0/task:0/device:CPU:0") "tf.AddV2"(%0, %0) {T = f32} : 1

	uaddr := uintptr(unsafe.Pointer(addr))
	addr32 := (*uint32)(unsafe.Pointer(uaddr &^ 3))
	word := uint32(v) << ((uaddr & 3) * 8) // little endian
	for {
		old := *addr32
		if Cas(addr32, old, old|word) {
			return
		}
	}
}

//go:nosplit
func And8(addr *uint8, v uint8) {
	// Align down to 4 bytes and use 32-bit CAS.
	uaddr := uintptr(unsafe.Pointer(addr))
	addr32 := (*uint32)(unsafe.Pointer(uaddr &^ 3))

  %tmp2 = "tf.AddV2"(%tmp1, %tmp1) { device = "" } : (tensor<1xf32>, tensor<1xf32>) -> (tensor<1xf32>)
  %tmp3 = "tfl.add"(%const, %tmp2) { fused_activation_function = "NONE" } : (tensor<1xf32>, tensor<1xf32>) -> (tensor<1xf32>)
  %tmp4 = "tf.AddV2"(%tmp2, %tmp2) { device = "" } : (tensor<1xf32>, tensor<1xf32>) -> (tensor<1xf32>)

  %9 = "tf.AddV2"(%8, %cst_6) : (tensor<1x98x98x1xf32>, tensor<1xf32>) -> tensor<1x98x98x1xf32>
  %10 = "tf.Mul"(%9, %cst_5) : (tensor<1x98x98x1xf32>, tensor<1xf32>) -> tensor<1x98x98x1xf32>
  %11 = "tf.AddV2"(%10, %cst_8) : (tensor<1x98x98x1xf32>, tensor<f32>) -> tensor<1x98x98x1xf32>

	var packetConnTests = []struct {
		net   string
		addr1 string
		addr2 string
	}{
		{"udp", "127.0.0.1:0", "127.0.0.1:0"},
		{"unixgram", testUnixAddr(t), testUnixAddr(t)},
	}

	closer := func(c PacketConn, net, addr1, addr2 string) {
		c.Close()
		switch net {
		case "unixgram":
			os.Remove(addr1)
			os.Remove(addr2)
		}
	}

	for _, tt := range packetConnTests {