// CHECK-NEXT: tf_mlrt.executeop([[y]]
  %w = "tf.AddV2"(%y, %z) {__op_key = 2: i32}: (tensor<i32>, tensor<i32>) -> tensor<i32>
  // CHECK-NEXT: tf_mlrt.executeop([[y]]
  %u = "tf.AddV2"(%y, %z) {__op_key = 3: i32} : (tensor<i32>, tensor<i32>) -> tensor<i32>
  // CHECK-NEXT: tf_mlrt.executeop([[y]]
  %v = "tf.AddV2"(%y, %z) {__op_key = 4: i32}: (tensor<i32>, tensor<i32>) -> tensor<i32>

	{name: "Add8", argLength: 2, commutative: true}, // arg0 + arg1
	{name: "Add16", argLength: 2, commutative: true},
	{name: "Add32", argLength: 2, commutative: true},
	{name: "Add64", argLength: 2, commutative: true},
	{name: "AddPtr", argLength: 2}, // For address calculations.  arg0 is a pointer and arg1 is an int.
	{name: "Add32F", argLength: 2, commutative: true},
	{name: "Add64F", argLength: 2, commutative: true},

	hi, lo1b := bits.Mul32(uint32(x>>32), n)
	lo1, c := bits.Add32(lo1a, lo1b, 0)
	hi += c
	if lo1 == 0 && lo0 < uint32(n) {
		n64 := uint64(n)
		thresh := uint32(-n64 % n64)
		for lo1 == 0 && lo0 < thresh {
			x := r.Uint64()
			lo1a, lo0 = bits.Mul32(uint32(x), n)
			hi, lo1b = bits.Mul32(uint32(x>>32), n)
			lo1, c = bits.Add32(lo1a, lo1b, 0)
			hi += c
		}
	}
	return hi
}

  %next_index = "tf.AddV2"(%index, %one) {device = "/job:localhost/replica:0/task:0/device:CPU:0"} : (tensor<i32>, tensor<i32>) -> tensor<i32>

//	struct timespec *timeout, int32 *uaddr2, int32 val2);
TEXT runtime·futex(SB),NOSPLIT|NOFRAME,$0
	MOVD	addr+0(FP), R3
	MOVW	op+8(FP), R4
	MOVW	val+12(FP), R5
	MOVD	ts+16(FP), R6
	MOVD	addr2+24(FP), R7
	MOVW	val3+32(FP), R8
	SYSCALL	$SYS_futex
	BVC	2(PC)
	NEG	R3	// caller expects negative errno
	MOVW	R3, ret+40(FP)
	RET

// int64 clone(int32 flags, void *stk, M *mp, G *gp, void (*fn)(void));

		{name: "LoweredAtomicLoadPtr", argLength: 2, reg: gpload, typ: "Int64", aux: "Int64", clobberFlags: true, faultOnNilArg0: true},

		// atomic add32, 64
		// LWSYNC
		// LDAR         (Rarg0), Rout
		// ADD		Rarg1, Rout
		// STDCCC       Rout, (Rarg0)
		// BNE          -3(PC)
		// return new sum