%cst = arith.constant dense<0.0> : tensor<2x2xf32>
  %0 = "tf.AddV2"(%arg0, %cst) : (tensor<1x2xf32>, tensor<2x2xf32>) -> tensor<2x2xf32>
  func.return %0 : tensor<2x2xf32>

  // CHECK-LABEL: DontRemoveTrivialAdd
  // CHECK: %[[CONST:.*]] = arith.constant dense<0.000000e+00> : tensor<2x2xf32>
  // CHECK: %[[RESULT:.*]] = "tf.AddV2"(%arg0, %[[CONST]]) : (tensor<1x2xf32>, tensor<2x2xf32>) -> tensor<2x2xf32>

#include "tensorflow/c/experimental/ops/array_ops.h"
#include "tensorflow/c/experimental/ops/math_ops.h"
#include "tensorflow/c/experimental/ops/nn_ops.h"

using std::vector;
using tensorflow::ops::AddV2;
using tensorflow::ops::Div;
using tensorflow::ops::DivNoNan;
using tensorflow::ops::MatMul;
using tensorflow::ops::Mul;
using tensorflow::ops::Neg;
using tensorflow::ops::OnesLike;
using tensorflow::ops::SqrtGrad;

        %1 = "tf.Const"() {_xla_outside_compilation = "0", value = dense<1.0> : tensor<f32>} : () -> (tensor<f32>)
        %2 = "tf.Identity"(%1) {_xla_outside_compilation = "0"} : (tensor<f32>) -> (tensor<f32>)
        %3 = "tf.AddV2"(%1, %2) : (tensor<f32>, tensor<f32>) -> (tensor<f32>)
        tf_device.return %3 : tensor<f32>
      }) {num_cores_per_replica = 1, topology =  "", device_assignment =  []} : () -> tensor<f32>
      return %0 : tensor<f32>

	MOV	$SYS_madvise, A7
	ECALL
	MOVW	A0, ret+24(FP)
	RET

// func futex(addr unsafe.Pointer, op int32, val uint32, ts, addr2 unsafe.Pointer, val3 uint32) int32
TEXT runtime·futex(SB),NOSPLIT|NOFRAME,$0
	MOV	addr+0(FP), A0
	MOVW	op+8(FP), A1
	MOVW	val+12(FP), A2
	MOV	ts+16(FP), A3
	MOV	addr2+24(FP), A4
	MOVW	val3+32(FP), A5
	MOV	$SYS_futex, A7
	ECALL
	MOVW	A0, ret+40(FP)
	RET

	MOVV	$SYS_madvise, R11
	SYSCALL
	MOVW	R4, ret+24(FP)
	RET

// func futex(addr unsafe.Pointer, op int32, val uint32, ts, addr2 unsafe.Pointer, val3 uint32) int32
TEXT runtime·futex(SB),NOSPLIT|NOFRAME,$0
	MOVV	addr+0(FP), R4
	MOVW	op+8(FP), R5
	MOVW	val+12(FP), R6
	MOVV	ts+16(FP), R7
	MOVV	addr2+24(FP), R8
	MOVW	val3+32(FP), R9
	MOVV	$SYS_futex, R11
	SYSCALL
	MOVW	R4, ret+40(FP)
	RET

		case PrefixPT:
			op += ",pt"
			continue

		case PrefixAddrSize, PrefixAddr16, PrefixAddr32:
			// For unknown reasons, if the addr16 prefix is repeated,
			// libopcodes displays all but the last as addr32, even though
			// the addressing form used in a memory reference is clearly
			// still 16-bit.
			n := 32
			if inst.Mode == 32 {
				n = 16
			}
			numAddr++

      %outputs_4, %control_5 = tf_executor.island wraps "tf.RandomUniform"(%outputs_2) {device = "", seed = 87654321 : i64, seed2 = 0 : i64} : (tensor<1xi32>) -> tensor<*xf32>
      %outputs_6, %control_7 = tf_executor.island wraps "tf.AddV2"(%arg2, %outputs) {device = ""} : (tensor<i32>, tensor<i32>) -> tensor<*xi32>
      %outputs_8, %control_9 = tf_executor.island wraps "tf.Identity"(%outputs_6) {device = ""} : (tensor<*xi32>) -> tensor<*xi32>

  RunTest({x1, x2}, {x1_shape, x2_shape}, {y}, {x1_shape});
}

TEST_F(NaryGradTest, AddV2) {
  TensorShape x1_shape({3, 2, 5});
  TensorShape x2_shape({2, 5});
  auto x1 = Placeholder(scope_, DT_FLOAT, Placeholder::Shape(x1_shape));
  auto x2 = Placeholder(scope_, DT_FLOAT, Placeholder::Shape(x2_shape));
  auto y = AddV2(scope_, x1, x2);
  RunTest({x1, x2}, {x1_shape, x2_shape}, {y}, {x1_shape});
}

  (TF_TensorScatterUpdateOp $input, $indices,
     (TF_AddOp $updates, (CreateGatherNdOp $updates, $input, $indices)))>;

//===----------------------------------------------------------------------===//
// AddV2 op patterns.
//===----------------------------------------------------------------------===//

// This is mostly safe, we can use i64 but it will break Darwinn case.

	save, restore string
}

func (l *layout) add(op, reg string, size int) {
	l.regs = append(l.regs, regPos{saveOp: op, restoreOp: op, reg: reg, pos: l.stack})
	l.stack += size
}

func (l *layout) add2(sop, rop, reg string, size int) {
	l.regs = append(l.regs, regPos{saveOp: sop, restoreOp: rop, reg: reg, pos: l.stack})
	l.stack += size
}

func (l *layout) addSpecial(save, restore string, size int) {