%2 = corert.create_dense_tensor.i32 {shape = [0], value = []}
  // CHECK-NEXT: %[[RESULT_3:.*]] = corert.executeop(%[[RESULT_0]]) "tf.AddV2"(%[[RESULT_2]], %[[RESULT_2]])
  %3 = corert.executeop(%0) "tf.AddV2"(%2, %2) {T = f32, device = "/device:CPU:0"} : 1
  // CHECK-NEXT: %[[RESULT_4:.*]] = tfrt.get_device %[[ARG_0]] {device_name = "/device:GPU:0"}

		addr1 := v1.UnsafeAddr()
		addr2 := v2.UnsafeAddr()
		if addr1 > addr2 {
			// Canonicalize order to reduce number of entries in visited.
			addr1, addr2 = addr2, addr1
		}

		// Short circuit if references are identical ...
		if addr1 == addr2 {
			return true
		}

		// ... or already seen
		typ := v1.Type()
		v := visit{addr1, addr2, typ}
		if visited[v] {
			return true
		}

  // CHECK: %[[ARG1_TRANSPOSE:[0-9]*]] = "tf.Transpose"(%arg1, %[[ARG_PERM]])
  // CHECK: %[[ADD:[0-9]*]] = "tf.AddV2"(%[[ARG0_TRANSPOSE]], %[[ARG1_TRANSPOSE]]) {{.*}} tensor<1x8x4x4xf32>
  // CHECK: return %[[ADD]]

  %0 = "tf.AddV2"(%arg0, %arg1) : (tensor<1x4x4x8xf32>, tensor<1x4x4x8xf32>) -> tensor<1x4x4x8xf32>
  %1 = "tf.Const"() {value = dense<[0, 3, 1, 2]> : tensor<4xi32>} : () -> tensor<4xi32>

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

func.func @else(%x: tensor<i32>, %y: tensor<i32>) -> (tensor<i32>) {
  %0 = "tf.Const"() {value = dense<1> : tensor<i32> } : () -> tensor<i32>
  %1 = "tf.AddV2"(%x, %0) : (tensor<i32>, tensor<i32>) -> tensor<i32>
  %2 = "tf.AddV2"(%y, %1) : (tensor<i32>, tensor<i32>) -> tensor<i32>

//go:build arm || wasm

// Export some functions via linkname to assembly in sync/atomic.
//
//go:linkname And32
//go:linkname Or32
//go:linkname And64
//go:linkname Or64
//go:linkname Anduintptr
//go:linkname Oruintptr

package atomic

import _ "unsafe" // For linkname

//go:nosplit
func And32(ptr *uint32, val uint32) uint32 {
	for {
		old := *ptr
		if Cas(ptr, old, old&val) {
			return old
		}
	}

// CHECK-EMPTY:
// CHECK-NEXT:        },
// CHECK-NEXT:        has_rank: true
// CHECK-NEXT:      }, {
// CHECK-NEXT:        shape: [ 3, 2 ],
// CHECK-NEXT:        buffer: 2,
// CHECK-NEXT:        name: "tf.AddV2",
// CHECK-NEXT:        quantization: {
// CHECK-EMPTY:
// CHECK-NEXT:        },
// CHECK-NEXT:        has_rank: true
// CHECK-NEXT:      } ],
// CHECK-NEXT:      inputs: [ 0 ],
// CHECK-NEXT:      outputs: [ 1 ],

  %add = "tf.AddV2"(%arg0, %y#0) : (tensor<8x8x8x8xf32>, tensor<8x8x8x8xf32>) -> tensor<8x8x8x8xf32>
  %relu = "tf.Relu"(%add) : (tensor<8x8x8x8xf32>) -> tensor<8x8x8x8xf32>
  func.return %relu : tensor<8x8x8x8xf32>
}

  // CHECK: [[handle:%.*]] = "tf.VarHandleOp"()
  // CHECK: "tf.ReadVariableOp"([[handle]])
  %0 = "tf.ReadVariableOp"(%arg1) {device = "/device:CPU:0"} : (tensor<*x!tf_type.resource>) -> tensor<1x3xf32>
  %1 = "tf.AddV2"(%arg0, %0) {device = "/device:CPU:0"} : (tensor<1x3xf32>, tensor<1x3xf32>) -> tensor<1x3xf32>
  %2 = "tf.Identity"(%1) {device = "/device:CPU:0"} : (tensor<1x3xf32>) -> tensor<1x3xf32>
  func.return %2 : tensor<1x3xf32>
}

// CANON:             [[VAL_14:%.*]] = "tf.AddV2"([[VAL_12]], [[VAL_4]])
// CANON:             [[VAL_15:%.*]] = "tf.AddV2"([[VAL_13]], [[VAL_5]])
// CANON:             [[VAL_16:%.*]] = "tf.AddV2"([[VAL_11]], [[VAL_4]])
// CANON:             "tfl.yield"([[VAL_16]], [[VAL_14]], [[VAL_15]]) : (tensor<*xi32>, tensor<*xi32>, tensor<*xf32>) -> ()

// CHECK-DAG: %[[cst_0:.*]] = "tf.Const{{.*}} dense<1.200000e+01> : tensor<16xf32>
// CHECK-NEXT: %[[conv:.*]] = "tf.Conv2D"(%arg0, %[[cst]])
// CHECK-NEXT: %[[bias:.*]] = "tf.AddV2"(%[[conv]], %[[cst_0]])
// CHECK-NEXT: return %[[bias]] : tensor<256x8x7x16xf32>
}

// CHECK-LABEL: convaddv2mul
func.func @convaddv2mul(%arg: tensor<256x32x32x3xf32>) -> tensor<256x8x7x16xf32> {