{name: "FNMSUBS", argLength: 3, reg: fp31, asm: "FNMSUBS"}, // -arg0 + (arg1 * arg2)
		{name: "FNMSUBD", argLength: 3, reg: fp31, asm: "FNMSUBD"}, // -arg0 + (arg1 * arg2)
		{name: "MADD", argLength: 3, reg: gp31, asm: "MADD"},       // +arg0 + (arg1 * arg2)
		{name: "MADDW", argLength: 3, reg: gp31, asm: "MADDW"},     // +arg0 + (arg1 * arg2), 32-bit
		{name: "MSUB", argLength: 3, reg: gp31, asm: "MSUB"},       // +arg0 - (arg1 * arg2)

    %stack_push:2 = tf_executor.island wraps "tf.StackPushV2"(%stack_handle#0, %arg1) : (!tf_res, tensor<f32>) -> tensor<f32>
    %add:2 = tf_executor.island wraps "tf.Add"(%arg1, %arg1) : (tensor<f32>, tensor<f32>) -> tensor<f32>
    %stack_push2:2 = tf_executor.island(%stack_push#1) wraps "tf.StackPushV2"(%stack_handle#0, %add#0) : (!tf_res, tensor<f32>) -> tensor<f32>

  %0 = "tf.BatchMatMul"(%arg0, %arg1) : (tensor<4x5xf32>, tensor<5x6xf32>) -> tensor<4x6xf32>
  func.return %0 : tensor<4x6xf32>

  // CHECK-LABEL: batchMatMulMatrix
  // CHECK: %[[MATMUL_1:.*]] = "tf.MatMul"(%arg0, %arg1) <{grad_a = false, grad_b = false, transpose_a = false, transpose_b = false}> : (tensor<4x5xf32>, tensor<5x6xf32>) -> tensor<4x6xf32>

  %0 = "tfl.dequantize"(%arg0) : (tensor<2x3x5x!quant.uniform<u8:f32, 0.1>>) -> tensor<2x3x5xf32>
  %1 = "tfl.slice"(%0, %arg1, %arg2) : (tensor<2x3x5xf32>, tensor<3xi32>, tensor<3xi32>) -> tensor<?x3x5xf32>
  func.return %1 : tensor<?x3x5xf32>

// CHECK: %0 = "tfl.dequantize"(%arg0)
// CHECK: %1 = "tfl.slice"(%0, %arg1, %arg2)
// CHECK: %2 = "tfl.quantize"(%1)

^bb0(%arg0: tensor<256x32x32x3xf32>, %arg1: tensor<3x3x3x16xf32>, %arg2: tensor<256x3x32x32xf32>) :
   // OK
   %0 = "tf.Conv2D"(%arg0, %arg1) {T = "tfdtype$DT_FLOAT", data_format = "NHWC", dilations = [1, 2, 3, 1], padding = "SAME", strides = [1, 4, 5, 1]} : (tensor<256x32x32x3xf32>, tensor<3x3x3x16xf32>) -> tensor<256x8x7x16xf32>
   // Unsupported data format

func.func private @"map/while_cond"(%arg0: tensor<i32>, %arg1: tensor<i32>, %arg2: tensor<!tf_type.variant<tensor<*xf32>>>, %arg3: tensor<?xf32>) -> tensor<i1> {
  %cst = "tf.Const"() {device = "/job:localhost/replica:0/task:0/device:CPU:0", value = dense<3> : tensor<i32>} : () -> tensor<i32>
  %0 = "tf.Less"(%arg0, %cst) {device = "/job:localhost/replica:0/task:0/device:CPU:0"} : (tensor<i32>, tensor<i32>) -> tensor<i1>

  func.return %0 : tensor<3x4xi1>
}

// CHECK-LABEL: func @fill
// CHECK-SAME: (%[[ARG0:.*]]: tensor<*xi64>, %[[ARG1:.*]]: tensor<*xf32>)
func.func @fill(%arg0: tensor<*xi64>, %arg1: tensor<*xf32>) -> tensor<*xf32> {
  // CHECK: "tf.BroadcastTo"(%[[ARG1]], %[[ARG0]])
  %0 = "tf.Fill"(%arg0, %arg1) : (tensor<*xi64>, tensor<*xf32>) -> tensor<*xf32>
  func.return %0 : tensor<*xf32>
}

    %0 = "tf.MatMul"(%arg0, %arg1) {attr_map = "0:transpose_a,1:transpose_b", device = "", transpose_a = false, transpose_b = false} : (tensor<2x2xf32>, tensor<2x2xf32>) -> tensor<2x2xf32>
    return %0 : tensor<2x2xf32>
  }

// CHECK-LABEL: func @reachable_function
func.func @reachable_function(%arg0: tensor<*x!tf_type.resource<tensor<2x8xi32>>>, %arg1: tensor<i32>) {
  // CHECK-NOT: "tf.AssignAddVariableOp"
  "tf.AssignAddVariableOp"(%arg0, %arg1) {dtype = "tfdtype$DT_INT32"} : (tensor<*x!tf_type.resource<tensor<2x8xi32>>>, tensor<i32>) -> ()
  "tf.OpA"(%arg0, %arg1) {f = @reachable_from_reachable_function} : (tensor<*x!tf_type.resource<tensor<2x8xi32>>>, tensor<i32>) -> ()
  func.return

  "tf.AssignVariableOp"(%arg1, %add0) : (tensor<*x!tf_type.resource<tensor<f32>>>, tensor<f32>) -> ()
  %read1 = "tf.ReadVariableOp"(%arg1) : (tensor<*x!tf_type.resource<tensor<f32>>>) -> tensor<f32>
  // CHECK-NEXT: %[[ADD1:.*]] = "tf.AddV2"(%[[ADD0]], %[[ADD0]])
  %add1 = "tf.AddV2"(%read1, %read1) : (tensor<f32>, tensor<f32>) -> tensor<f32>