op: "Mul"
  input: "loss_1/fc1000/bias/Regularizer/mul/x"
  input: "loss_1/ArithmeticOptimizer/HoistCommonFactor_Add_AddN"
  device: "/job:localhost/replica:0/task:0/device:GPU:0"
  attr {
    key: "T"
    value {
      type: DT_FLOAT
    }
  }
}
node {
  name: "bn_conv1_1/AssignMovingAvg/mul"
  op: "Mul"
  input: "bn_conv1_1/AssignMovingAvg/sub_1"

    %cast = "tf.Cast"(%accumulation) {Truncate = false} : (tensor<*xi32>) -> tensor<*xf32>
    %mul = "tf.Mul"(%cast, %rescale_factor) : (tensor<*xf32>, tensor<*xf32>) -> tensor<*xf32>
    %add = "tf.AddV2"(%mul, %float_out_zp) : (tensor<*xf32>, tensor<*xf32>) -> tensor<*xf32>
    func.return %add : tensor<*xf32>
  }

    %5 = "tf.AddV2"(%4, %arg2) : (tensor<1x3x2x2xi32>, tensor<2xi32>) -> tensor<1x3x2x2xi32>
    %6 = "tf.Mul"(%arg3, %arg5) : (tensor<f32>, tensor<f32>) -> tensor<f32>
    %7 = "tf.Div"(%6, %arg9) : (tensor<f32>, tensor<f32>) -> tensor<f32>
    %8 = "tf.Cast"(%5) {Truncate = false} : (tensor<1x3x2x2xi32>) -> tensor<1x3x2x2xf32>
    %9 = "tf.Mul"(%7, %8) : (tensor<f32>, tensor<1x3x2x2xf32>) -> tensor<1x3x2x2xf32>

	MOV	3*8(X7), X17	// x[3]

	MULHU	X8, X6, X9	// z_hi[0] = x[0] * y
	MUL	X8, X6, X8	// z_lo[0] = x[0] * y
	ADD	X8, X10, X21	// z_lo[0] = x[0] * y + z[0]
	SLTU	X8, X21, X22
	ADD	X9, X22, X9	// z_hi[0] = x[0] * y + z[0]
	ADD	X21, X29, X10	// z_lo[0] = x[0] * y + z[0] + c
	SLTU	X21, X10, X22
	ADD	X9, X22, X29	// next c

	MULHU	X11, X6, X12	// z_hi[1] = x[1] * y
	MUL	X11, X6, X11	// z_lo[1] = x[1] * y

  %a = "tf.Neg"(%arg0) : (tensor<4xf32>) -> tensor<4xf32>
  %b = "tf.Abs"(%arg0) : (tensor<4xf32>) -> tensor<4xf32>
  // Verify that the Div and Mul are hoisted out of the body
  // CHECK: "tf.Div"
  // CHECK: constant dense<2.200000e+01>
  // CHECK: "tf.Mul"
  // CHECK: "tf.WhileRegion"
  // Verify that Add and Sub is not hoisted out
  // CHECK: "tf.Add"
  // CHECK: "tf.Sub"
  %0:2 = "tf.WhileRegion"(%arg0, %arg1) (

    %2 = "tf.MatMul"(%arg0, %1) {attr_map = "0:transpose_a,1:transpose_a", device = "", transpose_a = false, transpose_b = false} : (tensor<1x2x2x2xf32>, tensor<2x1024xf32>) -> tensor<*xf32>
    %3 = "tf.Mul"(%2, %cst_0) : (tensor<*xf32>, tensor<f32>) -> tensor<*xf32>
    return %3 : tensor<*xf32>
  }

// CHECK-LABEL: func @not_propagate_matmul

// CHECK: (%[[ARG:.*]]: tensor<4xf32>) -> (tensor<4xf32>, tensor<4xf32>, tensor<4xf32>)
func.func @main(%arg0: tensor<4xf32>) -> tensor<4xf32> attributes {tf.entry_function = {inputs = "mul"}} {

// CHECK:         %[[SUB:.*]] = s32[] subtract(s32[] %[[CONVERT_1]],
// CHECK:         %[[CONVERT_2:.*]] = f32[] convert(s32[] %[[SUB]])
// CHECK:         %[[MUL:.*]] = f32[] multiply(f32[] %[[CONVERT_2]],

  %1 = "tfl.squared_difference"(%arg0, %0) {fused_activation_function = "NONE"} : (tensor<4xf32>, tensor<4xf32>) -> tensor<4xf32> loc("squared_difference")
  %2 = "tfl.mul"(%arg0, %1) {fused_activation_function = "NONE"} : (tensor<4xf32>, tensor<4xf32>) -> tensor<4xf32> loc("mul")
  %3 = "tfl.div"(%2, %1) {fused_activation_function = "NONE"} : (tensor<4xf32>, tensor<4xf32>) -> tensor<4xf32> loc("div")

  %1 = "tfl.mul"(%0, %cst_2) {fused_activation_function = "NONE"} : (tensor<1x256xbf16>, tensor<32x1x256xbf16>) -> tensor<32x1x256xbf16>
  func.return %1 : tensor<32x1x256xbf16>

// CHECK:  %[[V0:.*]] = "tfl.fully_connected"(%arg0, {{.*}}) <{{{.*}}}> : (tensor<1x10368xbf16>, tensor<256x10368xbf16>, none) -> tensor<1x256xbf16>