// CHECK-DAG: %[[SHIFT:.*]] = "tf.AddV2"(%[[CAST]], %[[HALF_RANGE]])
  // CHECK-DAG: %[[DRANGE:.*]] = "tf.Sub"(%arg2, %arg1)
  // CHECK-DAG: %[[SCALE:.*]] = "tf.Div"(%[[DRANGE]], %[[C255:.*]])
  // CHECK-DAG: %[[SS:.*]] = "tf.Mul"(%[[SHIFT]], %[[SCALE]])
  // CHECK-DAG: %[[RESULT:.*]] = "tf.AddV2"(%[[SS]], %arg1)
  %0 = "tf.Dequantize"(%arg0, %min_range, %max_range) : (tensor<2x3x!tf_type.qint8>, tensor<f32>, tensor<f32>) -> tensor<2x3xf32>

func.func @main(tensor<4xf32>) -> tensor<4xf32> {
^bb0(%arg0: tensor<4xf32>):
  %0 = "tfl.pseudo_const" () {name = "Const", value = dense<1.0> : tensor<4xf32>} : () -> tensor<4xf32>
  %1 = "tfl.mul"(%arg0, %0) {fused_activation_function = "NONE", name = "mul"} : (tensor<4xf32>, tensor<4xf32>) -> tensor<4xf32>
  %2:2 = "tf.MyCustomOp"(%1, %0) {name = "MyCustomOp"} : (tensor<4xf32>, tensor<4xf32>) -> (tensor<4xf32>, tensor<3xf32>)

	SLD	$32, R1, R1
	LDGR	R1, F2
	WFMDB	V2, V2, V4
	WFMDB	V4, V0, V6
	WFMSDB	V4, V6, V2, V4
	FMOVD	40(R9), F6
	FMSUB	F1, F4, F2
	FMOVD	32(R9), F4
	WFMDB	V2, V2, V3
	FMOVD	24(R9), F1
	FMUL	F3, F0
	FMOVD	16(R9), F3
	WFMADB	V2, V0, V5, V2
	FMOVD	8(R9), F5
	FMADD	F6, F2, F4
	WFMADB	V2, V1, V3, V1
	WFMDB	V2, V2, V6
	FMOVD	0(R9), F3
	WFMADB	V4, V6, V1, V4
	WFMADB	V2, V5, V3, V2

        type: DT_FLOAT
      }
    }
    node_def {
      name: "tf.Mul"
      op: "Mul"
      input: "cond_false_arg0"
      input: "cond_false_arg3"
      attr {
        key: "T"
        value {
          type: DT_FLOAT
        }
      }
      experimental_debug_info {
        original_node_names: "tf.Mul"
      }
    }
    ret {
      key: "cond_false_ret"

	SUB	$2, R0
loop:
	CBZ	R0, done
	LDP.P	32(R2), (R5, R6)
	LDP	-16(R2), (R7, R8)

	MUL	R3, R5, R10
	UMULH	R3, R5, R11
	ADDS	R4, R10
	MUL	R3, R6, R12
	UMULH	R3, R6, R13
	ADCS	R11, R12

	MUL	R3, R7, R14
	UMULH	R3, R7, R15
	ADCS	R13, R14
	MUL	R3, R8, R16
	UMULH	R3, R8, R17
	ADCS	R15, R16
	ADC	$0, R17, R4

	STP.P	(R10, R12), 32(R1)
	STP	(R14, R16), -16(R1)
	SUB	$4, R0

    %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>

  %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

  %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>