{"1.38.1", rel2, "go1.39.2"},
	{"1.38.1", relRC, "go1.39.2"},
	{"1.39", rel0, "go1.39.2"},
	{"1.39", rel2, "go1.39.2"},
	{"1.39", relRC, "go1.39.2"},
	{"1.39.2", rel0, "go1.39.2"},
	{"1.39.2", rel2, "go1.39.2"},
	{"1.39.2", relRC, "go1.39.2"},
	{"1.39.3", rel0, "go1.40.0"},
	{"1.39.3", rel2, "go1.40.2"},
	{"1.39.3", relRC, "go1.40.2"},
	{"1.40", rel0, "go1.40.0"},
	{"1.40", rel2, "go1.40.2"},

	//				0: R_LARCH_PCALA_HI20	local.moduledata
	o(0x1a000004)
	rel, _ := initfunc.AddRel(objabi.R_LOONG64_ADDR_HI)
	rel.SetOff(0)
	rel.SetSiz(4)
	rel.SetSym(ctxt.Moduledata)

	//	4:	02c00084	addi.d	$a0, $a0, 0
	//				4: R_LARCH_PCALA_LO12	local.moduledata
	o(0x02c00084)
	rel2, _ := initfunc.AddRel(objabi.R_LOONG64_ADDR_LO)
	rel2.SetOff(4)
	rel2.SetSiz(4)
	rel2.SetSym(ctxt.Moduledata)

	//	8:	50000000	b	0

    attrs=['act: {"", "RELU", "RELU6", "TANH"} = ""'],
    derived_attrs=['T: {float, int8}'],
    outputs=['o: T'])
def _composite_fully_connected(input_, filter_, bias, act):
  res = tf.raw_ops.MatMul(
      a=input_, b=filter_, transpose_a=False, transpose_b=True)
  res = tf.raw_ops.Add(x=res, y=bias)
  if act == 'RELU':
    return tf.raw_ops.Relu(features=res)
  elif act == 'RELU6':

' ##################################

!define Rel_(e_alias1, e_alias2, e_label, e_direction="") e_alias1 e_direction e_alias2 : "===e_label"
!define Rel_(e_alias1, e_alias2, e_label, e_techn, e_direction="") e_alias1 e_direction e_alias2 : "===e_label\n//<size:TECHN_FONT_SIZE>[e_techn]</size>//"

!define Rel(e_from,e_to, e_label) Rel_(e_from,e_to, e_label, "-->")
!define Rel(e_from,e_to, e_label, e_techn) Rel_(e_from,e_to, e_label, e_techn, "-->")

// Since the tf.AddV2 op has two uses, we have a _FusedBatchNormEx without the
// Relu activation and we only fuse the add.
// CHECK-NEXT: %[[Y:[a-z0-9]*]], {{.*}}_FusedBatchNormEx
// CHECK-NEXT: %[[relu:[a-z0-9]*]] ={{.*}}Relu"(%[[Y]]
// CHECK-NEXT: return %[[relu]]

        dim {
          size: 1
        }
      }
    }
  }
}
node {
  name: "Relu"
  op: "Relu"
  input: "input"
  device: "/device:CPU:0"
  attr {
    key: "T"
    value {
      type: DT_FLOAT
    }
  }
}
node {
  name: "out"
  op: "IdentityN"
  input: "Relu"
  input: "Relu"
  attr: {
    key: "T"
    value: {
      list: {
        type: DT_FLOAT

    // CHECK: tfl.add
    // CHECK-SAME: tac.skip_target_annotation
    %1 = "tfl.add"(%arg0, %0) {fused_activation_function = "RELU"} : (tensor<1xf32>, tensor<1xf32>) -> tensor<1xf32>
    // CHECK: tfl.relu
    // CHECK-SAME: tac.skip_target_annotation
    %2 = "tfl.relu"(%arg0) : (tensor<1xf32>) -> tensor<1xf32>
    func.return
  }
}

// -----

Russ Cox <******@****.***> 1410149331 -0400

  // CHECK: device = "cpu"
  %2 = "tf.Relu"(%1) {T = f32, _output_shapes = ["tfshape$dim { size: 3 } dim { size: 3 }"], device = "cpu"} : (tensor<3x3xf32>) -> tensor<3x3xf32>
  // CHECK: device = "xpu"
  %3 = "tf.Relu"(%2) {T = f32, _output_shapes = ["tfshape$dim { size: 3 } dim { size: 3 }"]} : (tensor<3x3xf32>) -> tensor<3x3xf32>
  func.return %3 : tensor<3x3xf32>

  // will be preserved since it has uses.
  %2 = "tf.MyCustomOp"(%1, %0) {fused_activation_function = "RELU", int_attr = 2 : i32}  : (tensor<4xf32>, tensor<4xf32>) -> tensor<4xf32>
  // will be preserved since it has side-effect.
  "tf.MyCustomOp"(%1, %0) {fused_activation_function = "RELU", int_attr = 2 : i32}  : (tensor<4xf32>, tensor<4xf32>) -> tensor<4xf32>
  func.return %2 : tensor<4xf32>