{"Section64.Name", Field, 0},
		{"Section64.Nreloc", Field, 0},
		{"Section64.Offset", Field, 0},
		{"Section64.Reloff", Field, 0},
		{"Section64.Reserve1", Field, 0},
		{"Section64.Reserve2", Field, 0},
		{"Section64.Reserve3", Field, 0},
		{"Section64.Seg", Field, 0},
		{"Section64.Size", Field, 0},
		{"SectionHeader", Type, 0},
		{"SectionHeader.Addr", Field, 0},
		{"SectionHeader.Align", Field, 0},

  %1 = "tf.Transpose"(%arg0, %0) : (tensor<1x64x112x112xf32>, tensor<4xi32>) -> tensor<1x112x112x64xf32>

  // Compute FusedBatchNormV3 in NHWC format
  %2, %batch_mean, %batch_var, %reserve_1, %reserve_2, %reserve_3
    = "tf.FusedBatchNormV3"(%1, %arg1, %arg1, %arg1, %arg1)
       {
         data_format = "NHWC",
         epsilon = 1.001 : f32,
         exponential_avg_factor = 1.0 : f32,

	// Reserve some Ids for compatible expansion
	tReserved7 = bootstrapType("_reserved1", (*struct{ r7 int })(nil))
	tReserved6 = bootstrapType("_reserved1", (*struct{ r6 int })(nil))
	tReserved5 = bootstrapType("_reserved1", (*struct{ r5 int })(nil))
	tReserved4 = bootstrapType("_reserved1", (*struct{ r4 int })(nil))
	tReserved3 = bootstrapType("_reserved1", (*struct{ r3 int })(nil))

pkg syscall (darwin-arm64), type IfData struct, Recvquota uint8
pkg syscall (darwin-arm64), type IfData struct, Recvtiming uint32
pkg syscall (darwin-arm64), type IfData struct, Reserved1 uint32
pkg syscall (darwin-arm64), type IfData struct, Reserved2 uint32
pkg syscall (darwin-arm64), type IfData struct, Type uint8
pkg syscall (darwin-arm64), type IfData struct, Typelen uint8
pkg syscall (darwin-arm64), type IfData struct, Unused1 uint8

  // CHECK-SAME: <{value = dense<[0, 2, 3, 1]> : tensor<4xi64>}>
  // CHECK: %[[RES_TRANSPOSE:[0-9]*]] = "tf.Transpose"(%y, %[[RES_PERM]])
  // CHECK: return %[[RES_TRANSPOSE]]

  %y, %batch_mean, %batch_var, %reserve_1, %reserve_2, %reserve_3
    = "tf.FusedBatchNormV3"(%arg0, %arg1, %arg1, %arg1, %arg1)
       {
         data_format = "NHWC",
         epsilon = 1.001 : f32,
         exponential_avg_factor = 1.0 : f32,

  // CHECK-SAME: <{value = dense<[0, 3, 1, 2]> : tensor<4xi64>}>
  // CHECK: %[[RES_TRANSPOSE:[0-9]*]] = "tf.Transpose"(%y, %[[RES_PERM]])
  // CHECK: return %[[RES_TRANSPOSE]]

  %y, %batch_mean, %batch_var, %reserve_1, %reserve_2, %reserve_3
    = "tf.FusedBatchNormV3"(%arg0, %arg1, %arg1, %arg1, %arg1)
       {
         data_format = "NCHW",
         epsilon = 1.001 : f32,
         exponential_avg_factor = 1.0 : f32,
         is_training = true

//
// Developed at SunPro, a Sun Microsystems, Inc. business.
// Permission to use, copy, modify, and distribute this
// software is freely granted, provided that this notice
// is preserved.
// ====================================================
//
// __ieee754_remainder(x,y)
// Return :
//      returns  x REM y  =  x - [x/y]*y  as if in infinite
//      precision arithmetic, where [x/y] is the (infinite bit)

//
// Developed at SunPro, a Sun Microsystems, Inc. business.
// Permission to use, copy, modify, and distribute this
// software is freely granted, provided that this notice
// is preserved.
// ====================================================
//
//
// asinh(x)
// Method :
//	Based on
//	        asinh(x) = sign(x) * log [ |x| + sqrt(x*x+1) ]
//	we have
//	asinh(x) := x  if  1+x*x=1,

	}
	if p.typeList[i] == reserved {
		p.errorf("internal error: %d already reserved in parseSavedType", i)
	}
	if p.typeList[i] == nil {
		p.reserve(i)
		p.parseTypeSpec(pkg, append(nlist, i))
	}
	if p.typeList[i] == nil || p.typeList[i] == reserved {
		p.errorf("internal error: parseSavedType(%d,%v) reserved/nil", i, nlist)
	}
}

    (*f.mutable_attr())[attr.first] = attr.second;
  }

  std::vector<Output> func_inputs;
  std::vector<DataType> input_dtypes;
  const int num_inputs = op.num_inputs();
  func_inputs.reserve(num_inputs + grad_inputs.size());
  input_dtypes.reserve(num_inputs);

  for (int i = 0; i < num_inputs; i++) {
    func_inputs.push_back(op.input(i));
    input_dtypes.push_back(op.input_type(i));
  }