testCases = append(testCases,
		testCase{capitalLetters, "brad", "BrAd"},
		testCase{capitalLetters, Repeat("a", (32<<10)+123), Repeat("A", (32<<10)+123)},
		testCase{capitalLetters, "", ""},

		testCase{inc, "brad", "csbe"},
		testCase{inc, "\x00\xff", "\x01\x00"},
		testCase{inc, "", ""},

		testCase{NewReplacer("a", "1", "a", "2"), "brad", "br1d"},
	)

	// repeat maps "a"->"a", "b"->"bb", "c"->"ccc", ...

# CL 134210043 archive/zip: add Writer.Flush, Brad Fitzpatrick <******@****.***>
pkg archive/zip, method (*Writer) Flush() error

# CL 97140043 compress/flate: add Reset() to allow reusing large buffers to compress multiple buffers, James Robinson <******@****.***>
pkg compress/flate, type Resetter interface { Reset }
pkg compress/flate, type Resetter interface, Reset(io.Reader, []uint8) error
pkg compress/zlib, type Resetter interface { Reset }

    auto multiply_result = Multiply(scope, subtraction_result, logits_softmax);
    grad = Add(scope, grad, multiply_result);
  }
  auto minus_log_softmax = Multiply(scope, LogSoftmax(scope, logits), -1.0f);
  grad_outputs->push_back(grad);
  grad_outputs->push_back(BroadcastMul(scope, grad_loss, minus_log_softmax));
  return scope.status();
}

     (CreateTFReadVariableOp $src_op, $grad, $ms_resource),
     (TF_AddV2Op:$ms_new
       (TF_MulOp
         (TF_MulOp $grad, $grad),
         (TF_SubOp $one, $rho)
       ),
       (TF_MulOp
          (CreateTFReadVariableOp $src_op, $grad, $ms_resource),
          $rho
       )
     ),
     (TF_AssignVariableOp $ms_resource, $ms_new, (CreateConstBoolAttrFalse)),
     // mg = grad * (one - rho) + mg * rho;

    /* Given upstream grad U and a Sub op A-B, the gradients are:
     *
     *    dA =  U
     *    dB = -U
     *
     */

    // Grad for A
    DCHECK(grad_outputs[0]);
    grad_inputs[0] = grad_outputs[0];
    grad_inputs[0]->Ref();

    // Grad for B
    // negate the upstream grad
    std::string name = "Neg_Sub_Grad_B";
    TF_RETURN_IF_ERROR(

LogicalResult HandleTensorArrayGradV3Op(
    TF::TensorArrayGradV3Op grad,
    llvm::SmallDenseMap<Value, TensorArrayStats>* stats) {
  auto local_var = grad.getHandle();
  OpBuilder builder(grad);
  Value grad_var;
  auto sit = stats->find(local_var);
  if (sit == stats->end()) return grad.emitOpError("unknown tensor array");
  auto emplace_res =
      sit->getSecond().grads.try_emplace(grad.getSource().str(), Value());
  if (!emplace_res.second) {

  tensorflow::Output y = op.input(1);
  if (DataTypeIsComplex(grad.type())) {
    x = Conj(scope, x);
    y = Conj(scope, y);
  }

  const auto x_shape = Shape(scope, x);
  const auto y_shape = Shape(scope, y);
  Output grad_x =
      EinsumGradWrt(scope, grad, y, x_shape, x_subs, y_subs, output_subs);
  Output grad_y =
      EinsumGradWrt(scope, grad, x, y_shape, y_subs, x_subs, output_subs);

        return s;
      }
    } else {
      if (!persistent_) {
        trace.backward_function_deleter(trace.backward_function);
      }
      for (Gradient* grad : out_gradients) {
        if (grad != nullptr) {
          vspace.DeleteGradient(grad);
        }
      }
    }
    for (int i = 0, end = in_gradients.size(); i < end; ++i) {
      const int64_t id = trace.input_tensor_id[i];

  if (TF_GetCode(status) != TF_OK) return;
  if (!grad) return;

  status->status = g->graph.AddFunctionDef(grad->record->fdef(),
                                           grad->record->stack_traces());
  if (TF_GetCode(status) != TF_OK) return;

  tensorflow::GradientDef gdef;
  gdef.set_function_name(func->record->fdef().signature().name());
  gdef.set_gradient_func(grad->record->fdef().signature().name());

  func.func @max_pool_grad_valid(%orig_input: tensor<10x24x24x64xf32>, %orig_output: tensor<10x12x12x64xf32>, %grad: tensor<10x12x12x64xf32>) -> tensor<10x24x24x64xf32> {
    // CHECK: %[[ZERO:.*]] = mhlo.constant dense<0.000000e+00> : tensor<f32>
    // CHECK: %[[RESULT:.*]] = "mhlo.select_and_scatter"(%[[INPUT]], %[[GRAD]], %[[ZERO]]) <{