EXPECT_EQ(num_compute_calls, 5);
  cache2.LookupOrCompute("c", &value, compute_func, &status);
  TF_EXPECT_OK(status.status);
  EXPECT_EQ(value, 2);
  EXPECT_EQ(num_compute_calls, 5);
  cache2.LookupOrCompute("d", &value, compute_func, &status);
  TF_EXPECT_OK(status.status);
  EXPECT_EQ(value, 3);
  EXPECT_EQ(num_compute_calls, 5);

  // Re-read "a", ensure it is re-computed.
  cache2.LookupOrCompute("a", &value, compute_func, &status);

std::string getTF_OutputDebugString(TF_Output node);

// Set whether to propagate assigned device information when constructing a new
// Graph from a GraphDef. By default assigned device information is not copied
// and is re-computed by the runtime.
inline void TF_ImportGraphDefOptionsSetPropagateDeviceSpec(
    TF_ImportGraphDefOptions* opts, unsigned char propagate_device_spec) {
  opts->opts.propagate_device_spec = propagate_device_spec;
}

// is used as an input to an op (so we know when we're done computing gradients
// for that Tensor). We also count, for each tape entry, how many of its output
// Tensors need gradients to be computed (Tensors which are not used do not need
// any gradients to be computed).
//
// Finally, we start a backprop stack with a set of tape entries for which we
// have all gradients available. This set usually is a subset of the set of

// will block till the operation that produces `h` has completed.
TF_CAPI_EXPORT extern const char* TFE_TensorHandleDeviceName(
    TFE_TensorHandle* h, TF_Status* status);

// Returns the name of the device in whose memory `h` resides.
//
// This function will block till the operation that produces `h` has completed.
TF_CAPI_EXPORT extern const char* TFE_TensorHandleBackingDeviceName(

                 TF_Message(first_bad_status.get()));
    return result;
  }
  // For each output of the original operation, pack the per-device
  // TensorHandles we've computed into a single parallel TensorHandle.
  std::vector<std::unique_ptr<ParallelTensor>> per_device_outputs;
  per_device_outputs.reserve(first_op_output_count);
  for (int i = 0; i < first_op_output_count; ++i) {

  TF_Status* s = TF_NewStatus();
  TF_OpKernelConstruction_GetAttrType(ctx, "foobar", &type, s);
  TF_DeleteStatus(s);
  return NULL;
}

// A compute function. This will never actually get called in this test, it's
// just nice to know that it compiles.
void compute(void* kernel, TF_OpKernelContext* ctx) {
  TF_Tensor* input;
  TF_Status* s = TF_NewStatus();
  TF_GetInput(ctx, 0, &input, s);
  TF_DeleteTensor(input);

      : forward_outputs_(f_outputs) {
    for (auto output : forward_outputs_) {
      if (output) {
        output->Ref();
      }
    }
  }

  Status Compute(AbstractContext* ctx,
                 absl::Span<AbstractTensorHandle* const> grad_outputs,
                 absl::Span<AbstractTensorHandle*> grad_inputs) override {
    AbstractTensorHandle* upstream_grad = grad_outputs[0];

 
-     The choice of symbol is arbitrary. The static address we obtain
-     by constructing a non GOT reference to the symbol, the dynamic
-     address of the symbol we compute using adrp/add to compute the
-     symbol's address relative to the PC. */
-
-  ElfW(Addr) static_addr;
-  ElfW(Addr) dynamic_addr;
-
-  asm ("					\n\
-	adrp	%1, _dl_start;			\n\

  TF_DeleteBuffer(op_list_buf);
}

class DummyKernel : public tensorflow::OpKernel {
 public:
  explicit DummyKernel(tensorflow::OpKernelConstruction* context)
      : OpKernel(context) {}
  void Compute(tensorflow::OpKernelContext* context) override {}
};

// Test we can query kernels
REGISTER_OP("TestOpWithSingleKernel")
    .Input("a: float")
    .Input("b: float")
    .Output("o: float");

class TestUnavailableErrorOp : public tensorflow::OpKernel {
 public:
  explicit TestUnavailableErrorOp(tensorflow::OpKernelConstruction* ctx)
      : tensorflow::OpKernel(ctx) {}
  void Compute(tensorflow::OpKernelContext* ctx) override {
    ctx->SetStatus(tensorflow::errors::Unavailable("Test error."));
  }
};
REGISTER_KERNEL_BUILDER(
    Name("TestNonCommUnavailable").Device(tensorflow::DEVICE_DEFAULT),