virtual serialization: "Error, could not find corresponding entry in derived for base" when serializing non constructible classes

[cz4rs]

Scenario based on serialization of vt::objgroup::holder::HolderBase (required for vt::objgroup::ObjGroupManager component).

Given following class hierarchy:

struct HolderBase {
  checkpoint_virtual_serialize_root()

  virtual ~HolderBase() = default;

  template <typename Serializer>
  void serialize(Serializer& s) {}
};

template <typename ObjT>
struct HolderObjBase : HolderBase {
  checkpoint_virtual_serialize_derived_from(HolderBase)

  virtual ObjT* get() = 0;

  template <typename Serializer>
  void serialize(Serializer& s) {}
};

template <typename ObjT>
struct HolderBasic final : HolderObjBase<ObjT> {
  checkpoint_virtual_serialize_derived_from(HolderObjBase<ObjT>)

  ObjT* get() override { return obj_; }
  ObjT* obj_ = nullptr;

  template <typename Serializer>
  void serialize(Serializer& s) {}
};

When virtual serialization is attempted:

  std::unique_ptr<HolderBase> ptr = std::make_unique<HolderBasic<int>>();
  checkpoint::getMemoryFootprint(ptr);

Test crashes with following error:

-----------------------------------------------------------------------
Assertion failed in Checkpoint library:
-----------------------------------------------------------------------
   Reason: Error, could not find corresponding entry in derived for base
Condition: false
     File: ../src/checkpoint/dispatch/vrt/serializer_registry.h
     Line: 156
     Func: getBaseIdx
-----------------------------------------------------------------------

Backtrace after crash:

(...)
#4  in checkpoint::debug::assertOut (cond=0x5555555f6de6 "false", str=0x5555555f6da8 "Error, could not find corresponding entry in derived for base", 
    file=0x5555555f6a90 "../src/checkpoint/dispatch/vrt/serializer_registry.h", line=156, func=0x5555555f6d97 "getBaseIdx") at ../src/checkpoint/common.h:103
#5  in checkpoint::dispatch::vrt::serializer_registry::getBaseIdx<checkpoint::tests::unit::HolderObjBase<int> > (base_idx=0)
    at ../src/checkpoint/dispatch/vrt/serializer_registry.h:156
#6  in checkpoint::tests::unit::HolderObjBase<int>::_checkpointDynamicSerialize (this=0x5555556634a0, s=0x7fffffffd7b0, base_ser_idx=0, expected_idx=1)
    at ../tests/unit/test_footprinter.cc:543
#7  in checkpoint::tests::unit::HolderBasic<int>::_checkpointDynamicSerialize (this=0x5555556634a0, s=0x7fffffffd7b0, base_ser_idx=0, expected_idx=-1)
    at ../tests/unit/test_footprinter.cc:553
#8  in checkpoint::dispatch::vrt::virtualSerialize<checkpoint::tests::unit::HolderBase, checkpoint::Footprinter> (base=@0x7fffffffd4c8: 0x5555556634a0, s=...)
    at ../src/checkpoint/dispatch/vrt/virtual_serialize.h:66
(...)

[cz4rs]

@lifflander @PhilMiller can you please have a look at this?
I have pushed a commit that reproduces this in #144 (the test case can probably be further simplified and moved into test_virtual_serialize.cc), but I have not been able to fix this so far

[PhilMiller]

I'm seeing the exact same issue in some code I'm working on

[PhilMiller]

I just checked that this is not the result of the changes in #133 - checking out its predecessor and running the test on that (modified to use the older-style macros with more arguments) still fails

[PhilMiller]

It actually still fails for me even when commenting out the templating on the classes in question!

[PhilMiller]

It seems that the intermediate class in the inheritance hierarchy plays some role in the problem - when I skip that, the test passes.

[cz4rs]

It seems that the intermediate class in the inheritance hierarchy plays some role in the problem - when I skip that, the test passes.

also, when the get() method in the middle class in not abstract, it passes ok

-  virtual ObjT* get() = 0;
+  virtual ObjT* get() { return nullptr; };

[cz4rs]

ok, so clearly it's not a template issue, seems like an abstract class in the middle is not ok
I will try to play around with different hierarchies and produce a truly minimal example

[PhilMiller]

I think that potentially is the hint that we need to unravel this - there's maybe an instantiation failure around the constructor functor, which the compiler doesn't complain about because it finds something less specialized to instantiate instead (SFINAE)

[PhilMiller]

Perhaps dispatch::Reconstructor<ObjT>::constructAllowFail(buf) in object_registry.h isn't behaving as intended

[PhilMiller]

Nope, looks like that's designed to always instantiate, and fail an assertion if it's ever called

[lifflander]

I have a fix in place now on #144

[cz4rs]

renaming after the fix, since templates were not the source of the issue here