Add non-template ref-to-value converting constructor to __basic_any by bdice · Pull Request #8320 · NVIDIA/cccl

bdice · 2026-04-07T20:20:09Z

Description

Add a non-template converting constructor __basic_any(__basic_any<_Interface&> const&) to the __basic_any value specialization, enabling implicit conversion from proxy types that wrap a __basic_any reference type.

The __basic_any value specialization provides a forwarding constructor __basic_any(_Tp&&) and template converting constructors __basic_any(__basic_any<_OtherInterface> const&). When a proxy type wraps a __basic_any<I&> (e.g. a resource_ref) and provides operator __basic_any<I&>&(), neither path works: the forwarding constructor deduces the proxy type which doesn't satisfy the interface, and the template converting constructors can't deduce _OtherInterface from the proxy since implicit user-defined conversions are not considered during template argument deduction.

This occurs concretely with Cython's __Pyx_FakeReference<T>, which wraps intermediate C++ expression results and provides operator T&(). Constructing any_resource<P...> from a Cython-wrapped resource_ref<P...> fails at compile time.

The fix adds a non-template constructor for the same-interface ref-to-value conversion. Because the parameter is non-template, the compiler considers implicit conversion sequences through the proxy's operator T&(). The existing template converting constructor gains an additional !same_as<_OtherInterface, _Interface&> constraint to avoid ambiguity.

This single fix in __basic_any covers all downstream types: any_resource, any_synchronous_resource, and any future __basic_any-based value types.

Testing: A new test was added based on the minimal reproducer in the issue. RMM Cython bindings build and pass tests using direct any_resource construction from resource_ref without workaround helpers.

Checklist

New or existing tests cover these changes.
The documentation is up to date with these changes.

The __basic_any value specialization inherits converting constructors from templates that accept __basic_any<_OtherInterface>. When a proxy type wraps a __basic_any reference type (e.g. __basic_any<I&>) and provides operator __basic_any<I&>&(), these template constructors cannot match: template argument deduction deduces the proxy type, not the underlying __basic_any, and implicit user-defined conversions are not considered during deduction. Add a non-template constructor __basic_any(__basic_any<_Interface&> const&) for the ref-to-value conversion case. Non-template parameters participate in implicit conversion sequences, allowing proxy types to convert through their user-defined conversion operator. This fixes any_resource construction from proxy-wrapped resource_ref and any_synchronous_resource from proxy-wrapped synchronous_resource_ref without requiring per-type overloads. Fixes NVIDIA#8316

…d refs Add regression tests for NVIDIA#8316 verifying that any_resource and any_synchronous_resource can be constructed from a value_proxy<T> that wraps resource_ref or synchronous_resource_ref via operator T&(), mimicking Cython's __Pyx_FakeReference proxy pattern.

ericniebler

very nice 👍

Co-authored-by: Eric Niebler <eniebler@boost.org>

Work around CCCL issue where any_resource cannot be constructed from Cython's __Pyx_FakeReference proxy type wrapping resource_ref. Declare any_resource and device_accessible in .pxd, add inline make_any_device_resource helper. Update all 12 constructor call sites in .pyx to use the helper. Ref: NVIDIA/cccl#8320

- Add enable_if<__copyable> to the non-template converting constructors from __basic_any<_Interface&>, since converting a ref to a value requires copying the referenced object. - Add the corresponding move constructor from __basic_any<_Interface&>&&. - Update the template move converting constructor's same_as guard to use reference comparison (same_as<T&, U&>) for consistency with the copy converting constructor. - In tests, use CHECK((expr)) to prevent Catch2 expression decomposition from triggering an nvcc auto NTTP SFINAE bug via ADL. - Guard proxy construction tests with _CCCL_CUDA_COMPILER(NVCC, <, 12, 9) since older nvcc treats auto NTTP deduction failures as hard errors. - Add rvalue proxy test sections.

…f-constructor

bdice · 2026-04-10T17:18:36Z

+#  if !_CCCL_CUDA_COMPILER(NVCC, <, 12, 9)
+  // nvcc before CTK 12.9 has a bug where auto NTTP deduction failures in
+  // __satisfies are hard errors instead of SFINAE during overload resolution.


Just raising a flag here. It seems like CUDA 12.0 didn't work with these changes. I wasn't able to come to a good workaround, but maybe someone else will have ideas. This is the MRE (sadly not very minimal): https://gist.github.com/bdice/0704486a9d32bceb782acd931c823123

I can't give a Godbolt reproducer link because it requires the changes from this PR...

seems reasonable to exclude the tests that trigger a since-fixed compiler bug. particularly one that is hard to work around, like this one.

EDIT: i wonder if it is the same issue that i worked around here: https://github.com/NVIDIA/cccl/blob/main/libcudacxx/include/cuda/__utility/__basic_any/basic_any_value.h#L254-L255

github-actions · 2026-04-10T18:25:59Z

🥳 CI Workflow Results

🟩 Finished in 2h 08m: Pass: 100%/108 | Total: 2d 10h | Max: 2h 08m | Hits: 95%/281975

See results here.

pciolkosz · 2026-04-10T21:31:48Z

+    __convert_from(__other);
+  }
+
+  template <bool _Copyable = __copyable, ::cuda::std::enable_if_t<_Copyable, int> = 0>


Should this be _CCCL_REQUIRES?

No, this was intentional. The __basic_any class uses raw enable_if_t instead of _CCCL_TEMPLATE/_CCCL_REQUIRES to maintain ABI compatibility across C++17 and C++20. The _CCCL_REQUIRES macro expands to a requires-clause in C++20 but to additional SFINAE template parameters in C++17, which produces different mangled names depending on the dialect. This caused runtime failures (vtable cast assertions) when any_resource was used across DSO boundaries compiled with different -std= flags. See #7397 for the original bug report and #7401 / #7405 for the fix.

we only needed to use enable_if for the functions whose ptrs actually end up in the pseudo-vtable (the ones marked _CCCL_PUBLIC_API). for these constructors, _CCCL_REQUIRES would not cause the ABI issue described in #7397, and it would be consistent with the other constructors in this class.

Change all resource/adaptor constructors from device_async_resource_ref to cuda::mr::any_resource<cuda::mr::device_accessible> taken by value, with std::move into member storage. This follows the sink-parameter idiom (like shared_ptr), enabling move semantics for rvalue arguments and making ownership transfer explicit. Rename upstream_mr/upstream_resource to upstream for consistency. Cython bindings use a make_any_device_resource inline helper to work around CCCL template deduction issue with Cython's FakeReference proxy. Ref: NVIDIA/cccl#8320

…rs (#2354) ## Description Change all resource/adaptor constructors from `device_async_resource_ref` to `cuda::mr::any_resource<cuda::mr::device_accessible>` taken **by value**, with `std::move` into member storage. This follows the sink-parameter idiom (like `std::shared_ptr`), enabling move semantics for rvalue arguments and making ownership transfer explicit. The central motivation is that `device_async_resource_ref` is a non-owning reference that cannot bind to temporaries (rvalues/xvalues). Code like `pool_memory_resource(cuda_memory_resource(), size)` is impossible with `device_async_resource_ref` because the temporary `cuda_memory_resource` is destroyed before the adaptor can use it. `any_resource<device_accessible>` by value solves this: callers pass a resource that gets moved into type-erased owned storage, so the adaptor owns its upstream and no external lifetime management is needed. Also renames `upstream_mr` / `upstream_resource` to `upstream` for consistency across all adaptors. Cython bindings use a `make_any_device_resource` inline helper to work around a CCCL template deduction issue where `any_resource` cannot be constructed from Cython's `__Pyx_FakeReference` proxy type wrapping `resource_ref` ([NVIDIA/cccl#8320](NVIDIA/cccl#8320)). This workaround should be removed once CCCL merges the upstream fix. **Testing:** - C++ build: 178/178 targets - C++ tests: 103/103 passed - Python tests: 1785 passed, 1 skipped - All against stock (unpatched) CCCL ## Checklist - [x] I am familiar with the [Contributing Guidelines](https://github.com/rapidsai/rmm/blob/HEAD/CONTRIBUTING.md). - [x] New or existing tests cover these changes. - [x] The documentation is up to date with these changes.

…VIDIA#8320) * Add non-template ref-to-value converting constructor to __basic_any The __basic_any value specialization inherits converting constructors from templates that accept __basic_any<_OtherInterface>. When a proxy type wraps a __basic_any reference type (e.g. __basic_any<I&>) and provides operator __basic_any<I&>&(), these template constructors cannot match: template argument deduction deduces the proxy type, not the underlying __basic_any, and implicit user-defined conversions are not considered during deduction. Add a non-template constructor __basic_any(__basic_any<_Interface&> const&) for the ref-to-value conversion case. Non-template parameters participate in implicit conversion sequences, allowing proxy types to convert through their user-defined conversion operator. This fixes any_resource construction from proxy-wrapped resource_ref and any_synchronous_resource from proxy-wrapped synchronous_resource_ref without requiring per-type overloads. Fixes NVIDIA#8316 * Add tests for any_resource/any_synchronous_resource from proxy-wrapped refs Add regression tests for NVIDIA#8316 verifying that any_resource and any_synchronous_resource can be constructed from a value_proxy<T> that wraps resource_ref or synchronous_resource_ref via operator T&(), mimicking Cython's __Pyx_FakeReference proxy pattern. * Update libcudacxx/include/cuda/__utility/__basic_any/basic_any_value.h Co-authored-by: Eric Niebler <eniebler@boost.org> * Constrain ref-to-value constructors, add move overload, fix tests - Add enable_if<__copyable> to the non-template converting constructors from __basic_any<_Interface&>, since converting a ref to a value requires copying the referenced object. - Add the corresponding move constructor from __basic_any<_Interface&>&&. - Update the template move converting constructor's same_as guard to use reference comparison (same_as<T&, U&>) for consistency with the copy converting constructor. - In tests, use CHECK((expr)) to prevent Catch2 expression decomposition from triggering an nvcc auto NTTP SFINAE bug via ADL. - Guard proxy construction tests with _CCCL_CUDA_COMPILER(NVCC, <, 12, 9) since older nvcc treats auto NTTP deduction failures as hard errors. - Add rvalue proxy test sections. --------- Co-authored-by: Eric Niebler <eniebler@boost.org>

This PR updates CCCL to the latest 3.4.0 pre-release commit. Comparison of CCCL commits: NVIDIA/cccl@c5594eb...c936ad8 We specifically need these features/fixes for RAPIDS projects: - NVIDIA/cccl#8353 - NVIDIA/cccl#8272 - NVIDIA/cccl#8320 - NVIDIA/cccl#8486 Authors: - Niranda Perera (https://github.com/nirandaperera) - Bradley Dice (https://github.com/bdice) Approvers: - Bradley Dice (https://github.com/bdice) URL: #1008

github-project-automation Bot added this to CCCL Apr 7, 2026

github-project-automation Bot moved this to Todo in CCCL Apr 7, 2026

bdice requested a review from a team as a code owner April 7, 2026 20:20

bdice requested a review from griwes April 7, 2026 20:20

cccl-authenticator-app Bot moved this from Todo to In Review in CCCL Apr 7, 2026

ericniebler approved these changes Apr 7, 2026

View reviewed changes

Comment thread libcudacxx/include/cuda/__utility/__basic_any/basic_any_value.h Outdated

Update libcudacxx/include/cuda/__utility/__basic_any/basic_any_value.h

1074754

Co-authored-by: Eric Niebler <eniebler@boost.org>

This comment has been minimized.

Sign in to view

miscco requested changes Apr 8, 2026

View reviewed changes

Comment thread libcudacxx/include/cuda/__utility/__basic_any/basic_any_value.h Outdated

github-project-automation Bot moved this from In Review to In Progress in CCCL Apr 8, 2026

bdice mentioned this pull request Apr 9, 2026

Use any_resource<device_accessible> for upstream constructor parameters rapidsai/rmm#2354

Merged

3 tasks

bdice added 2 commits April 10, 2026 16:14

Merge remote-tracking branch 'upstream/main' into fix/any-resource-re…

de03818

…f-constructor

bdice commented Apr 10, 2026

View reviewed changes

pciolkosz reviewed Apr 10, 2026

View reviewed changes

pciolkosz approved these changes Apr 14, 2026

View reviewed changes

miscco approved these changes Apr 14, 2026

View reviewed changes

github-project-automation Bot moved this from In Progress to In Review in CCCL Apr 14, 2026

bdice merged commit b0f99b3 into NVIDIA:main Apr 14, 2026
130 of 134 checks passed

github-project-automation Bot moved this from In Review to Done in CCCL Apr 14, 2026

bdice mentioned this pull request Apr 17, 2026

Update CCCL version rapidsai/rapids-cmake#1008

Merged

7 tasks

bdice mentioned this pull request Jun 1, 2026

Remove Cython make_any_device_resource helper rapidsai/rmm#2434

Closed

Navigation Menu

Uh oh!

Add non-template ref-to-value converting constructor to __basic_any - #8320

Add non-template ref-to-value converting constructor to __basic_any#8320
bdice merged 5 commits into
NVIDIA:mainfrom
bdice:fix/any-resource-ref-constructor

bdice commented Apr 7, 2026 •

edited

Loading

Uh oh!

ericniebler left a comment

Uh oh!

Uh oh!

This comment has been minimized.

Uh oh!

bdice Apr 10, 2026

Uh oh!

ericniebler Apr 14, 2026 •

edited

Loading

Uh oh!

Uh oh!

github-actions Bot commented Apr 10, 2026

Uh oh!

pciolkosz Apr 10, 2026

Uh oh!

bdice Apr 11, 2026 •

edited

Loading

Uh oh!

ericniebler Apr 14, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

Uh oh!

Conversation

bdice commented Apr 7, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Description

Checklist

Uh oh!

ericniebler left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

This comment has been minimized.

Uh oh!

bdice Apr 10, 2026

Choose a reason for hiding this comment

Uh oh!

ericniebler Apr 14, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

github-actions Bot commented Apr 10, 2026

🥳 CI Workflow Results

🟩 Finished in 2h 08m: Pass: 100%/108 | Total: 2d 10h | Max: 2h 08m | Hits: 95%/281975

Uh oh!

pciolkosz Apr 10, 2026

Choose a reason for hiding this comment

Uh oh!

bdice Apr 11, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

ericniebler Apr 14, 2026

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

bdice commented Apr 7, 2026 •

edited

Loading

ericniebler Apr 14, 2026 •

edited

Loading

bdice Apr 11, 2026 •

edited

Loading