NVIDIA
[BUG]: constraint recursion in the `cuda::mr::resource` concept
Is this a duplicate?
- [X] I confirmed there appear to be no duplicate issues for this bug and that I agree to the Code of Conduct
Type of Bug
Compile-time Error
Component
libcu++
Describe the bug
The resource concept is defined as follows:
template <class _Resource>
_LIBCUDACXX_CONCEPT_FRAGMENT(
__resource_,
requires(_Resource& __res, void* __ptr, size_t __bytes, size_t __alignment)(
requires(_CUDA_VSTD::same_as<void*, decltype(__res.allocate(__bytes, __alignment))>),
requires(_CUDA_VSTD::same_as<void, decltype(__res.deallocate(__ptr, __bytes, __alignment))>),
requires(_CUDA_VSTD::equality_comparable<_Resource>)));
template <class _Resource>
_LIBCUDACXX_CONCEPT resource = _LIBCUDACXX_FRAGMENT(__resource_, _Resource);
A type must be equality_comparable to satisfy resource. But some resource types that are heterogeneously comparable have operator== overloads that are indirectly constrained with resource (like below), which causes constrain recursion and a hard error.
template <class _Resource>
_CCCL_NODISCARD_FRIEND auto operator==(cuda_memory_resource const& __lhs, _Resource const& __rhs) noexcept
_LIBCUDACXX_TRAILING_REQUIRES(bool)(__different_resource<cuda_memory_resource, _Resource>)
{
return resource_ref<>{const_cast<cuda_memory_resource&>(__lhs)} == resource_ref<>{const_cast<_Resource&>(__rhs)};
}
The problem here is the __different_resource concept, which subsumes resource.
My suggestion is to drop support for heterogeneous comparisons.
How to Reproduce
#include <cuda/memory_resource>
struct my_resource : cuda::mr::cuda_memory_resource {
using cuda::mr::cuda_memory_resource::cuda_memory_resource;
};
static_assert(cuda::mr::resource<my_resource>);
result:
/opt/compiler-explorer/libs/cccl/trunk/libcudacxx/include/cuda/__memory_resource/resource.h(104): error: expression must have a constant value
(!::cuda::std::__4::same_as< ::cuda::std::__4::decay_t<_Resource>, ::cuda::std::__4::decay_t<_OtherResource>>)
^
/opt/compiler-explorer/libs/cccl/trunk/libcudacxx/include/cuda/__memory_resource/resource.h(105): note #2689-D: the value of variable "cuda::mr::__4::resource [with _Resource=my_resource]" (declared at line 58) cannot be used as a constant
&& resource<_OtherResource>;
^
detected during:
instantiation of "const bool cuda::mr::__4::__different_resource [with _Resource=cuda::mr::__4::cuda_memory_resource, _OtherResource=my_resource]" at line 75 of /opt/compiler-explorer/libs/cccl/trunk/libcudacxx/include/cuda/std/__concepts/equality_comparable.h
instantiation of "const bool cuda::std::__4::__weakly_equality_comparable_with [with _Tp=my_resource, _Up=my_resource]" at line 78 of /opt/compiler-explorer/libs/cccl/trunk/libcudacxx/include/cuda/std/__concepts/equality_comparable.h
instantiation of "const bool cuda::std::__4::equality_comparable [with _Tp=my_resource]" at line 58
instantiation of "const bool cuda::mr::__4::resource [with _Resource=my_resource]" at line 7 of <source>
<source>(7): error: static assertion failed
static_assert(cuda::mr::resource<my_resource>);
^
2 errors detected in the compilation of "<source>".
Compiler returned: 2
https://godbolt.org/z/4o97PM7oz
Expected behavior
the static_assert should not cause a hard error.
Reproduction link
https://godbolt.org/z/4o97PM7oz
Operating System
No response
nvidia-smi output
No response
NVCC version
No response
I encountered this problem in one of my tests an I thought that i "fixed" it through the ``__different_from` concept, that should short circuit the concept evaluation if both resources have the same type. With that the type should be equality comparable or am I missing something.
