https://gcc.gnu.org/bugzilla/show_bug.cgi?id=127830
Bug ID: 127830
Summary: [OpenMP] need_device_ptr doesn't handle dependent
types
Product: gcc
Version: 17.0
Status: UNCONFIRMED
Keywords: openmp, rejects-valid
Severity: normal
Priority: P3
Component: c++
Assignee: unassigned at gcc dot gnu.org
Reporter: gabravier at gmail dot com
Target Milestone: ---
#include <type_traits>
template <class T> struct P { using pointer = T *; };
template <class T> void v2 (typename P<T>::pointer x);
#pragma omp declare variant (v2<T>) match (construct={dispatch}) adjust_args
(need_device_ptr: x)
template <class T> void b2 (typename P<T>::pointer x);
template <class T> void v3 (std::add_pointer_t<T> x);
#pragma omp declare variant (v3<T>) match (construct={dispatch}) adjust_args
(need_device_ptr: x)
template <class T> void b3 (std::add_pointer_t<T> x);
My reading of the OpenMP is that this code is valid (it doesn't appear to have
any issues with dependent types as a general matter, and e.g. ยง15.5.2.1 "in the
declaration of the function variant" would seem to imply that it has to be
checked post-resolution of dependent types since the declaration doesn't exist
until template instantiation), yet GCC (with -fopenmp) complains:
<source>:6:96: error: 'x' is not a C pointer
6 | #pragma omp declare variant (v2<T>) match (construct={dispatch})
adjust_args (need_device_ptr: x)
|
^
<source>:10:96: error: 'x' is not a C pointer
10 | #pragma omp declare variant (v3<T>) match (construct={dispatch})
adjust_args (need_device_ptr: x)
|
^
...this reproduces on trunk, 16 and 15 (14 and below do not implement
dispatch).
This code seems to be at fault:
if (strcmp (p, "need_device_ptr") == 0)
{
bool is_ptr_or_template
= TEMPLATE_PARM_P (TREE_TYPE (arg))
|| POINTER_TYPE_P (TREE_TYPE (arg));
if (!is_ptr_or_template)
{
error_at (arg_loc, "%qD is not a C pointer",
decl);
continue;
}
}
...I assume it should allow dependent types, and then later do the check
properly when the type is actually resolved.
