https://gcc.gnu.org/bugzilla/show_bug.cgi?id=127273
--- Comment #2 from Tobias Burnus <burnus at gcc dot gnu.org> ---
OK - I messed partially up: It fails also with 'amdgcn-amdhsa',
i.e. it is likely a generic issue.
* * *
The testcase is now:
testsuite/libgomp.c/usm_env_handling-1-self.c
(which enables ''#pragma omp requires self_maps' and includes
testsuite/libgomp.c/usm_env_handling-1.c)
That testcase is for now XFAILED on nvptx (and will be xfailed for amdgcn as
well)
* * *
Simplified testcase [note that OpenACC rejects 'link' at compile time]
int A[1] = {1};
#pragma omp declare target link(A)
//#pragma acc declare device_resident(A) // OK
//#pragma acc declare link(A) // error: ‘A’ with ‘link’ clause used in
‘routine’ function
//#pragma acc routine
void
checkA (int expected, int also_ok, int other)
{
if (expected != A[0] && other != A[0])
__builtin_printf ("%d\n", A[0]);
if (expected != A[0] && also_ok != A[0])
__builtin_abort ();
}
int
main ()
{
A[0] = 2;
//#pragma acc parallel copyin(A)
#pragma omp target map(always, to: A)
checkA (2, 2, 2);
}
* * *
The failing code is the statement:
_14 = A$linkptr;
with
(gdb) p debug_bb(bb)
<bb 5> [local count: 1073741824]:
_14 = A$linkptr;
_2 = (*_14)[0];
if (_2 != expected_5(D))
goto <bb 6>; [66.00%]
else
goto <bb 8>; [34.00%]
* * *
The code:
symtab_node::get (t)
is
497 gcc_checking_assert (TREE_CODE (decl) == FUNCTION_DECL
498 || (TREE_CODE (decl) == VAR_DECL
499 && (TREE_STATIC (decl) || DECL_EXTERNAL
500 || in_lto_p)));
501 /* Check that the mapping is sane - perhaps this check can go away,
502 but at the moment frontends tends to corrupt the mapping by
calling
503 memcpy/memset on the tree nodes. */
504 gcc_checking_assert (!decl->decl_with_vis.symtab_node
505 || decl->decl_with_vis.symtab_node->decl ==
decl);
506 return decl->decl_with_vis.symtab_node;
And here decl->decl_with_vis.symtab_node == NULL, which obviously will segfault
when getting dereferenced.
