https://gcc.gnu.org/bugzilla/show_bug.cgi?id=122517
--- Comment #1 from Tobias Burnus <burnus at gcc dot gnu.org> ---
The fail is for the assert:
gcc_assert ((n->value & GOVD_SEEN) == 0 | ...
as n == NULL.
If one at this location invokes 'return', the program finishes with the (valid)
message:
error: OpenMP constructs other than ‘ordered simd’, ‘simd’, ‘loop’ or ‘atomic’
may not be nested inside ‘simd’ region
16 | #pragma omp interop init(target : intop)
Still, the question is: why does the code assume that the symbol can be found
(n != NULL) why does this now fail?
* * *
The following variant fails with:
#0 scan_sharing_clauses (clauses=0x7ffff720d828, ctx=0x357f590) at
../../../repos/gcc/gcc/omp-low.cc:1217
1217 gcc_assert (is_taskreg_ctx (ctx));
#1 0x0000000001040db8 in scan_omp_target (stmt=0x7ffff73e6cb8, outer_ctx=0x0)
at ../../../repos/gcc/gcc/omp-low.cc:3151
#2 0x0000000001041f89 in scan_omp_1_stmt (gsi=0x7fffffffc860,
handled_ops_p=0x7fffffffc82f, wi=0x7fffffffc9d0) at
../../../repos/gcc/gcc/omp-low.cc:4272
as is_taskreg_ctx (ctx) == false with:
is_taskreg_ctx (omp_context *ctx)
{
return is_parallel_ctx (ctx) || is_task_ctx (ctx) || is_host_teams_ctx
(ctx);
with scan_omp_target's (gdb) p debug(stmt)
#pragma omp target num_teams(-2) thread_limit(0) shared(intop) shared(cnt)
shared(n) map(tofrom:cnt [len: 4]) [child fn: f._omp_fn.0 (???)]
{
#pragma omp parallel firstprivate(x) firstprivate(i) shared(cnt) shared(n)
WHY ARE THERE SHARED clauses on TARGET?
// ...
#pragma omp target map(tofrom: cnt) //, intop)
#pragma omp parallel
for (int i = 1; i < n; i++)
{
int x = 1;
if (omp_is_initial_device())
{
intop = omp_interop_nonequit
#pragma omp interop init(target : intop)
}
// ...
Presumably, INTEROP causes that the context gets messed up?
