On Thu, Oct 15, 2020 at 12:50:18PM +0200, Jakub Jelinek via Gcc-patches wrote:
> If nothing of this is valid (i.e. mandatory forbids all the fallback), then
> your patch implements it, but if e.g. only if (false) is allowed and nothing
> else, then we are ATM in trouble (we pass -2 for if false as well as it is
> returned from omp_get_initial_device (); thankfully, OpenMP 5.0 changed it
> (something that should be implemented in libgomp too), so that
> omp_get_initial_device () is now required to be omp_get_num_devices (),
> i.e. the host device is always the last one; so, we would do fallback
> for -2 and not for anything else), if also omp_get_initial_device ()
> number is valid, then we should allow fallback for devices -2 and
> omp_get_initial_device () and punt on everything else.
Ok, the first response indicates that both if (false) and
omp_get_initial_device () are valid.
Therefore, I think until omp_get_initial_device () value is changed, we
want in resolve_device:
if (device_id < 0 || device_id >= gomp_get_num_devices ())
{
if (device_id != GOMP_DEVICE_HOST_FALLBACK)
gomp_fatal (...);
return NULL;
}
and do gomp_fatal also for further return NULLs in the function.
And then in
if (devicep == NULL
|| !(devicep->capabilities & GOMP_OFFLOAD_CAP_OPENMP_400)
/* All shared memory devices should use the GOMP_target_ext function. */
|| devicep->capabilities & GOMP_OFFLOAD_CAP_SHARED_MEM
|| !(fn_addr = gomp_get_target_fn_addr (devicep, fn)))
return gomp_target_fallback (fn, hostaddrs);
and similar do gomp_fatal before the gomp_target_fallback call
if target-offload-var ICV is mandatory and devicep != NULL.
And when we change omp_get_initial_device (), we'd use
device_id > gomp_get_num_devices () above instead, and
call gomp_fatal there only for
(device_id != GOMP_DEVICE_HOST_FALLBACK
&& device_id != gomp_get_num_devices ())
Jakub
