On Wed, Jul 08, 2020 at 12:21:57PM +0200, Thomas Schwinge wrote:
> Andrew and I are currently trying to sort out zero-length array sections
> behavior in the OpenACC context. From that fell out a testcase that I
> created to learn/verify (some of?) the OpenMP behavior. OK to push?
>
>
> Oh, and any thought on this:
>
> int dev = omp_get_default_device();
> bool shared_mem = false;
> #pragma omp target map(alloc:shared_mem)
> shared_mem = true;
> #if 1 //TODO
> /* 'omp_target_is_present' (and a few others) have behavior different from
> most others, where 'devicep == NULL' is handled the same as
> 'device_num ==
> GOMP_DEVICE_HOST_FALLBACK'. Is that difference intentional? */
> /* Given that GCC doesn't support shared-memory offload devices, we fake
> the
> expected thing as follows. */
> if (shared_mem)
> dev = omp_get_initial_device(); // GOMP_DEVICE_HOST_FALLBACK
> #endif
The difference is in quite unclear cases, where the device is either not
valid (outside of range) or the device used to be valid, but has been
unloaded already.
I don't remember why have I done it that way, we could as well treat also
the device == NULL cases as fallback, but any expectations how it will
behave in those cases are wrong IMHO.
> int *a = malloc(n * sizeof *a);
> assert(a);
I'd probably just return 0; from the test instead if a is NULL, that is not
something invalid.
Otherwise LGTM.
Jakub
