https://gcc.gnu.org/bugzilla/show_bug.cgi?id=111707
Bug ID: 111707
Summary: omp_target_alloc and omp_target_is_present - corner
case handling & spec conformance
Product: gcc
Version: 14.0
Status: UNCONFIRMED
Keywords: documentation, wrong-code
Severity: normal
Priority: P3
Component: libgomp
Assignee: unassigned at gcc dot gnu.org
Reporter: burnus at gcc dot gnu.org
CC: jakub at gcc dot gnu.org
Target Milestone: ---
To be checked / updated.
Besides fixing the code, also libgomp.texi needs to be updated.
omp_target_alloc:
For size == 0, the spec meanwhile requires that NULL is returned.
However, we call 'malloc' on the host which might return a
a non-NULL pointer that can be passed to 'free'.
And on the device side, an GOMP_error diagnostic is printed
to stderr for nvptx and gcn if the allocation fails.
Both the HSA and CUDA documentation imply that size==0 is
an error.
omp_target_is_present
OpenMP states that this functions checks whether the pointer is
mapped.
In GCC, we also return 'true' for shared memory systems and
for the device_num being the host.
The question is whether that's OK or not.
Does a SELF MAP qualify or not?
(self map - A mapping operation that, when corresponding storage is not
already present, results in corresponding storage that is the same as
its original storage.)
[The current behavior is about to be documented, but we can
change it.]
