First *ping* (two weeks old but all during the holiday season).
Side note, for what it is worth: Thomas K regarded the Fortran part as
reasonable/OK – but those OpenACC changes also affect
C/C++/omp-low.c/libgomp :-) – cf.
https://gcc.gnu.org/ml/gcc-patches/2019-12/msg01596.html
Cheers,
Tobias
On 12/24/19 3:23 PM, Tobias Burnus wrote:
On the front-end side, adding "if" and "if_present" to the "acc
host_data" directive is simple as other directives already support
those clauses.
The 'if_present' status has to be passed along the use_device_ptr
flag; for this a new flag has been introduced, using the gap in the
gomp_map_kind enum (16 was still free; now used for
GOMP_MAP_USE_DEVICE_PTR_IF_PRESENT, use_device_ptr has 14).
There was some forth and back whether a new enum should be used or a
flag (passed in the device argument, as OpenACC only uses host and
device and not a device number). Initial version was with new enum,
then with flag and now again a new enum.
On the libgomp side, one simply needs to skip the if-not-present error
and otherwise handle it as use_device_ptr.
Build on x86-64-gnu-linux without offloading and with nvptx offloading.
OK for the trunk?
Cheers,
Tobias
PS: History: The initial version (not public) used an enum but Thomas
was wondering whether a flag is not the better solution, cf.
discussion at https://gcc.gnu.org/ml/gcc/2018-12/msg00118.html –
Hence, the first public version has switched to a flag and submitted
for the OG8 (openacc-gcc-8-branch GIT) branch at
https://gcc.gnu.org/ml/gcc-patches/2018-12/msg01559.html and has been
committed as Rev. b52c8d006581b2a657ac3d81ab2527bb398b6615 (this
commit is also part of OG9. — A minor fix was committed to OG8?/OG9 as
Rev. 995f9680a46c3a7246fe465faa847f8009e47ed8.
Thomas now wonders whether the flag will be future safe, hence, it now
uses again an enum as in the original version.
PPS: Changes to OG9: Using again an enum; Fortran run-time test and
rediffs.
