Committed as r11-2395,
https://gcc.gnu.org/g:269322ece17202632bc354e9c510e4a5bd6ad84b
with review comments applied plus a follow-up commit for an indentation issue I
missed:
r11-2399, https://gcc.gnu.org/g:6de5600a8bd1ef0ad3d57670efdcc68bb3484276
Tobias
On 7/29/20 10:23 AM, Jakub Jelinek wrote
On Wed, Jul 29, 2020 at 10:10:35AM +0200, Tobias Burnus wrote:
> + case AB_OMP_REQ_REVERSE_OFFLOAD:
> +gfc_omp_requires_add_clause (OMP_REQ_REVERSE_OFFLOAD,
> + "reverse_offload",
> + &gfc_current_lo
And another update; I removed the global variable, which
was only used for checking at the end – and
reconstruct its value in a local variable just before calling
gfc_check_omp_requires.
(Following a suggestion by Jakub and removing a kind-of left
over from the first implementation.)
Tobias
On
I just realized that supporting 'acq_rel' is simple; while
'omp atomic' parsing needs to be updated quite a bit for the
OpenMP 5 changes, just adding ACQ_REL support for 'requires'
is trivial.
Hence, I updated the requires-9.c testcase for 'acq_rel',
adjusted trans-openmp.c and did some openmp.c
Attached is an updated version – whether that will be fully
in line with OpenMP 5.1 remains to be seen. But in any case,
it now handles properly:
"If a directive appears in the declarative part of a module then the
behavior is as if that directive appears after any references to that
module."
Not
As with C/C++, the offloading requires are not implemented
an rejected early with a 'sorry'.
The mem-setting default for atomic is handled, but only 'seq_cst'
as currently gfortran's only supports this one. (C/C++ support more.)
In C/C++, the 'requires' has to be at file scope. For Fortran,
"mu
