On Thu, Dec 21, 2023 at 01:31:19PM +0100, Thomas Schwinge wrote:
> These three: implicitly, or explicit '#pragma omp declare target' etc.,
> or inside '#pragma omp begin declare target' region are the only OpenMP
> facilities to get things 'omp declare target'ed, right?
I think so.
> That doesn't generally work, as the gimplification-level code re
> 'Static locals [...] need to be "omp declare target"' runs *after*
> 'omp_discover_implicit_declare_target'. Thus my "move" idea above.
Can't we mark the static locals already during that discovery?
The addition during gimplification was probably made when we didn't have
that at all.
> OK to push, for a start, the attached
> "GCN, nvptx: Basic '__cxa_guard_{acquire,abort,release}' for C++ static local
> variables support"?
> That's now in libgcc not libgomp, so that it's also usable for GCN, nvptx
> target testing, where we thus see a number of FAIL -> PASS progressions.
> For now, for single-threaded GCN, nvptx target use only; extension for
> multi-threaded offloading use to follow later.
>
> libgcc/
> * c++-minimal/README: New.
> * c++-minimal/guard.c: New.
> * config/gcn/t-amdgcn (LIB2ADD): Add it.
> * config/nvptx/t-nvptx (LIB2ADD): Likewise.
> +/* Copy'n'paste/edit from 'libstdc++-v3/libsupc++/cxxabi.h'. */
> +
> + int
> + __cxa_guard_acquire(__guard*);
> +
> + void
> + __cxa_guard_release(__guard*);
> +
> + void
> + __cxa_guard_abort(__guard*);
When all this isn't inside a namespace, shouldn't it be indented by
2 spaces less?
> +
> +/* Copy'n'paste/edit from 'libstdc++-v3/libsupc++/guard.cc'. */
> +
> +# undef _GLIBCXX_GUARD_TEST_AND_ACQUIRE
> +# undef _GLIBCXX_GUARD_SET_AND_RELEASE
> +# define _GLIBCXX_GUARD_SET_AND_RELEASE(G) _GLIBCXX_GUARD_SET (G)
And without a space after # here?
Otherwise LGTM, but hope that one day we'll get rid of it again.
Jakub
