[PATCH 7/7] [og10] WIP GOMP_MAP_ATTACH_ZERO_LENGTH_ARRAY_SECTION changes

[Julian Brown]

On Mon, 17 May 2021 21:14:26 +0800
Chung-Lin Tang <cltang@codesourcery.com> wrote:

> On 2021/5/11 4:57 PM, Julian Brown wrote:
> > This work-in-progress patch tries to get
> > GOMP_MAP_ATTACH_ZERO_LENGTH_ARRAY_SECTION to behave more like
> > GOMP_MAP_ATTACH_DETACH -- in that the mapping is made to form groups
> > to be processed by build_struct_group/build_struct_comp_map.  I
> > think that's important to integrate with how groups of mappings for
> > array sections are handled in other cases.
> > 
> > This patch isn't sufficient by itself to fix a couple of broken
> > test cases at present (libgomp.c++/target-lambda-1.C,
> > libgomp.c++/target-this-4.C), though.  
> 
> No, GOMP_MAP_ATTACH_ZERO_LENGTH_ARRAY_SECTION is supposed to be just
> a slightly different behavior version of GOMP_MAP_ATTACH; it
> tolerates an unmapped pointer-target and assigns NULL on the device,
> instead of just gomp_fatal(). (see its handling in libgomp/target.c)
> 
> In case OpenACC can have the same such zero-length array section
> behavior, we can just share one GOMP_MAP_ATTACH map. For now it is
> treated as separate cases.

OK, understood. But, I'm a bit concerned that we're ignoring some
"hidden rules" with regards to OMP pointer clause ordering/grouping that
certain code (at least the bit that creates GOMP_MAP_STRUCT node
groups, and parts of omp-low.c) relies on. I believe those rules are as
follows:

 - an array slice is mapped using two or three pointers -- two for a
   normal (non-reference) base pointer, and three if we have a
   reference to a pointer (i.e. in C++) or an array descriptor (i.e. in
   Fortran). So we can have e.g.

   GOMP_MAP_TO
   GOMP_MAP_ALWAYS_POINTER

   GOMP_MAP_TO
   GOMP_MAP_.*_POINTER
   GOMP_MAP_ALWAYS_POINTER

   GOMP_MAP_TO
   GOMP_MAP_TO_PSET
   GOMP_MAP_ALWAYS_POINTER

 - for OpenACC, we extend this to allow (up to and including
   gimplify.c) the GOMP_MAP_ATTACH_DETACH mapping. So we can have (for
   component refs):

   GOMP_MAP_TO
   GOMP_MAP_ATTACH_DETACH

   GOMP_MAP_TO
   GOMP_MAP_TO_PSET
   GOMP_MAP_ATTACH_DETACH

   GOMP_MAP_TO
   GOMP_MAP_.*_POINTER
   GOMP_MAP_ATTACH_DETACH

For the scanning in insert_struct_comp_map (as it is at present) to
work right, these groups must stay intact.  I think the current
behaviour of omp_target_reorder_clauses on the og10 branch can break
those groups apart though!

(The "prev_list_p" stuff in the loop in question in gimplify.c just
keeps track of the first node in these groups.)

For OpenACC, the GOMP_MAP_ATTACH_DETACH code does *not* depend on the
previous clause when lowering in omp-low.c. But GOMP_MAP_ALWAYS_POINTER
does! And in one case ("update" directive), GOMP_MAP_ATTACH_DETACH is
rewritten to GOMP_MAP_ALWAYS_POINTER, so for that case at least, the
dependency on the preceding mapping node must stay intact.

OpenACC also allows "bare" GOMP_MAP_ATTACH and GOMP_MAP_DETACH nodes
(corresponding to the "attach" and "detach" clauses). Those are handled
a bit differently to GOMP_MAP_ATTACH_DETACH in gimplify.c -- but
GOMP_MAP_ATTACH_Z_L_A_S doesn't quite behave like that either, I don't
think?

Anyway: I've not entirely understood what omp_target_reorder_clauses is
doing, but I think it may need to try harder to keep the groups
mentioned above together.  What do you think?

Thanks,

Julian