[Bug middle-end/95270] New: OpenACC 'enter data attach(data_p)' fails for 'int *data_p'

[Bug middle-end/95270] New: OpenACC 'enter data attach(data_p)' fails for 'int *data_p'

[tschwinge at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=95270

            Bug ID: 95270
           Summary: OpenACC 'enter data attach(data_p)' fails for 'int
                    *data_p'
           Product: gcc
           Version: 10.0
            Status: UNCONFIRMED
          Keywords: openacc
          Severity: normal
          Priority: P3
         Component: middle-end
          Assignee: unassigned at gcc dot gnu.org
          Reporter: tschwinge at gcc dot gnu.org
                CC: jakub at gcc dot gnu.org, jules at gcc dot gnu.org
  Target Milestone: ---

Created attachment 48581
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=48581&action=edit
'mdc-2b_.c'

See test case attached: 'acc_attach(&data_p)' works, '#pragma enter data
attach(data_p)' doesn't.

In such a case, I suppose we either need to explicitly set 'OMP_CLAUSE_SIZE' in
all front ends, or otherwise make sure to skip/invalidate the
'gcc/gimplify.c:gimplify_scan_omp_clauses' handling:

    if (OMP_CLAUSE_SIZE (c) == NULL_TREE)
      OMP_CLAUSE_SIZE (c) = DECL_P (decl) ? DECL_SIZE_UNIT (decl)
                            : TYPE_SIZE_UNIT (TREE_TYPE (decl));

..., or elsewhere?

Given that for 'GOMP_MAP_ATTACH' etc. this is a *bias*, not a *size*, I'd like
to see this explicitly mentioned/handled in the code, under the assumption that
for 'GOMP_MAP_STRUCT'-related 'GOMP_MAP_ATTACH', we do need this set non-zero. 
Also need to update the comments in 'include/gomp-constants.h:enum
gomp_map_kind' to explain that (see 'GOMP_MAP_POINTER' example?).

[Bug middle-end/95270] OpenACC 'enter data attach' looks up target memory object displaced by pointer size

[tschwinge at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=95270

Thomas Schwinge <tschwinge at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Assignee|unassigned at gcc dot gnu.org      |jules at gcc dot gnu.org
             Status|UNCONFIRMED                 |ASSIGNED
   Target Milestone|---                         |10.0
            Summary|OpenACC 'enter data         |OpenACC 'enter data attach'
                   |attach(data_p)' fails for   |looks up target memory
                   |'int *data_p'               |object displaced by pointer
                   |                            |size
   Last reconfirmed|                            |2020-06-08
     Ever confirmed|0                           |1

--- Comment #1 from Thomas Schwinge <tschwinge at gcc dot gnu.org> ---
I just hit this one again, started to debug it once again (until I finally
remembered this existing issue...), and cried a little bit.  ;'-|

Julian, this is a bug, right?  During gimplification, or in the front end(s)? 
Please have a look.

This issue is also visible in the (current) 'omplower' tree dump scans in
'c-c++-common/goacc/mdc-1.c': expecting 'bias: 8'.  (Pointer size, by the way;
cross-checked with '-m32' where we get 'bias: 4'.)

I'd thought I remember that I may work around this by wrapping pointers as
'struct' members, which would then 'attach' correctly, without size/bias
confusion, but apparently that's not true; the attached 'mdc-2b_.c' fails in
the same way even if altered as follows:

    -  int *data_p = &data;
    +#if 0
    +  int *data_p;
    +#else
    +  struct s {
    +    int *data_p;
    +  } s;
    +# define data_p s.data_p
    +#endif
    +  data_p = &data;

What you instead have to do is to make sure that your target memory object
(here: 'data') is bigger than the bias (pointer size):

    -  int data;
    +  struct { char d[sizeof (void *) + 1]; } data;

Then it works, in all variants I've now posted.

Execution test cases that would ("might") exercise/notice this behavior:

  - 'libgomp.oacc-c-c++-common/deep-copy-2.c' -- works despite the bug, because
the '#pragma acc data copy(s.a[0:10])' already synthesizes an implicit
'map(attach:s.a [bias: 0])' (correct) and thus the incorrect 'map(attach:s.a
[bias: 8])' synthesized for the inner '#pragma acc parallel loop attach(s.a)'
doesn't cause a problem, as 'bias' is only used in the first call of the
respective 'gomp_attach_pointer'.
  - 'libgomp.oacc-c-c++-common/deep-copy-4.c' -- works despite the bug, because
for a struct bigger than the bias ('struct node' here), a 'splay_tree_lookup'
with start offset by bias will then still look up the expected object,
apparently.  With the 'val'/'sum' business removed, this test case quickly
fails: 'libgomp: pointer target not mapped for attach'.
  - 'libgomp.oacc-fortran/deep-copy-6.f90' -- correctly synthesizes
'map(attach:var.a [bias: 0])' for '!$acc enter data attach(var%a)'.  However,
for the earlier '!$acc enter data copyin(var%a(5:n - 5), var%b(5:n - 5))'
synthesizes the two 'gomp_attach_pointer' calls both with 'bias = 16' -- seems
to work, but is that correct?
  - 'libgomp.oacc-fortran/deep-copy-8.f90' -- correctly synthesizes
'map(attach:var.a [bias: 0])' for '!$acc enter data attach(var%a)'.

That's all with an explicit 'attach' clause, folks.

[Bug middle-end/95270] OpenACC 'enter data attach' looks up target memory object displaced by pointer size

[jules at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=95270

--- Comment #2 from jules at gcc dot gnu.org ---
TBH I've suspected problems with misuse of the bias for attach/detach before,
but I've not come up with a test case. I'll have a look.

Thanks!

[Bug middle-end/95270] OpenACC 'enter data attach' looks up target memory object displaced by pointer size

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=95270

--- Comment #3 from CVS Commits <cvs-commit at gcc dot gnu.org> ---
The master branch has been updated by Julian Brown <jules@gcc.gnu.org>:

https://gcc.gnu.org/g:0d00fe404c162ad0cf922ca8455aa23a74042b63

commit r11-1976-g0d00fe404c162ad0cf922ca8455aa23a74042b63
Author: Julian Brown <julian@codesourcery.com>
Date:   Tue Jun 9 06:21:34 2020 -0700

    openacc: Set bias to zero for explicit attach/detach clauses in C and C++

    This is a fix for the pointer (or array) size inadvertently being used
    for the bias with attach and detach mapping kinds, for both C and C++.

    2020-07-09  Julian Brown  <julian@codesourcery.com>
                Thomas Schwinge  <thomas@codesourcery.com>

    gcc/c/
            PR middle-end/95270
            * c-typeck.c (c_finish_omp_clauses): Set OMP_CLAUSE_SIZE (bias) to
zero
            for standalone attach/detach clauses.

    gcc/cp/
            PR middle-end/95270
            * semantics.c (finish_omp_clauses): Likewise.

    include/
            PR middle-end/95270
            * gomp-constants.h (gomp_map_kind): Expand comment for
attach/detach
            mapping kinds.

    gcc/testsuite/
            PR middle-end/95270
            * c-c++-common/goacc/mdc-1.c: Update expected dump output for zero
            bias.

    libgomp/
            PR middle-end/95270
            * testsuite/libgomp.oacc-c-c++-common/pr95270-1.c: New test.
            * testsuite/libgomp.oacc-c-c++-common/pr95270-2.c: New test.

[Bug middle-end/95270] OpenACC 'enter data attach' looks up target memory object displaced by pointer size

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=95270

--- Comment #4 from CVS Commits <cvs-commit at gcc dot gnu.org> ---
The releases/gcc-10 branch has been updated by Julian Brown
<jules@gcc.gnu.org>:

https://gcc.gnu.org/g:2974e1eee053496e1d43f7f2e62f5feac2aa0315

commit r10-8475-g2974e1eee053496e1d43f7f2e62f5feac2aa0315
Author: Julian Brown <julian@codesourcery.com>
Date:   Tue Jun 9 06:21:34 2020 -0700

    openacc: Set bias to zero for explicit attach/detach clauses in C and C++

    This is a fix for the pointer (or array) size inadvertently being used
    for the bias with attach and detach mapping kinds, for both C and C++.

    2020-07-09  Julian Brown  <julian@codesourcery.com>
                Thomas Schwinge  <thomas@codesourcery.com>

    gcc/c/
            PR middle-end/95270
            * c-typeck.c (c_finish_omp_clauses): Set OMP_CLAUSE_SIZE (bias) to
zero
            for standalone attach/detach clauses.

    gcc/cp/
            PR middle-end/95270
            * semantics.c (finish_omp_clauses): Likewise.

    include/
            PR middle-end/95270
            * gomp-constants.h (gomp_map_kind): Expand comment for
attach/detach
            mapping kinds.

    gcc/testsuite/
            PR middle-end/95270
            * c-c++-common/goacc/mdc-1.c: Update expected dump output for zero
            bias.

    libgomp/
            PR middle-end/95270
            * testsuite/libgomp.oacc-c-c++-common/pr95270-1.c: New test.
            * testsuite/libgomp.oacc-c-c++-common/pr95270-2.c: New test.

    (cherry picked from commit 0d00fe404c162ad0cf922ca8455aa23a74042b63)

[Bug middle-end/95270] OpenACC 'enter data attach' looks up target memory object displaced by pointer size

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=95270

Andrew Pinski <pinskia at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
         Resolution|---                         |FIXED
   Target Milestone|10.0                        |10.2
             Status|ASSIGNED                    |RESOLVED

--- Comment #5 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
Fixed.