Fortran: OpenMP/OpenACC avoid uninit access in size calc for mapping

  Hi Thomas & Jakub,

found when working on the deep-mapping patch* with OpenMP code
(and part of that patch) but it already shows up in an existing
OpenACC testcase. I think it makes sense to fix it already for GCC 12.

Problem: Also for unallocated allocatables, their size was
calculated - the 'if(desc.data == NULL)' check was only added
for pointers.

Result after the patch: When compiling with -O (which is the default
for goacc.exp), the warning now disappears. Thus, I now use '-O0'
and the previous "is uninitialized" is now "may be uninitialized".

Unrelated to the patch and the testcase, I added some
'allocate'**/'if(allocated())' to the testcase - as otherwise
uninit vars would be accessed. (Not relevant for the warning
or the patch - but I prefer no invalid code in testcases,
if it can be avoided.)

OK for mainline?

Tobias
* https://gcc.gnu.org/pipermail/gcc-patches/2022-March/591144.html

** I am actually not sure whether 'acc update(b)' will/should map a
previous allocated variable - or whether it should. But that's
unrelated to this bug fix. See also: https://gcc.gnu.org/PR96668
for the re-mapping in OpenMP (works for arrays but not scalars).
-----------------
Siemens Electronic Design Automation GmbH; Anschrift: Arnulfstraße 201, 80634 München; Gesellschaft mit beschränkter Haftung; Geschäftsführer: Thomas Heurung, Frank Thürauf; Sitz der Gesellschaft: München; Registergericht München, HRB 106955
  

Hi Tobias!

On 2022-03-08T15:25:07+0100, Tobias Burnus <tobias@codesourcery.com> wrote:
> found when working on the deep-mapping patch* with OpenMP code
> (and part of that patch) but it already shows up in an existing
> OpenACC testcase. I think it makes sense to fix it already for GCC 12.
>
> Problem: Also for unallocated allocatables, their size was
> calculated - the 'if(desc.data == NULL)' check was only added
> for pointers.
>
> Result after the patch: When compiling with -O (which is the default
> for goacc.exp), the warning now disappears. Thus, I now use '-O0'
> and the previous "is uninitialized" is now "may be uninitialized".

I recently added that checking in
commit 4bd8b1e881f0c26a5103cd1919809b3d63b60ef2
"Document current '-Wuninitialized'/'-Wmaybe-uninitialized' diagnostics
for OpenACC test cases", to document the status quo.

I'll leave it to you to decide what is more appropriate: (1), as you have
proposed, add '-O0' (but with source code comment, please); something
like:

     ! { dg-additional-options -Wuninitialized }
    +! Trigger "may be used uninitialized".
    +! { dg-additional-options -O0 }

..., or (2): update the test cases to simply reflect diagnostics that are
now (no longer) seen with (default) '-O' (rationale: the test cases
haven't originally been written for the '-Wuninitialized' diagnostics;
that's just tested additionally, and using '-O0' instead of '-O' may be
disturbing what they originally meant to test?), or (3): duplicate the
test cases to account for both (1) and (2) (in other words: write
dedicated test cases for your GCC/Fortran front end changes (for example,
based on the ones you've modified here), and for the existing test cases
apply (2)).  The latter, (3), would be my approach.

> Unrelated to the patch and the testcase, I added some
> 'allocate'**/'if(allocated())' to the testcase - as otherwise
> uninit vars would be accessed. (Not relevant for the warning
> or the patch - but I prefer no invalid code in testcases,
> if it can be avoided.)

Agreed in principle, but again: I don't know what these test cases
originally have been testing?

> OK for mainline?

I can't comment on the GCC/Fortran front end changes -- so unless
somebody else speaks up, that's an implicit approval for those, I
suppose.  ;-)

> Tobias
> * https://gcc.gnu.org/pipermail/gcc-patches/2022-March/591144.html

> ** I am actually not sure whether 'acc update(b)' will/should map a
> previous allocated variable - or whether it should.

(Are the typos here: in "will/should map": 's%map%update', and in
"or whether it should": 's%should%shouldn't'?)

> But that's
> unrelated to this bug fix. See also: https://gcc.gnu.org/PR96668
> for the re-mapping in OpenMP (works for arrays but not scalars).

I don't quickly dig that, sorry.  Do we need to first clarify that with
OpenACC Technical Committee, or is this just a GCC/OpenACC implementation
issue?


Grüße
 Thomas


> Fortran: OpenMP/OpenACC avoid uninit access in size calc for mapping
>
> gcc/fortran/ChangeLog:
>
>       * trans-openmp.cc (gfc_trans_omp_clauses, gfc_omp_finish_clause):
>       Obtain size for mapping only if allocatable array is allocated.
>
> gcc/testsuite/ChangeLog:
>
>       * gfortran.dg/goacc/array-with-dt-1.f90: Run with -O0 and
>       update dg-warning.
>       * gfortran.dg/goacc/pr93464.f90: Likewise.
>
>  gcc/fortran/trans-openmp.cc                         |  6 ++++--
>  gcc/testsuite/gfortran.dg/goacc/array-with-dt-1.f90 | 12 +++++++++---
>  gcc/testsuite/gfortran.dg/goacc/pr93464.f90         |  8 ++++----
>  3 files changed, 17 insertions(+), 9 deletions(-)
>
> diff --git a/gcc/fortran/trans-openmp.cc b/gcc/fortran/trans-openmp.cc
> index 4d56a771349..fad76a4791f 100644
> --- a/gcc/fortran/trans-openmp.cc
> +++ b/gcc/fortran/trans-openmp.cc
> @@ -1597,7 +1597,8 @@ gfc_omp_finish_clause (tree c, gimple_seq *pre_p, bool openacc)
>        tree size = create_tmp_var (gfc_array_index_type);
>        tree elemsz = TYPE_SIZE_UNIT (gfc_get_element_type (type));
>        elemsz = fold_convert (gfc_array_index_type, elemsz);
> -      if (GFC_TYPE_ARRAY_AKIND (type) == GFC_ARRAY_POINTER
> +      if (GFC_TYPE_ARRAY_AKIND (type) == GFC_ARRAY_ALLOCATABLE
> +       || GFC_TYPE_ARRAY_AKIND (type) == GFC_ARRAY_POINTER
>         || GFC_TYPE_ARRAY_AKIND (type) == GFC_ARRAY_POINTER_CONT)
>       {
>         stmtblock_t cond_block;
> @@ -3208,7 +3209,8 @@ gfc_trans_omp_clauses (stmtblock_t *block, gfc_omp_clauses *clauses,
>
>                     /* We have to check for n->sym->attr.dimension because
>                        of scalar coarrays.  */
> -                   if (n->sym->attr.pointer && n->sym->attr.dimension)
> +                   if ((n->sym->attr.pointer || n->sym->attr.allocatable)
> +                       && n->sym->attr.dimension)
>                       {
>                         stmtblock_t cond_block;
>                         tree size
> diff --git a/gcc/testsuite/gfortran.dg/goacc/array-with-dt-1.f90 b/gcc/testsuite/gfortran.dg/goacc/array-with-dt-1.f90
> index 136e42acd59..f6880238c89 100644
> --- a/gcc/testsuite/gfortran.dg/goacc/array-with-dt-1.f90
> +++ b/gcc/testsuite/gfortran.dg/goacc/array-with-dt-1.f90
> @@ -1,4 +1,4 @@
> -! { dg-additional-options -Wuninitialized }
> +! { dg-additional-options "-Wuninitialized -O0" }
>
>  type t
>     integer, allocatable :: A(:,:)
> @@ -8,9 +8,15 @@ type(t), allocatable :: b(:)
>  ! { dg-note {'b' declared here} {} { target *-*-* } .-1 }
>
>  !$acc update host(b)
> -! { dg-warning {'b\.dim\[0\]\.ubound' is used uninitialized} {} { target *-*-* } .-1 }
> -! { dg-warning {'b\.dim\[0\]\.lbound' is used uninitialized} {} { target *-*-* } .-2 }
> +! { dg-warning {'b\.dim\[0\]\.ubound' may be used uninitialized} {} { target *-*-* } .-1 }
> +! { dg-warning {'b\.dim\[0\]\.lbound' may be used uninitialized} {} { target *-*-* } .-2 }
> +
> +allocate(b(1))
> +!$acc update host(b)
>  !$acc update host(b(:))
> +
> +!$acc update host(b(1)%A)
> +allocate(b(1)%A(1,1))
>  !$acc update host(b(1)%A)
>  !$acc update host(b(1)%A(:,:))
>  end
> diff --git a/gcc/testsuite/gfortran.dg/goacc/pr93464.f90 b/gcc/testsuite/gfortran.dg/goacc/pr93464.f90
> index c92f1d3d8b2..18531abdf77 100644
> --- a/gcc/testsuite/gfortran.dg/goacc/pr93464.f90
> +++ b/gcc/testsuite/gfortran.dg/goacc/pr93464.f90
> @@ -2,17 +2,17 @@
>  !
>  ! Contributed by G. Steinmetz
>
> -! { dg-additional-options -Wuninitialized }
> +! { dg-additional-options "-Wuninitialized -O0" }
>
>  program p
>     character :: c(2) = 'a'
>     character, allocatable :: z(:)
>     ! { dg-note {'z' declared here} {} { target *-*-* } .-1 }
>     !$acc parallel
> -   ! { dg-warning {'z\.dim\[0\]\.ubound' is used uninitialized} {} { target *-*-* } .-1 }
> -   ! { dg-warning {'z\.dim\[0\]\.lbound' is used uninitialized} {} { target *-*-* } .-2 }
> +   ! { dg-warning {'z\.dim\[0\]\.ubound' may be used uninitialized} {} { target *-*-* } .-1 }
> +   ! { dg-warning {'z\.dim\[0\]\.lbound' may be used uninitialized} {} { target *-*-* } .-2 }
>     !$omp target
> -   z = c
> +   if (allocated(z)) z = c
>     !$acc end parallel
>     !$omp end target
>     print *, z
-----------------
Siemens Electronic Design Automation GmbH; Anschrift: Arnulfstraße 201, 80634 München; Gesellschaft mit beschränkter Haftung; Geschäftsführer: Thomas Heurung, Frank Thürauf; Sitz der Gesellschaft: München; Registergericht München, HRB 106955
  

Hi Thomas, hi all,

(Updated patch attached – only changes the goacc testcases.
I intent to commit the patch tomorrow, unless more comments come up.)

On 10.03.22 10:00, Thomas Schwinge wrote:

[OpenACC testcases:]

> I recently added that checking in ... [...] to document the status quo.
> (1), as you have proposed, add '-O0' (but with source code comment, please); [...]
> ..., or (2): update the test cases [...] (rationale: the test cases
> haven't originally been written for the '-Wuninitialized' diagnostics; [...]
> (3): duplicate the test cases to account for both (1) and (2)
> [...] (3), would be my approach.

Attached patch does (3). I also remove the code tweaking, added in previous
patch. - But added a bunch of comments.

And I have to admit that I did not realize that the -Wuninitialized was only
added later. (I did not expect that new flags get added to existing patches.)

>> ** I am actually not sure whether 'acc update(b)' will/should map a
>> previous allocated variable - or whether it should. [...]
>> testcases.

Should be: "previously *un*allocated" (+ Thomas s%...%.. comments).

> I don't quickly dig that, sorry.  Do we need to first clarify that with
> OpenACC Technical Committee, or is this just a GCC/OpenACC implementation
> issue?

That's mostly an OpenACC spec question.
But I did not check what the spec says. Thus, I don't know
* whether the spec needs to be improved
* what the implications are on the implementation.

I assume that the implementation for OpenMP does also make sense for
OpenACC (i.e. either works as required or does more but in a sensible
way) - but I don't know.

I think either/both of us should check the OpenACC spec.

  * * *

On the OpenMP side (clarified in 5.1 or 5.2):
* Map first an unallocated allocatable
=> That one is regarded as mapped/present
* Allocate it and map it again (e.g. implicit/explicit
   mapping for a target region)
=> Spec: Implementation may or may not update the item.
    However, (only) with 'always' the spec guarantees that it
    will also show up as allocated on the device.
=> Sentiment: should also work without 'always' modifier
    if previously unallocated. (Discussion postponed.)
* I have not checked 'update' but I think it behaves
   like 'map(always,to/from:' (except for ref counting)

OpenMP GCC implementation: there is a known issue for
scalar allocatables (PR96668). It does work for arrays
(also without 'always') - and 'omp update' has not been
checked. (Ref counting issues?)

OpenACC GCC implementation: I think this code is shared
with OpenMP and, thus, works likewise. But I have have
also not checked this.

Tobias
-----------------
Siemens Electronic Design Automation GmbH; Anschrift: Arnulfstraße 201, 80634 München; Gesellschaft mit beschränkter Haftung; Geschäftsführer: Thomas Heurung, Frank Thürauf; Sitz der Gesellschaft: München; Registergericht München, HRB 106955