[Bug middle-end/90859] [OMP] Mappings for VLA different depending on 'target { c && { ! lp64 } }'

[Bug middle-end/90859] [OMP] Mappings for VLA different depending on 'target { c && { ! lp64 } }'

[tschwinge at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=90859

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Keywords|                            |openmp
   Last reconfirmed|2019-06-20 00:00:00         |2020-4-29

--- Comment #3 from Thomas Schwinge <tschwinge at gcc dot gnu.org> ---
I've come across this again.  And again, I'm just documenting this here, don't
have bandwidth right now to figure out what's going on.

(In reply to Thomas Schwinge from comment #0)
> I noticed the following (with OpenACC, not yet tried to reproduce with
> OpenMP 'target').

Now also reproduced for OpenMP.

> Is there something wrong, or can that behavior be
> explained?
> 
> Only 'sizeof array' appears in the offloaded region, yet for 'target { c &&
> { ! lp64 } }' (only!), the gimplifier also creates a mapping for the array
> itself.

Now also reproduced on x86_64-pc-linux-gnu with default '-m64' instead of
'-m32'.

(In reply to Thomas Schwinge from comment #2)
> Experimenting on x86_64-pc-linux-gnu with '-m32', this depends on the data
> type used for 'array_so'.  If I change 'uint32_t' to 'unsigned int', I see
> the same strange behavior.  If however I change it to 'unsigned long', the
> issue goes away, as it does for 'unsigned short', for example.  The code
> inside the region is the same (aside from some type casting); in particular
> there aren't any 'array' references.

Similarly, with test case:

    #include <stdint.h>

    static void
    vla (int array_li)
    {
      _Complex double array[array_li];
      TYPE array_so;
    #pragma omp target \
      map(from:array_so)
      {
        array_so = sizeof array;
      }
      if (array_so != sizeof array)
        __builtin_abort ();
    }

... compiled with '-fopenmp -O0 -fdump-tree-gimple -DTYPE=int64_t', we get:

    #pragma omp target num_teams(1) thread_limit(0) map(from:array_so [len: 8])
firstprivate(array_li.0)
      {
        sizetype D.2157;
        sizetype D.2158;
        long int array_so.2;

        D.2157 = (sizetype) array_li.0;
        D.2158 = D.2157 * 16;
        array_so.2 = (long int) D.2158;
        array_so = array_so.2;
      }

..., but compiled with '-fopenmp -O0 -fdump-tree-gimple -DTYPE=uint64_t', we
get:

    #pragma omp target num_teams(1) thread_limit(0) map(from:array_so [len: 8])
firstprivate(array_li.0) map(tofrom:(*array.1) [len: D.2153])
map(firstprivate:array [pointer assign, bias: 0])
      {
        sizetype D.2157;
        sizetype array_so.2;

        D.2157 = (sizetype) array_li.0;
        array_so.2 = D.2157 * 16;
        array_so = array_so.2;
      }

..., that is, in the latter case, the 'array' itself gets mapped additionally,
even though unused inside the region.

[Bug middle-end/90859] [OMP] Mappings for VLA different depending on 'target { c && { ! lp64 } }'

[burnus at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=90859

Tobias Burnus <burnus at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |burnus at gcc dot gnu.org

--- Comment #4 from Tobias Burnus <burnus at gcc dot gnu.org> ---
Using
// here: float array[array_li];
...
//#pragma omp target defaultmap(none)  map(from:array_so_openmp_target)
#pragma omp parallel default(none)  shared(array_so_openmp_target)
    array_so_openmp_target = sizeof array;

Shows that the latter works – with an implicit 'firstprivate(array_li.0)' while
the former has no firstprivate and fails.

[Bug middle-end/90859] [OMP] Mappings for VLA different depending on 'target { c && { ! lp64 } }'

[burnus at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=90859

--- Comment #5 from Tobias Burnus <burnus at gcc dot gnu.org> ---
(In reply to Tobias Burnus from comment #4)
Ignore; this comment should have gone into PR94874 (whose test case has a
comment about this PR).

[Bug middle-end/90859] [OMP] Mappings for VLA different depending on 'target { c && { ! lp64 } }'

[burnus at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=90859

Tobias Burnus <burnus at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
         Depends on|                            |95002

--- Comment #6 from Tobias Burnus <burnus at gcc dot gnu.org> ---
My just created PR c/95002 (has a patch) is either a duplicate or at least
covers partially this PR. → mark as 'depends on'.


Referenced Bugs:

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=95002
[Bug 95002] VLA:   'var = sizeof array' gives spurous '= array, <size expr>'
instead of just '= <size expr>'

[Bug middle-end/90859] [OMP] Mappings for VLA different depending on 'target { c && { ! lp64 } }'

[dje at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=90859

David Edelsohn <dje at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |dje at gcc dot gnu.org

--- Comment #7 from David Edelsohn <dje at gcc dot gnu.org> ---
I'm seeing a new failure that this string no longer appears in the dump for 32
bit AIX.

[Bug middle-end/90859] [OMP] Mappings for VLA different depending on 'target { c && { ! lp64 } }'

[tschwinge at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=90859

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |jsm28 at gcc dot gnu.org

--- Comment #8 from Thomas Schwinge <tschwinge at gcc dot gnu.org> ---
(In reply to David Edelsohn from comment #7)
> I'm seeing a new failure that this string no longer appears in the dump for
> 32 bit AIX.

According to <gcc-testresults@gcc.gnu.org> posts, this has been FAILing forever
for powerpc-ibm-aix7.2.0.0 (so, not a "new failure" as far as I can tell?),
ever since my 2019-06-19 r272452 '[PR90859] Document status quo for "[OMP]
Mappings for VLA different depending on 'target { c && { ! lp64 } }'"'.  The
problem here relates to front end 'sizetype' internals, for which 'c && { !
lp64 }' doesn't seem to be quite the right conditional.

In PR95002, there is discussion about the actual C/C++ front end and/or
gimplifier changes assumed to be necessary to fix this for good.  Spending time
on that would be more useful than figuring out the right testsuite-level
conditional -- but I don't have a good understanding of the relevant C/C++
language-level requirements, and the relevant front end vs. gimple-level
folding.