[Bug target/102215] New: [GCN offloading] Missing '__atomic_compare_exchange_1' etc.

[Bug target/102215] New: [GCN offloading] Missing '__atomic_compare_exchange_1' etc.

[tschwinge at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=102215

            Bug ID: 102215
           Summary: [GCN offloading] Missing '__atomic_compare_exchange_1'
                    etc.
           Product: gcc
           Version: 12.0
            Status: UNCONFIRMED
          Keywords: openmp
          Severity: normal
          Priority: P3
         Component: target
          Assignee: unassigned at gcc dot gnu.org
          Reporter: tschwinge at gcc dot gnu.org
                CC: ams at gcc dot gnu.org, jakub at gcc dot gnu.org, jules at gcc dot gnu.org
  Target Milestone: ---
            Target: GCN

The recent commit 090f0d78f194e3cda23fe904016db77ea36c38fa "openmp: Improve
expand_omp_atomic_pipeline" regresses GCN offloading testing as follows:

    [-PASS:-]{+FAIL:+} libgomp.c/../libgomp.c-c++-common/reduction-16.c (test
for excess errors)
    [-PASS:-]{+UNRESOLVED:+} libgomp.c/../libgomp.c-c++-common/reduction-16.c
[-execution test-]{+compilation failed to produce executable+}

    ld: error: undefined symbol: __atomic_compare_exchange_1
    >>> referenced by /tmp/cce2YauE.o:(test_char._omp_fn.1)
    >>> referenced by /tmp/cce2YauE.o:(test_char._omp_fn.1)
    >>> referenced by /tmp/cce2YauE.o:(test_char._omp_fn.0)
    >>> referenced by /tmp/cce2YauE.o:(test_char._omp_fn.0)

    ld: error: undefined symbol: __atomic_compare_exchange_2
    >>> referenced by /tmp/cce2YauE.o:(test_short._omp_fn.1)
    >>> referenced by /tmp/cce2YauE.o:(test_short._omp_fn.1)
    >>> referenced by /tmp/cce2YauE.o:(test_short._omp_fn.0)
    >>> referenced by /tmp/cce2YauE.o:(test_short._omp_fn.0)
    collect2: error: ld returned 1 exit status
    mkoffload: fatal error:
[...]/gcc/x86_64-pc-linux-gnu-accel-amdgcn-amdhsa-gcc returned 1 exit status

Similar for 'libgomp.c++/../libgomp.c-c++-common/reduction-16.c'.

And:

    [-PASS:-]{+FAIL:+} libgomp.c/target-43.c (test for excess errors)
    [-PASS:-]{+UNRESOLVED:+} libgomp.c/target-43.c [-execution
test-]{+compilation failed to produce executable+}

    ld: error: undefined symbol: __atomic_compare_exchange_1
    >>> referenced by /tmp/cc1Kwlmj.o:(main._omp_fn.0)
    >>> referenced by /tmp/cc1Kwlmj.o:(main._omp_fn.0)
    collect2: error: ld returned 1 exit status
    mkoffload: fatal error:
[...]/gcc/x86_64-pc-linux-gnu-accel-amdgcn-amdhsa-gcc returned 1 exit status

These all do:

    { dg-additional-options "-foffload-options=nvptx-none=-latomic" { target {
offload_target_nvptx } } }

Do we now also need libatomic for GCN or should this be handled in the GCN back
end?

[Bug target/102215] [GCN offloading] Missing '__atomic_compare_exchange_1' etc.

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=102215

--- Comment #1 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
So what did it do for __sync_val_compare_and_exchange_1 ?
Was that expanded inline, or do we have such entrypoint somewhere outside of
libatomic, something else?

[Bug target/102215] [GCN offloading] Missing '__atomic_compare_exchange_1' etc.

[burnus at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=102215

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

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

--- Comment #2 from Tobias Burnus <burnus at gcc dot gnu.org> ---
(In reply to Jakub Jelinek from comment #1)
> So what did it do for __sync_val_compare_and_exchange_1 ?
> Was that expanded inline, or do we have such entrypoint somewhere outside of
> libatomic, something else?

For nvptx, it became more exposed as libgomp.c-c++-common/reduction-16.c
required __sync_val_compare_and_swap_16 – but only on PowerPC – such that
config/nvptx/atomic.c was added, which in turn required
__atomic_compare_exchange_n.
Cf. (libgomp) r11-3182 and (libatomic + config/nvptx/nvptx.md) r11-3145 w/
PR96898 + PR96964.

GCN: It seems as everything required so far was handled inline. I did note that
one could create atomic testcases which are not handled and require __atomic_ /
the nonexisting libatomic library.

Regarding libatomic, it is currently disabled for gcn via
libatomic/configure.tgt:

  *)
        # Who are you?
        UNSUPPORTED=1
        ;;
esac

I don't know what's needed to support it, but for nvptx, it were only a few
lines, cf. r11-3145

[Bug target/102215] [GCN offloading] Missing '__atomic_compare_exchange_1' etc.

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=102215

--- Comment #3 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
Actually, looking at GCN, it is inline e.g. for both:

int
foo (int *p)
{
  return __sync_val_compare_and_swap_4 (p, 1, 2);
}

int
bar (int *p)
{
  int e = 1;
  __atomic_compare_exchange_4 (p, &e, 2, 0, __ATOMIC_SEQ_CST,
__ATOMIC_SEQ_CST);
  return e;
}

but for 1 or 2 byte atomics it is never inline:

int
foo (char *p)
{
  return __sync_val_compare_and_swap_1 (p, 1, 2);
}

int
bar (char *p)
{
  char e = 1;
  __atomic_compare_exchange_1 (p, &e, 2, 0, __ATOMIC_SEQ_CST,
__ATOMIC_SEQ_CST);
  return e;
}


The reason why foo in the second testcase works is libgcc/config/gcn/atomic.c
which defines __sync_{val,bool}_compare_and_swap_{1,2}.

So, either it should also define __atomic_compare_exchange_{1,2} perhaps with
calling __sync_val_compare_and_swap_{1,2} under the hood or the other way
around, or the backend somewhere needs to rewrite calls to
__atomic_compare_exchange_{1,2} into calls to __sync_*_compare_and_swap_{1,2}.

The reason why I've changed omp-expand is that by using __atomic_* APIs one can
use the user-requested memory models, and once the 5.1 atomics support is done,
there will be more - weak flag too, and user specified failure memory models.

[Bug target/102215] [GCN offloading] Missing '__atomic_compare_exchange_1' etc.

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=102215

--- Comment #4 from CVS Commits <cvs-commit at gcc dot gnu.org> ---
The master branch has been updated by Tobias Burnus <burnus@gcc.gnu.org>:

https://gcc.gnu.org/g:450526551dcb97b7c0513699d4333efb79b8b490

commit r12-7567-g450526551dcb97b7c0513699d4333efb79b8b490
Author: Tobias Burnus <tobias@codesourcery.com>
Date:   Wed Mar 9 19:34:48 2022 +0100

    GCN: Implement __atomic_compare_exchange_{1,2} in libgcc [PR102215]

    libgcc/ChangeLog:

            PR target/102215
            * config/gcn/atomic.c (__sync_val_compare_and_swap_##SIZE): Move
            a line up to non-arg-dependent value first.
            (__ATOMIC_COMPARE_EXCHANGE): Define + call to generate
            __atomic_compare_exchange_{1,2}.

[Bug target/102215] [GCN offloading] Missing '__atomic_compare_exchange_1' etc.

[burnus at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=102215

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
         Resolution|---                         |FIXED
             Status|UNCONFIRMED                 |RESOLVED

--- Comment #5 from Tobias Burnus <burnus at gcc dot gnu.org> ---
FIXED with commit r12-7567 on mainline (GCC 12).

(Close as the issue was only exposed by the OpenMP change r12-3325. While
supporting more atomic ops is nice, this does not imply the need to back-port
it.)