[Bug libstdc++/102882] New: [AIX] 23_containers 96088 testsuite failures

[Bug libstdc++/102882] New: [AIX] 23_containers 96088 testsuite failures

[dje at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=102882

            Bug ID: 102882
           Summary: [AIX] 23_containers 96088 testsuite failures
           Product: gcc
           Version: unknown
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: libstdc++
          Assignee: unassigned at gcc dot gnu.org
          Reporter: dje at gcc dot gnu.org
  Target Milestone: ---

The 96088 testcases are failing on AIX.

Does this testcase require overriding operator new[] in the library itself, not
only the testcase? The default build options for libstdc++ on AIX do not permit
operator overloading. Maybe it requires XFAIL on AIX.

/nasfarm/edelsohn/src/src/libstdc++-v3/testsuite/23_containers/unordered_set/960
88.cc:42: void test01(): Assertion '__gnu_test::counter::count() == 3' failed.
operator new is called 
operator new is called 
FAIL: 23_containers/unordered_set/96088.cc execution test

/nasfarm/edelsohn/src/src/libstdc++-v3/testsuite/23_containers/unordered_map/960
88.cc:42: void test01(): Assertion '__gnu_test::counter::count() == 3' failed.
operator new is called 
operator new is called 
FAIL: 23_containers/unordered_map/96088.cc execution test

/nasfarm/edelsohn/src/src/libstdc++-v3/testsuite/23_containers/unordered_multima
p/96088.cc:43: void test01(): Assertion '__gnu_test::counter::count() == 3'
fail
ed.
operator new is called 
operator new is called 
FAIL: 23_containers/unordered_multimap/96088.cc execution test

/nasfarm/edelsohn/src/src/libstdc++-v3/testsuite/23_containers/unordered_multise
t/96088.cc:43: void test01(): Assertion '__gnu_test::counter::count() == 3'
fail
ed.
operator new is called 
operator new is called 
FAIL: 23_containers/unordered_multiset/96088.cc execution test

[Bug libstdc++/102882] [AIX] 23_containers 96088 testsuite failures

[redi at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=102882

--- Comment #1 from Jonathan Wakely <redi at gcc dot gnu.org> ---
(In reply to David Edelsohn from comment #0)
> Does this testcase require overriding operator new[] in the library itself,
> not only the testcase?

Yes, I think so. The allocations done by the std::string objects will be done
by the explicit instantiations of std::string inside the library.

Maybe the real solution is to stop explicitly instantiating class templates in
the library. If std::string was instantiated in the tests then all the
allocations would be done in the tests, and the 'operator new' replacement
would work.

That seems like it would remove a whole class of problems where the allocations
and deallocations might happen in different binary objects. The downside would
be slower compilation, because every object file would re-instantiate the same
templates. Avoiding that cost is the purpose of the explicit instantiations,
but if they cause incorrect behaviour, maybe the cost is worth it.

[Bug libstdc++/102882] [AIX] 23_containers 96088 testsuite failures

[redi at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=102882

--- Comment #2 from Jonathan Wakely <redi at gcc dot gnu.org> ---
Created attachment 51646
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=51646&action=edit
Change test to not use std::string

This patch changes the testcase to use a custom instantiation of
std::basic_string, so it doesn't use the explicit instantiation definitions in
the library. Does this make the patched test PASS?

But avoiding using std::string in our own testsuite is a major pain, we should
just make it so it works.

[Bug libstdc++/102882] [AIX] 23_containers 96088 testsuite failures

[dje at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=102882

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
     Ever confirmed|0                           |1
             Status|UNCONFIRMED                 |NEW
   Last reconfirmed|                            |2021-10-22

--- Comment #3 from David Edelsohn <dje at gcc dot gnu.org> ---
The patch for unordered_set fixes that testcase.

[Bug libstdc++/102882] [AIX] 23_containers 96088 testsuite failures

[redi at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=102882

--- Comment #4 from Jonathan Wakely <redi at gcc dot gnu.org> ---
OK, that confirms that the allocations are being done by the std::string
objects inside the library.

But I really don't want to have to use custom allocators every time we use
std::string in such tests.