[committed] libstdc++: Fix compare_exchange_padding.cc test for std::atomic_ref

[committed] libstdc++: Fix compare_exchange_padding.cc test for std::atomic_ref

[Jonathan Wakely]

Tested x86_64-linux, pushed to trunk.

The test was only failing for me with -m32 (and not -m64), so I didn't
notice until now. That probably means we should make the test fail more
reliably if the padding isn't being cleared.

-- >8 --

This test was written assuming that std::atomic_ref clears its target's
padding on construction, but that could introduce data races. Change the
test to store a value after construction and check that its padding is
cleared by the store.

libstdc++-v3/ChangeLog:

	* testsuite/29_atomics/atomic_ref/compare_exchange_padding.cc:
	Store value with non-zero padding bits after construction.
---
 .../29_atomics/atomic_ref/compare_exchange_padding.cc | 11 ++++++-----
 1 file changed, 6 insertions(+), 5 deletions(-)

diff --git a/libstdc++-v3/testsuite/29_atomics/atomic_ref/compare_exchange_padding.cc b/libstdc++-v3/testsuite/29_atomics/atomic_ref/compare_exchange_padding.cc
index 1b1a12dddda..e9f8a4bdf2a 100644
--- a/libstdc++-v3/testsuite/29_atomics/atomic_ref/compare_exchange_padding.cc
+++ b/libstdc++-v3/testsuite/29_atomics/atomic_ref/compare_exchange_padding.cc
@@ -20,14 +20,15 @@ int
 main ()
 {
   S s;
-  fill_struct(s);
-  s.c = 'a';
-  s.s = 42;
-
   S ss{ s };
+  fill_struct(ss);
+  ss.c = 'a';
+  ss.s = 42;
+
   std::atomic_ref<S> as{ s };
+  as.store(ss);
   auto ts = as.load();
-  VERIFY( !compare_struct(ss, ts) ); // padding cleared on construction
+  VERIFY( !compare_struct(ss, ts) ); // padding cleared on store
   as.exchange(ss);
   auto es = as.load();
   VERIFY( compare_struct(ts, es) ); // padding cleared on exchange
-- 
2.37.3

Re: [committed] libstdc++: Fix compare_exchange_padding.cc test for std::atomic_ref

[Eric Botcazou]

> The test was only failing for me with -m32 (and not -m64), so I didn't
> notice until now. That probably means we should make the test fail more
> reliably if the padding isn't being cleared.

The tests fail randomly for me on SPARC64/Linux:

FAIL: 29_atomics/atomic/compare_exchange_padding.cc execution test
FAIL: 29_atomics/atomic_ref/compare_exchange_padding.cc execution test

/home/ebotcazou/src/libstdc++-v3/testsuite/29_atomics/atomic_ref/
compare_exchange_padding.cc:34: int main(): Assertion 'compare_struct(ts, es)' 
failed.
FAIL: 29_atomics/atomic_ref/compare_exchange_padding.cc execution test

  std::atomic<S> as{ s };
  auto ts = as.load();
  VERIFY( !compare_struct(s, ts) ); // padding cleared on construction
  as.exchange(s);
  auto es = as.load();
  VERIFY( compare_struct(ts, es) ); // padding cleared on exchange

How is it supposed to pass exactly?  AFAICS you have no control on the padding 
bits of ts or es and, indeed, at -O2 the loads are scalarized:

  __buf$c_81 = MEM[(struct S *)&__buf].c;
  __buf$s_59 = MEM[(struct S *)&__buf].s;
  __buf ={v} {CLOBBER(eol)};
  ts.c = __buf$c_81;
  ts.s = __buf$s_59;
[...]
  __buf$c_100 = MEM[(struct S *)&__buf].c;
  __buf$s_35 = MEM[(struct S *)&__buf].s;
  __buf ={v} {CLOBBER(eol)};
  es.c = __buf$c_100;
  es.s = __buf$s_35;
  _66 = MEM <unsigned int> [(char * {ref-all})&ts];
  _101 = MEM <unsigned int> [(char * {ref-all})&es];
  if (_66 != _101)
    goto <bb 5>; [0.04%]
  else
    goto <bb 6>; [99.96%]

so the result of the 4-byte comparison is random.

-- 
Eric Botcazou

Re: [committed] libstdc++: Fix compare_exchange_padding.cc test for std::atomic_ref

[Jonathan Wakely]

On Mon, 31 Oct 2022 at 15:34, Eric Botcazou <botcazou@adacore.com> wrote:
>
> > The test was only failing for me with -m32 (and not -m64), so I didn't
> > notice until now. That probably means we should make the test fail more
> > reliably if the padding isn't being cleared.
>
> The tests fail randomly for me on SPARC64/Linux:
>
> FAIL: 29_atomics/atomic/compare_exchange_padding.cc execution test
> FAIL: 29_atomics/atomic_ref/compare_exchange_padding.cc execution test
>
> /home/ebotcazou/src/libstdc++-v3/testsuite/29_atomics/atomic_ref/
> compare_exchange_padding.cc:34: int main(): Assertion 'compare_struct(ts, es)'
> failed.
> FAIL: 29_atomics/atomic_ref/compare_exchange_padding.cc execution test
>
>   std::atomic<S> as{ s };
>   auto ts = as.load();
>   VERIFY( !compare_struct(s, ts) ); // padding cleared on construction
>   as.exchange(s);
>   auto es = as.load();
>   VERIFY( compare_struct(ts, es) ); // padding cleared on exchange
>
> How is it supposed to pass exactly?  AFAICS you have no control on the padding
> bits of ts or es and, indeed, at -O2 the loads are scalarized:
>
>   __buf$c_81 = MEM[(struct S *)&__buf].c;
>   __buf$s_59 = MEM[(struct S *)&__buf].s;
>   __buf ={v} {CLOBBER(eol)};
>   ts.c = __buf$c_81;
>   ts.s = __buf$s_59;
> [...]
>   __buf$c_100 = MEM[(struct S *)&__buf].c;
>   __buf$s_35 = MEM[(struct S *)&__buf].s;
>   __buf ={v} {CLOBBER(eol)};
>   es.c = __buf$c_100;
>   es.s = __buf$s_35;
>   _66 = MEM <unsigned int> [(char * {ref-all})&ts];
>   _101 = MEM <unsigned int> [(char * {ref-all})&es];
>   if (_66 != _101)
>     goto <bb 5>; [0.04%]
>   else
>     goto <bb 6>; [99.96%]
>
> so the result of the 4-byte comparison is random.


I suppose we could use memcmp on the as variable itself, to inspect
the actual stored padding rather than the returned copy of it.

Re: [committed] libstdc++: Fix compare_exchange_padding.cc test for std::atomic_ref

[Eric Botcazou]

> I suppose we could use memcmp on the as variable itself, to inspect
> the actual stored padding rather than the returned copy of it.

Yes, that's probably the only safe stance when optimization is enabled.

-- 
Eric Botcazou

Re: [committed] libstdc++: Fix compare_exchange_padding.cc test for std::atomic_ref

[Jonathan Wakely]

On Mon, 31 Oct 2022 at 17:03, Eric Botcazou <botcazou@adacore.com> wrote:
>
> > I suppose we could use memcmp on the as variable itself, to inspect
> > the actual stored padding rather than the returned copy of it.
>
> Yes, that's probably the only safe stance when optimization is enabled.


Strictly speaking, it's not safe, because it's undefined to use memcmp
on an object of a non-trivial type. But it should work.

Re: [committed] libstdc++: Fix compare_exchange_padding.cc test for std::atomic_ref

[Jonathan Wakely]

On Mon, 31 Oct 2022 at 17:05, Jonathan Wakely <jwakely@redhat.com> wrote:
>
> On Mon, 31 Oct 2022 at 17:03, Eric Botcazou <botcazou@adacore.com> wrote:
> >
> > > I suppose we could use memcmp on the as variable itself, to inspect
> > > the actual stored padding rather than the returned copy of it.
> >
> > Yes, that's probably the only safe stance when optimization is enabled.
>
>
> Strictly speaking, it's not safe, because it's undefined to use memcmp
> on an object of a non-trivial type. But it should work.

Do those loads still get scalarized at -O0?

Re: [committed] libstdc++: Fix compare_exchange_padding.cc test for std::atomic_ref

[Eric Botcazou]

> Do those loads still get scalarized at -O0?

Presumably not at the GIMPLE level, but possibly at the RTL level.

-- 
Eric Botcazou

Re: [committed] libstdc++: Fix compare_exchange_padding.cc test for std::atomic_ref

[Jonathan Wakely]

On Mon, 31 Oct 2022 at 15:34, Eric Botcazou wrote:
>
> > The test was only failing for me with -m32 (and not -m64), so I didn't
> > notice until now. That probably means we should make the test fail more
> > reliably if the padding isn't being cleared.
>
> The tests fail randomly for me on SPARC64/Linux:
>
> FAIL: 29_atomics/atomic/compare_exchange_padding.cc execution test
> FAIL: 29_atomics/atomic_ref/compare_exchange_padding.cc execution test
>
> /home/ebotcazou/src/libstdc++-v3/testsuite/29_atomics/atomic_ref/
> compare_exchange_padding.cc:34: int main(): Assertion 'compare_struct(ts, es)'
> failed.
> FAIL: 29_atomics/atomic_ref/compare_exchange_padding.cc execution test
>
>   std::atomic<S> as{ s };
>   auto ts = as.load();
>   VERIFY( !compare_struct(s, ts) ); // padding cleared on construction
>   as.exchange(s);
>   auto es = as.load();
>   VERIFY( compare_struct(ts, es) ); // padding cleared on exchange
>
> How is it supposed to pass exactly?  AFAICS you have no control on the padding
> bits of ts or es and, indeed, at -O2 the loads are scalarized:
>
>   __buf$c_81 = MEM[(struct S *)&__buf].c;
>   __buf$s_59 = MEM[(struct S *)&__buf].s;
>   __buf ={v} {CLOBBER(eol)};
>   ts.c = __buf$c_81;
>   ts.s = __buf$s_59;
> [...]
>   __buf$c_100 = MEM[(struct S *)&__buf].c;
>   __buf$s_35 = MEM[(struct S *)&__buf].s;
>   __buf ={v} {CLOBBER(eol)};
>   es.c = __buf$c_100;
>   es.s = __buf$s_35;
>   _66 = MEM <unsigned int> [(char * {ref-all})&ts];
>   _101 = MEM <unsigned int> [(char * {ref-all})&es];
>   if (_66 != _101)
>     goto <bb 5>; [0.04%]
>   else
>     goto <bb 6>; [99.96%]
>
> so the result of the 4-byte comparison is random.

This should be fixed now. I rewrote the test to check the padding byte
directly, instead of inspecting a copy of it which might not preserve
the padding bits.

Re: [committed] libstdc++: Fix compare_exchange_padding.cc test for std::atomic_ref

[Eric Botcazou]

> This should be fixed now. I rewrote the test to check the padding byte
> directly, instead of inspecting a copy of it which might not preserve
> the padding bits.

Great, thanks!

-- 
Eric Botcazou