Re: [eigen] any remaining issues blocking beta1, speak up!

[Gael Guennebaud <gael.guennebaud@xxxxxxxxx>]

On Sun, Jul 4, 2010 at 5:42 AM, Benoit Jacob <jacob.benoit.1@xxxxxxxxx> wrote:
> Besides the -fstrict-aliasing issue, is there anything else blocking beta1 ?
>
> For the -fstrict-aliasing issue, do I understand correctly that it is
> ONLY with gcc 4.3, and not any other version?

yes this is really specific to this version of GCC. I spent almost a
day on it, and my intuition is that gcc gets confused by the deep
hierarchy and the static cast from the top level expression to, e.g.,
DenseBase, and back to expression via derived(). But that's only an
intuition.

>
> If yes, do you want us to go forward and just release beta1 with a
> "caveat", "exposes what seems to be a bug in gcc 4.3" ?

yes, and the solution is to compile with -O2 -fno-strict-aliasing....

gael

>
> Benoit
>
>
>
> 2010/7/3 Benoit Jacob <jacob.benoit.1@xxxxxxxxx>:
>> Gael: could you add your 1st example as a unit test? I understand that
>> other tests catch the issue already, but having this simple test case
>> can help in the future.
>>
>> 2010/7/3 Benoit Jacob <jacob.benoit.1@xxxxxxxxx>:
>>> I really can't reproduce on linux x86-64 with GCC 4.3.
>>>
>>> I tried compiling with -fstrict-aliasing and also with -O2
>>> I tried both your versions, reading the program's return value like this:
>>>
>>> ##### 10:53:39 ~/build/eigen$ g++-4.3 z.cpp -o z -I ~/eigen -O2 && ./z
>>> && echo "zero"
>>> zero
>>>
>>> Benoit
>>>
>>>
>>>
>>> 2010/7/3 Gael Guennebaud <gael.guennebaud@xxxxxxxxx>:
>>>> no warning.
>>>>
>>>>
>>>> here is an simpler version:
>>>>
>>>> #include <iostream>
>>>> #include <Eigen/Dense>
>>>>
>>>> int main()
>>>> {
>>>>  Eigen::Matrix<float,1,1> v;
>>>>  v.setOnes();
>>>>  return (v.cwiseAbs()-v).sum();
>>>> }
>>>>
>>>> I have also simplified CwiseBinaryOp, such that the coeff(int)
>>>> function is directly implemented in CwiseBinaryOp (not in the base
>>>> class CwiseBinaryOpImpl) and such that it simply does:
>>>>
>>>> return m_rhs.coeff(0);
>>>>
>>>> which, in this example, should return v.coeff(0) == 1;
>>>>
>>>>  (so the left hand side and the functor are not used, but I remove
>>>> them then the issue disappear).
>>>>
>>>> I've also changed the functor by an object of type:
>>>>
>>>> struct foo {
>>>>  foo() : x(1) {}
>>>>  int x;
>>>> };
>>>>
>>>> just be sure the issue does not come from our empty functors. No success.
>>>>
>>>> If the left hand side (here v.cwiseAbs(), which is not used) is nested
>>>> by reference, then the issue disappear.
>>>>
>>>> Also, m_rhs().data() is wrong.
>>>>
>>>> ...
>>>>
>>>> gael
>>>>
>>>> On Sat, Jul 3, 2010 at 3:55 PM, Benoit Jacob <jacob.benoit.1@xxxxxxxxx> wrote:
>>>>> Can you try compiling with
>>>>>
>>>>>    -Wstrict-aliasing=1
>>>>>
>>>>> indeed, the default value is 3, and 1 is the most aggressive warning
>>>>> level. It could however report a lot of false positives. According to
>>>>> `man gcc`.
>>>>>
>>>>> Benoit
>>>>>
>>>>> 2010/7/3 Gael Guennebaud <gael.guennebaud@xxxxxxxxx>:
>>>>>> here is a simplified version:
>>>>>>
>>>>>> #include <iostream>
>>>>>> #include <Eigen/Dense>
>>>>>>
>>>>>> template<typename A, typename B> EIGEN_DONT_INLINE int foo(const A& a,
>>>>>> const B& b)
>>>>>> {
>>>>>>  return (a - b).sum() <= 0;
>>>>>> }
>>>>>>
>>>>>> int main()
>>>>>> {
>>>>>>  Eigen::Matrix<float,1,1> v;
>>>>>>  v.setOnes();
>>>>>>  return foo(v.colwise().maxCoeff(),v);
>>>>>> }
>>>>>>
>>>>>> which, again, fails with gcc-4.3 + strict-aliasing only....
>>>>>> investigating to see if I can find a workaround.
>>>>>>
>>>>>> gael
>>>>>>
>>>>>> On Sat, Jul 3, 2010 at 1:03 PM, Jitse Niesen <jitse@xxxxxxxxxxxxxxxxx> wrote:
>>>>>>> On Sat, 3 Jul 2010, Benoit Jacob wrote:
>>>>>>>
>>>>>>>> All is in the title. Can we release beta1 this Saturday?
>>>>>>>
>>>>>>> There is also the stable_norm_1 test that is failing on gcc 4.3. This seems
>>>>>>> to be a mistake in Eigen itself.
>>>>>>>
>>>>>>> See the previous email I wrote:
>>>>>>>
>>>>>>> On Thu, 1 Jul 2010, Jitse Niesen wrote:
>>>>>>>>
>>>>>>>> Date: Thu, 1 Jul 2010 11:02:51 +0100 (BST)
>>>>>>>> From: Jitse Niesen <jitse@xxxxxxxxxxxxxxxxx>
>>>>>>>> To: Eigen mailinglist <eigen@xxxxxxxxxxxxxxxxxxx>
>>>>>>>> Subject: Failing stable_norm_1 test
>>>>>>>>
>>>>>>>> Hello,
>>>>>>>>
>>>>>>>> I just spent some time on the stable_norm_1 test, which fails on both my
>>>>>>>> computer and Gael's computer when using gcc 4.3. It looks similar to the
>>>>>>>> issue
>>>>>>>> which you solved by the empty struct magic (well done!).
>>>>>>>>
>>>>>>>> The details are as follows. I constructed the following (fairly) minimal
>>>>>>>> test
>>>>>>>> example:
>>>>>>>>
>>>>>>>> #include <iostream>
>>>>>>>> #include <Eigen/Dense>
>>>>>>>>
>>>>>>>> int main()
>>>>>>>> {
>>>>>>>>  Eigen::Matrix<float,1,1> v;
>>>>>>>>  v(0,0) = 0.1;
>>>>>>>>  std::cerr << "blue norm = " << v.colwise().blueNorm() << "\n";
>>>>>>>>  std::cerr << "status = " <<
>>>>>>>> v.colwise().hypotNorm().isApprox(v.colwise().norm(), 1e-4) << "\n";
>>>>>>>> }
>>>>>>>>
>>>>>>>> When compiling with g++ (version 4.3.3) -O2 -g and running under valgrind,
>>>>>>>> this
>>>>>>>> produces:
>>>>>>>>
>>>>>>>> Use of uninitialised value of size 4
>>>>>>>>    at 0x8048F28: bool
>>>>>>>> Eigen::DenseBase<Eigen::PartialReduxExpr<Eigen::Matrix<float, 1, 1, 0, 1,
>>>>>>>> 1>,
>>>>>>>> Eigen::ei_member_hypotNorm<float>, 0>
>>>>>>>> >::isApprox<Eigen::PartialReduxExpr<Eigen::Matrix<float, 1, 1, 0, 1, 1>,
>>>>>>>> Eigen::ei_member_norm<float>, 0>
>>>>>>>> >(Eigen::DenseBase<Eigen::PartialReduxExpr<Eigen::Matrix<float, 1, 1, 0,
>>>>>>>> > 1, 1>,
>>>>>>>> Eigen::ei_member_norm<float>, 0> > const&, float) const (in
>>>>>>>> /scratch/1/users/jitse/work/c/eigen/eigenfoo)
>>>>>>>>    by 0x8048AF5: main (in /scratch/1/users/jitse/work/c/eigen/eigenfoo)
>>>>>>>>
>>>>>>>> and the same error with the same stack trace a bit later, at 0x8048F2A.
>>>>>>>>
>>>>>>>> The result of the program is also wrong: it should print status = 1, but
>>>>>>>> it
>>>>>>>> says status = 0.
>>>>>>>>
>>>>>>>> The error goes away when compiling with -fno-strict-aliasing.
>>>>>>>> I tried compiling with -Wstrict-aliasing=2 but that does not produce any
>>>>>>>> warnings.
>>>>>>>>
>>>>>>>> Over to the gurus!
>>>>>>>>
>>>>>>>> Cheers,
>>>>>>>> Jitse
>>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>
>>>>>>
>>>>>>
>>>>>
>>>>>
>>>>>
>>>>
>>>>
>>>>
>>>
>>
>
>
>