Re: [eigen] can't reproduce gcc 4.3.4 errors

[Gael Guennebaud <gael.guennebaud@xxxxxxxxx>]

More news:

in util/XprHelper.h we have:

// just a workaround because GCC seems to not really like empty structs
#ifdef __GNUG__
  #define EIGEN_EMPTY_STRUCT_CTOR(X) \
    EIGEN_STRONG_INLINE X() {} \
    EIGEN_STRONG_INLINE X(const X&) {}
#else
  #define EIGEN_EMPTY_STRUCT_CTOR(X)
#endif

and all our empty functors are declared with it:

template<typename Scalar> struct ei_scalar_opposite_op {
  EIGEN_EMPTY_STRUCT_CTOR(ei_scalar_opposite_op)
  ...
};

If I remove this trick, i.e., by defining EIGEN_EMPTY_STRUCT_CTOR(X)
to nothing, then I get only 2 errors (with -O2):

guenneba@devel10:~/eigen/build/test> ./adjoint_1
Initializing random number generator with seed 1277814460
Repeating each test 10 times
Test adjoint(Matrix<float, 1, 1>()) failed in
"/home/guenneba/eigen/test/adjoint.cpp" (71)
    ei_isApprox((s1 * v1 + s2 * v2).dot(v3), ei_conj(s1) * v1.dot(v3)
+ ei_conj(s2) * v2.dot(v3), largerEps)

guenneba@devel10:~/eigen/build/test> ./eigensolver_complex_3
Initializing random number generator with seed 1277814463
Repeating each test 10 times
Test eigensolver(Matrix<std::complex<float>, 1, 1>()) failed in
"/home/guenneba/eigen/test/eigensolver_complex.cpp" (87)
    test_ei_isApprox(id.operatorNorm(), RealScalar(1))

I don't remember exactly why I've introduced this trick, probably
because I found GCC produced poor code without it. This is still very
strange.

Marking these empty constructors with EIGEN_DONT_INLINE also fix the errors..


Here is the short program I used to debug this:

#include <Eigen/Core>

typedef std::complex<float> T;
template <typename A> T foo(const A& a) { return (-a-a)(2); }

int main()
{
  Eigen::Matrix<T,3,1> n1(0,0,0);
  return int(Eigen::ei_real(foo(n1)));
}

(You need to run it under valgrind to see the issue)

if foo is marked with EIGEN_DONT_INLINE, then no error
if I manually inline foo, s/foo(n1)/(-n1-n1)(2), then no error
if I do only a+a or -a, then no error
if I output (-a-a)(2) at the beginning of foo, then the error still
occurs after the output.


gael

On Tue, Jun 29, 2010 at 1:14 PM, Benoit Jacob <jacob.benoit.1@xxxxxxxxx> wrote:
> anyway, it sounds like either Eigen or some of its unit tests are
> unsafe wrt strict aliasing!!
>
> If it's Eigen itself, we have no choice but to fix that, since strict
> aliasing is part of -O2 in gcc...
>
> Benoit
>
> 2010/6/29 Benoit Jacob <jacob.benoit.1@xxxxxxxxx>:
>> hehe, this sounds like the problem that Sune had the other day on IRC. CC'ing.
>>
>> Benoit
>>
>> 2010/6/29 Gael Guennebaud <gael.guennebaud@xxxxxxxxx>:
>>> hm,
>>>
>>> -O2 -fno-strict-aliasing (and removing the hardcoded -fstrict-aliasing
>>> from our cmakefile)
>>>
>>> 100% tests passed, 0 tests failed out of 453
>>>
>>> much better.
>>>
>>> gael
>>>
>>> On Tue, Jun 29, 2010 at 10:55 AM, Gael Guennebaud
>>> <gael.guennebaud@xxxxxxxxx> wrote:
>>>> ok here,  with -O2 -fno-guess-branch-probability, 4 tests fail as follow:
>>>>
>>>> guenneba@devel10:~/eigen/build/test> ./array_2
>>>> Initializing random number generator with seed 1277801661
>>>> Repeating each test 10 times
>>>> Test comparisons(Array22f()) failed in
>>>> "/home/guenneba/eigen/test/array.cpp" (136)
>>>>    test_ei_isApprox(((m1.abs()+1)>RealScalar(0.1)).colwise()..count(),
>>>> ArrayOfIndices::Constant(cols,rows).transpose())
>>>>
>>>> guenneba@devel10:~/eigen/build/test> ./array_3
>>>> Initializing random number generator with seed 1277801662
>>>> Repeating each test 10 times
>>>>
>>>>    actual   = 2.66729
>>>>    expected = 2.53391
>>>>
>>>>
>>>>    actual   = 1.17816
>>>>    expected = 1.93492
>>>>
>>>>
>>>>    actual   = -2.79358
>>>>    expected = -0.758141
>>>>
>>>>
>>>>    actual   = -2.06046
>>>>    expected = -2.54951
>>>>
>>>>
>>>>    actual   = -1.27569
>>>>    expected = 0.919771
>>>>
>>>>
>>>>    actual   = 5.67633
>>>>    expected = 4.85775
>>>>
>>>>
>>>>    actual   = -5.2517
>>>>    expected = -3.85892
>>>>
>>>>
>>>>    actual   = -1.33353
>>>>    expected = -0.700732
>>>>
>>>>
>>>>    actual   = -3.33972
>>>>    expected = 0.721723
>>>>
>>>>
>>>>    actual   = -1.61675
>>>>    expected = 0.256261
>>>>
>>>> Segmentation fault
>>>>
>>>> guenneba@devel10:~/eigen/build/test> ./geo_hyperplane_4
>>>> Initializing random number generator with seed 1277801668
>>>> Repeating each test 10 times
>>>> Segmentation fault
>>>>
>>>> guenneba@devel10:~/eigen/build/test> ./geo_parametrizedline_4
>>>> Initializing random number generator with seed 1277801673
>>>> Repeating each test 10 times
>>>> Segmentation fault
>>>>
>>>>
>>>> gael
>>>>
>>>> On Tue, Jun 29, 2010 at 10:29 AM, Hauke Heibel
>>>> <hauke.heibel@xxxxxxxxxxxxxx> wrote:
>>>>> Ok, I can try to run the complete tests too. Here are also the flags I
>>>>> used to run (so far just a few of) the failing tests.
>>>>>
>>>>> CXX:               /usr/bin/c++-4.exe
>>>>> CXX_VERSION:       c++-4 (GCC) 4.3.4 20090804 (release) 1
>>>>> CXX_FLAGS:          -Wnon-virtual-dtor -Wno-long-long -ansi -Wundef
>>>>> -Wcast-align -Wchar-subscripts -Wall -W -Wpointer-arith
>>>>> -Wwrite-strings -Wformat-s
>>>>> ecurity -fexceptions -fno-check-new -fno-common -fstrict-aliasing
>>>>> -Wno-variadic-macros -Wextra -pedantic
>>>>>
>>>>> - Hauke
>>>>>
>>>>> On Tue, Jun 29, 2010 at 10:25 AM, Gael Guennebaud
>>>>> <gael.guennebaud@xxxxxxxxx> wrote:
>>>>>> On Tue, Jun 29, 2010 at 10:14 AM, Hauke Heibel
>>>>>> <hauke.heibel@xxxxxxxxxxxxxx> wrote:
>>>>>>> On Mon, Jun 28, 2010 at 11:15 PM, Gael Guennebaud
>>>>>>> <gael.guennebaud@xxxxxxxxx> wrote:
>>>>>>>> I believe there is a bug in gcc, here is the related bug report:
>>>>>>>> http://gcc.gnu.org/bugzilla/show_bug.cgi?id=42255
>>>>>>>> but no fix yet.
>>>>>>>
>>>>>>> That means we can get rid of the errors by disabling -fprofile-arcs?
>>>>>>
>>>>>> A while ago, such errors appeared only with  -fprofile-arcs, i.e.,
>>>>>> when coverage testing was enabled. But now, even without
>>>>>> -fprofile-arcs we hit these weird errors. After some investigations, I
>>>>>> found that disabling the guess-branch-probability optimization that is
>>>>>> automatically enabled with -O2 solved a couple of errors:
>>>>>>
>>>>>> -O2 -fno-guess-branch-probability
>>>>>>
>>>>>> Actually, guess-branch-probability and -fprofile-arcs seems to be
>>>>>> quite related. I've updated the bug report with that respect, but news
>>>>>> so far.
>>>>>>
>>>>>> I'm currently compiling with -O2 -fno-guess-branch-probability to see
>>>>>> if there remain other errors....
>>>>>>
>>>>>>
>>>>>> gael
>>>>>>
>>>>>>> - Hauke
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>
>>>>>>
>>>>>>
>>>>>
>>>>>
>>>>>
>>>>
>>>
>>>
>>>
>>
>
>
>