Re: [eigen] Matrix product crashes when compiled with MSVC 2010 in release

[Benoit Jacob <jacob.benoit.1@xxxxxxxxx>]

2010/8/12 Ilya Baran <baran37@xxxxxxxxx>:
> Hi,
>
> I've run into a nasty crash with recent Eigen dev builds.  The
> following code reproduces it:
>
> void test()
> {
>    typedef Eigen::MatrixXd Mat;
>    Mat r = Mat::Random(5, 2);
>    Mat r2 = Mat::Random(2, 5);
>    Mat r3 = r * r2;
>    std::cout << r3;
> }
>
> It crashes only when the output matrix has an odd number of entries
> and only when built in release mode on MSVC 2010 and only when whole
> program optimization is turned off.  I suspect that this is a compiler
> bug, but perhaps it can be worked around.  I traced the crash to
> changeset 3260 (431547872cf7); 3259 works.

Thanks for your very detailed report, I am CC'ing Gael to ping him.

>
> Here's the backtrace (lines as in rev 3260, because it's a release
> build, debug info is unreliable):
>
> First-chance exception at 0x00ea34bd in EigenTest.exe: 0xC0000005:
> Access violation reading location 0xffffffff.

Hm, this is really, really strange. So MSVC claims that we are
dereferencing the 0xffffffff pointer. I've never seen such a bug in
Eigen.

>
>        EigenTest.exe!Eigen::ei_gebp_kernel<double,double,int,4,4,0,0>::operator()(double
> * res=0x004d3e20, int resStride=5, const double * blockA=0x0037fc70,
> const double * blockB=0x0037fc20, int rows=5, int depth=2, int cols=5,
> double alpha=1.0000000000000000, int strideA=2, int strideB=2, int
> offsetA=0, int offsetB=0, double * unpackedB=0x0037fba0)  Line 459 +
> 0x8 bytes       C++

If I understand well what you write above, this line number 459 is
unreliable because it's a release build? Indeed, there's no way that
this line could result in this 0xffffffff dereferencing:

                  C1 = ei_pmadd(C1, alphav, R1);

this should be manipulating just registers; at worst, variables on the stack.

If we want to investigate this further, we really need a reliable line
number. If MSVC can't give that, could you find it by stepping through
this function in a debugger, or by bisecting it by adding cerr's, or
whatever you favorite artisanal debugging tool is?

Cheers,
Benoit

>
>        EigenTest.exe!Eigen::ei_general_matrix_matrix_product<int,double,0,0,double,0,0,0>::run(int
> rows=5, int cols=5, int depth=2, const double * _lhs=0x004d3d20, int
> lhsStride=5, const double * _rhs=0x004d3da0, int rhsStride=2, double *
> res=0x004d3e20, int resStride=5, double alpha=1.0000000000000000,
> Eigen::ei_level3_blocking<double,double> & blocking={...},
> Eigen::GemmParallelInfo<int> * info=0x00000000)  Line 203 + 0x37
> bytes   C++
>
>        EigenTest.exe!Eigen::GeneralProduct<Eigen::Matrix<double,-1,-1,0,-1,-1>,Eigen::Matrix<double,-1,-1,0,-1,-1>,5>::scaleAndAddTo<Eigen::Matrix<double,-1,-1,0,-1,-1>
>>(Eigen::Matrix<double,-1,-1,0,-1,-1> & dst={...}, double
> alpha=1.0000000000000000)  Line 444     C++
>
>        EigenTest.exe!Eigen::ProductBase<Eigen::GeneralProduct<Eigen::Matrix<double,-1,-1,0,-1,-1>,Eigen::Matrix<double,-1,-1,0,-1,-1>,5>,Eigen::Matrix<double,-1,-1,0,-1,-1>,Eigen::Matrix<double,-1,-1,0,-1,-1>
>>::evalTo<Eigen::Matrix<double,-1,-1,0,-1,-1>
>>(Eigen::Matrix<double,-1,-1,0,-1,-1> & dst={...})  Line 110 + 0x83
> bytes   C++
>
>        EigenTest.exe!test()  Line 9 + 0x97 bytes       C++
>
>        EigenTest.exe!main()  Line 16   C++
>
>
> Figuring out ei_gebp_kernel looks daunting--do you (well, Gael is the
> person who needs to look at this, I expect) need any additional info?
>
> Thanks,
>
>   -Ilya
>
>
>