[eigen] Incorrect result when inverting 4x4 Matrix with gcc7.3 and -ffast-math

[Michael Mößner <michael.moessner@xxxxxx>]

Hello

I get incorrect results, when I invert a 4x4 matrix with gcc7.3 and the compiler option -ffast-math. Unfortunatly, I don't get access to the bug tracker, so I post it here.

Example code: (compile with "g++ test.cpp -O3 -ffast-math")

#include <Eigen/Dense>
#include <iostream>
using namespace std;

int main(int argc, char **argv)
{
    Eigen::Matrix<double,4,4> test;
    test(0,0) = 1.; test(0,1) = 0.; test(0,2) = 0.; test(0,3) = 0.;
    test(1,0) = 0.; test(1,1) = 1.; test(1,2) = 0.; test(1,3) = 0.;
    test(2,0) = 0.; test(2,1) = 0.; test(2,2) = 1.; test(2,3) = 0.;
    test(3,0) = 0.; test(3,1) = 0.; test(3,2) = 0.; test(3,3) = 1.;

    cerr << test.inverse() << endl;
    return 1;
}

------------------------------

This results in

1  0  0  0
-0 -1 -0 -0
 0  0  1  0
-0 -0 -0 -1

------------------------------

The expected result is

1 0 0 0
0 1 0 0
0 0 1 0
0 0 0 1

---------------------------------

I also tested this code with gcc4.8.5, gcc5.2 and clang6.0.0 where the problem does NOT occur.

I also tested with gcc7.2 where the problem occurs.

The problem does NOT occur when ommitting the -ffast-math or when compiling with -O0

My Eigen version:

3.3.4

My system data:

SUSE Linux Enterprise Desktop 12 SP2

Intel(R) Xeon(R) CPU E3-1276 v3 @ 3.60GHz

---------------------------

I also tested the following systems where the bug also occurs:

----------------------------

SUSE Linux Enterprise Desktop 11 SP2

Intel(R) Xeon(R) CPU E3-1276 v3 @ 3.60GHz

------------------------------

SUSE Linux Enterprise Desktop 12 SP2

Intel(R) Xeon(R) CPU E3-1230 @ 3.20GHz

------------------------------

Best regards

Michael