Created attachment 796 [details]
Minimal failing example
When defining EIGEN_USE_MKL and EIGEN_USE_MKL_ALL and linking against Intel MKL 2018.0, SelfAdjointEigenSolver will give wrong results for RowMajor matrices.
Attached is a minimal test case. It computes eigenvectors two times for the same matrix, one time stored in column-major order, the second time in row-major order. When compiled without MKL support, both results are equal. With MKL enabled, however, the result for the row-major matrix is the transpose of the correct result.