Bug in dlasd4 in 3.4.2

[andreasnoackjensen]

This small program gives either sigma=NaN or the wrong answer 0.5 in LAPACK 3.4.2 but not in previous versions. Also, the answer changes if I comment out the write statements.

program test

implicit none

double precision :: v(5), z(5), delta(5), work(5), sigma, rho, dnrm2
integer :: info

v = (/0.1d0, 0.2d0, 0.3d0, 0.4d0, 0.5d0/)
z = (/0.1d0, 0.2d0, 0.3d0, 0.4d0, 0.5d0/)
rho = dnrm2(5_8, z, 1_8)
write(*,*) "z", z
z = z / rho
rho = rho**2
write(*,*) "v", v
write(*,*) "z", z
write(*,*) "rho", rho
call dlasd4(5_8, 5_8, v, z, delta, rho, sigma, work, info)
write(*,*) "sigma", sigma
write(*,*) "info", info
end program

[f9JTu83o]

I'm getting wrong results from ZGESDD with LAPACK 3.4.2. The singular values are incorrect and there are some NANs as well. LAPACK 3.4.1 gives correct results. I found this thread so I tried to replace dlasd4.f with the 3.4.1 version. It has fixed the problem. So there really seems to be a serious bug in DLASD4.