https://gcc.gnu.org/bugzilla/show_bug.cgi?id=79720
--- Comment #4 from Thomas Koenig <tkoenig at gcc dot gnu.org> --- (In reply to Dominique d'Humieres from comment #3) > Where is computed 1./a? AFAICT the roundoff errors difference with > optimization is restricted to this computation. Yep, you're right... seems that constant of complex division is off. So here is a new, much shorter test case: $ cat r3.f90 program main complex :: c1, c2, r1, r2 character(len=50) :: inp c1 = (12e10, -3.2) inp = '(12e10, -3.2)' read (unit=inp, fmt=*) c2 ! Hide the assignment from the optimizer print *,'c1 - c2 = ', c1-c2 ! Should be zero r1 = 1/c1 ! Compile-time assignment r2 = 1/c2 ! Run-time assignment print *,'r1 = ', r1 print *,'r2 = ', r2 print *,'r1 - r2 = ', r1-r2 ! Should be zero, but is not. end program main $ gfortran -O r3.f90 $ ./a.out c1 - c2 = (0.00000000,0.00000000) r1 = (8.333333155E-12,2.222222180E-22) r2 = (8.333333155E-12,2.222221927E-22) r1 - r2 = (0.00000000,2.524354897E-29)
