Description of problem:

A complex arithmetic statement is compiled wrongly and produces the
wrong result


Version-Release number of selected component (if applicable):

gcc-g77-3.4.6-3.1

How to reproduce:

      program bugtest
      implicit none
      complex*16 z
      z = dcmplx(1.0, 2.0)
      call sub(z)
      stop
      end program bugtest

      subroutine sub(z)
      implicit none
      complex*16 z
      z = dcmplx(-dimag(z), dreal(z))
      write (6, *) z
      return
      end subroutine sub

If you compile this without optimization (g77 bugtest.f -o bugtest)
and run it you get the output (-2.,-2.) instead of (-2.,1.)

It is pretty obvious what is happening: the important line
(which implements z := iz) is being compiled as

x := -y
y := x

which is pretty stupid.  This problem does NOT occur for the more
modern gfortran compiler, nor does it occur if you use any
optimization flags (-O2, -O3 etc.)

Examination of the machine code shows that indeed the result of
x := -y overwrites x before the value is read for the calculation
of y.