Bug 15718 dispute: showstopper or WONTFIX?

[Joe Buck]

On Sat, May 29, 2004 at 07:24:22AM +0200, Gabriel Dos Reis wrote:
> Joe Buck <jbuck@welsh-buck.org> writes:
> 
> | I don't want to get into a battle with RTH, who I have enormous
> | respect for.  But it appears to me that the idea of closing bug
> | 15718 with WONTFIX is not an option.
> | 
> | This bug apparently has the effect that some significant number
> | of Fortran programs that use complex numbers (that is, pretty
> | much any physics or electrical engineering code) are likely to
> | hit the ICE.  This is a new ICE, not in 3.3.3.  Saying that we
> 
> The ICE is new, but the bug is an ABI bug that has been there,
> awaiting to be discovered by some constructs.  Basically 
> passing/returning complex numbers in registers was broken.

Since 3.3.4 shipped with this bug, we need a release note describing, in
user-level terms, how to avoid the bug.

We can't write "don't pass a __complex__ in a register".  After all,
the following Fortran program (g77.f-torture/execute/19990826-2.f)
gets an ICE on the Alpha:

      complex one, z
      real    a, f1
      f1(z) = real(z)
      one = (1.,0.)
      a = f1(one) 
      if ( abs(a-1.0) .gt. 1.0e-5 ) then
         write(6,*) 'A should be 1.0 but it is',a
         call abort()
      end if
      end

g77.f-torture/execute/20010116.f is very similar.

Evidently Fortran inline functions (I forget what these beasties are
officially called) pass complex arguments in registers on the Alpha,
therefore such functions are completely broken.