equivalence really broken :-)

[Jakub Jelinek]

On Fri, Oct 14, 2005 at 04:56:06PM +0200, Fran?ois-Xavier Coudert wrote:
> Well, Paul Thomas sent me the attached file (Polyhedron Benchmark
> Tests), which does segfault at execution with current CVS gfortran.
> The segfault doesn't happen any more if pool and pool3 are not
> equivalenced any more, but initialized separately (remove the
> equivalence and add "CALL RANDOM_NUMBER(pool3)" just after "CALL
> RANDOM_NUMBER(pool)").
> 
> Using nightly builds on i686-linux, the segfault appeared between
> 20051012 and 20051013. I don't have time to reduce it any further, but
> here it is.

Yeah, I can reproduce it.  Smaller testcase:
  integer, dimension (8388608) :: i, j, k
  integer :: t
  equivalence (i(1), j(10))
  i = 5
  k = i
  call foo (k)
  t = sum (k - i)
  write (*, fmt='(i10)') t
contains
  subroutine foo (x)
    integer, dimension (8388608) :: x
    x = 6
  end subroutine
end

It seems GCC decides to create some huge artifical temporaries
(not the vars created in trans-common.c) and if you don't have sufficiently
high ulimit -s, you will run out of stack.  Will look into this.

	Jakub