Allocatable components in gfortran

[FX Coudert]

Hi Matthew,

For the short story, please jump to the last paragaph of this mail.  
For the even shorter story: not a mingw-specific bug and already  
fixed by Paul & Erik.

I've trimmed down your testcase to:
-------------------------------
   implicit none
   type mumble
     integer, allocatable :: arr(:)
   end type mumble

   type(mumble) :: m
   integer :: i

   do i = 1, 100000
     m = make_mumble()
     deallocate(m%arr)
   enddo
   print *, i

contains
   function make_mumble() result(M)
     type(mumble) :: M
     allocate(M%arr(100000))
   end function make_mumble
end
-------------------------------

I diff-ed the code generated on i386-mingw32 (GCC mainline  
2006-10-11) and i686-linux (2006-10-21), and the content of the do- 
loop is translated into:

i686-linux:
>             {
>               int4 D.921;
>               void * * D.920;
>
>               D.920 = &m.arr.data;
>               _gfortran_deallocate (D.920, &D.921);
>             }
>             m = make_mumble ();
>             {
>               void * * D.922;
>
>               D.922 = &m.arr.data;
>               _gfortran_deallocate (D.922, 0B);
>             }

i386-mingw32:
>             {
>               int4 D.910;
>               void * * D.909;
>
>               D.909 = &m.arr.data;
>               _gfortran_deallocate (D.909, &D.910);
>             }
>             m = make_mumble ();
>             make_mumble ().arr.data = 0B;
>             {
>               void * * D.911;
>
>               D.911 = &m.arr.data;
>               _gfortran_deallocate (D.911, 0B);
>             }

So, the de-allocation of m%arr is indeed performed correctly, for  
some reason the code generated on i386-mingw is trying to nullify the  
return array, but does that by calling the make_mumble function once  
again, which is of course where the memory leak occurs.

Now, this appears to not be a mingw-specific problem, as a 2006-10-11  
i686-linux compiler (luckily, I had one handy) does emit the same  
wrong code. Thus, your problem will be fixed when I upload the next  
gfortran binaries, which is (if all goes well) in 4 to 10 hours from  
now.

Regards,
FX