Re: cray pointer behavior in 4.6 (bug?)

[Tobias Burnus <burnus at net-b dot de>]

Ryan S. Elliott wrote:
> I have observed some behavior in gfortran 4.6 that does not occur in 
> 4.5 and earlier (as far as I know).  The issue is associated with the 
> use of cray pointers.

The (technical) reason that 4.5 "works" and 4.6 doesn't is that GCC 4.6 
does more optimizations than 4.5 (cf. -fwhole-file, which is the default 
in 4.6).

The question is only whether the compiler has a bug by doing invalid 
optimizations or the program a bug as it relies on invalid assumptions. 
As Cray pointers are not formally defined, giving a definite statement 
is difficult.

It seems that the compiler - at higher optimization levels - does not 
see that the two "oh" and two "obj" alias.

In Fortran 90/95/2003/2008, one has to tell the compiler explicitly that 
a certain variable can alias; one does so by marking those variables 
with the TARGET attribute. If one does so for "oh", the program works. 
(Target also requires an explicit interface, e.g. via a module.)

* * *

However, given that Cray pointers precede Fortran 90, the question is 
how to solve the situation. I think marking automatically (internally) 
the Cray-pointer target as target, could be a viable solution.

The following patch attempts to do so - at least for simple assignments 
such as  "cray_pointer = var", it won't work for "cray_pointer = var + 
4", but I simply hope that that won't occur in practice. Alternatives 
are to walk possibility is to properly walk expressions or - as brute 
force method - to always assume aliasing with -fcray-pointer.

* * *

Early, completely untested draft patch. It probably won't work for the 
example as it sets the normal target attribute. However, the target 
attribute requires an explicit interface ...

--- a/gcc/fortran/resolve.c
+++ b/gcc/fortran/resolve.c
@@ -9303,2 +9303,9 @@ resolve_ordinary_assign (gfc_code *code, 
gfc_namespace *ns)
   gfc_check_assign (lhs, rhs, 1);
+
+  if (lhs->expr_type == EXPR_VARIABLE && rhs->expr_type == EXPR_VARIABLE
+      && lhs->ref == NULL
+      && lhs->symtree->n.sym->attr.cray_pointer
+      && !rhs->symtree->n.sym->attr.cray_pointer)
+    lhs->symtree->n.sym->attr.target = 1;
+
   return false;


Tobias