------- Comment #7 from rguenth at gcc dot gnu dot org 2009-03-20 10:31 -------
Hm? In the testcase the argument points to NULL. I do not see that
1) assigning sth else to the parameter is forbidden by the optimization switch
2) we prevent the compiler from coalescing an unrelated SSA_NAME to that
same base variable decl of the param.
The problem is that may_alias_p get's a pointer DECL, not an SSA_NAME and
checks
/* If -fargument-noalias-global is > 2, pointer arguments may
not point to anything else. */
if (flag_argument_noalias > 2 && TREE_CODE (ptr) == PARM_DECL)
{
alias_stats.alias_noalias++;
alias_stats.simple_resolved++;
return false;
}
which is very fragile.
We can just delete that check entirely and rely on PTA properly tracking if
a pointer still (only) points to the PARM_NOALIAS variable we created for
the points-to set of the incoming pointer SSA_NAME.
Anyway, I won't force a fix down your throat for 4.4 but will just remove
the above code pieces on a-i branch.
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=39506
