------- Comment #5 from fxcoudert at gcc dot gnu dot org 2008-02-04 22:51
-------
(In reply to comment #4)
> I did not test your patch, but without the common line, the dump contains
> "gfhzjf" until you remove the volatile. This saves you from reading assembler.
Thanks! This allowed me to realize that my patch didn't work: I was marking the
"fake" variable we create for debugging purposes, not the field of the
structure/union. The following patch works, this time:
Index: trans-common.c
===================================================================
--- trans-common.c (revision 131944)
+++ trans-common.c (working copy)
@@ -318,6 +318,15 @@ build_field (segment_info *h, tree union
GFC_DECL_ASSIGN_ADDR (field) = pushdecl_top_level (addr);
}
+ /* If this field is volatile, mark it. */
+ if (h->sym->attr.volatile_)
+ {
+ tree new;
+ TREE_THIS_VOLATILE (field) = 1;
+ new = build_qualified_type (TREE_TYPE (field), TYPE_QUAL_VOLATILE);
+ TREE_TYPE (field) = new;
+ }
+
h->field = field;
}
It makes the variable volatile in the following examples (common and
equivalence, since they are handled commonly by the codepath I modified):
subroutine wait4it ()
logical event
volatile event
common /dd/ event
event = .false.
do
if (event) print *, 'gfhzjf'
end do
end subroutine
subroutine wait4it2 ()
logical event
integer i
volatile event
equivalence(event,i)
event = .false.
do
if (event) print *, 'gfhzjf'
end do
end subroutine
We don't, however, handle the case where we mark volatile the other variable
involved in the equivalence:
subroutine wait4it3 ()
logical event
integer i
volatile i
equivalence(event,i)
event = .false.
do
if (event) print *, 'gfhzjf'
end do
end subroutine
Do we need it? I'm not sure. I've checked what other compilers do: Intel and
Sun do not optimize out the PRINT statement if variable "i" is marked as
volatile; g95 seems to never optimize out variables in equivalences. I've also
checked what GCC does for C code, and it is does behave the same as my patch
above, ie the following is indeed optimized out:
void wait4it ()
{
union { int event; volatile int i; } dd;
dd.event = 0;
while (1)
{
if (dd.event)
__builtin_puts ("gfhzjf");
}
}
>From the standard point of view, when you equivalence A and B, you're not
supposed to read B after having written to A, so I would think the behaviour I
propose here is standard-conforming. I've asked for opinions on c.l.f
(http://groups.google.com/group/comp.lang.fortran/browse_thread/thread/9602660696e0170e).
--
fxcoudert at gcc dot gnu dot org changed:
What |Removed |Added
----------------------------------------------------------------------------
AssignedTo|unassigned at gcc dot gnu |fxcoudert at gcc dot gnu dot
|dot org |org
Status|NEW |ASSIGNED
GCC build triplet|i686-pc-cygwin |
Keywords| |patch
Last reconfirmed|2008-01-31 13:59:37 |2008-02-04 22:51:44
date| |
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=35037
