------- Comment #2 from rguenth at gcc dot gnu dot org 2010-07-19 14:43 -------
Ick. The problem is that with MEM_REF we now have local variables that
are is_gimple_reg () but not in SSA form.
void
foo (void)
{
int i = 0;
bar (*(struct S *) &i);
}
is undefined and thus we remove the initializer and expand from
foo ()
{
int i;
# BLOCK 2 freq:10000
# PRED: ENTRY [100.0%] (fallthru,exec)
bar (MEM[(struct S *)&i]);
return;
# SUCC: EXIT [100.0%]
}
where when processing the remaining local decls we stop processing 'i' at
/* Expanded above already. */
if (is_gimple_reg (var))
{
TREE_USED (var) = 0;
goto next;
}
the issue is probably latent on the 4.5 branch if we manage to end up with
bar (VIEW_CONVERT <struct S>(i));
instead.
We could rewrite i into SSA form but would have troubles with even more
undefined code which would require a VIEW_CONVERT_EXPR to a different
sized object (which we now reject during gimple validation - the
view-converted use is the reason we don't rewrite it in the first place).
Another long-standing idea was to extend the use of DECL_GIMPLE_REG_P
to all variables (including those that are always registers). That
enables a check for whether a variable is in SSA form.
OTOH update-address-taken is to properly honor this and it does, it
is just missing handling of call arguments. I have a patch.
--
rguenth at gcc dot gnu dot org changed:
What |Removed |Added
----------------------------------------------------------------------------
AssignedTo|unassigned at gcc dot gnu |rguenth at gcc dot gnu dot
|dot org |org
Status|UNCONFIRMED |ASSIGNED
Component|rtl-optimization |tree-optimization
Ever Confirmed|0 |1
Last reconfirmed|0000-00-00 00:00:00 |2010-07-19 14:43:24
date| |
Target Milestone|--- |4.6.0
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=44988
