Serious problems with tree-ssa operand handling

[Diego Novillo]

On Tue, 2004-09-07 at 07:47, Richard Kenner wrote:

> As I said originally, this puts a pointer into the stack of that function
> into the operands of the statement.  That causes an access of junk when
> the operands are referenced.
> 
No, it does not.  VAR is not a GIMPLE register, follow what happens to
non-GIMPLE registers in add_stmt_operand:

   1424 add_stmt_operand (tree *var_p, tree stmt, int flags)
   1425 {
[ ... ]
   1430
   1431   var = *var_p;
   1432   STRIP_NOPS (var);
[ ... ]
   1445   is_real_op = is_gimple_reg (var);
   1446   if (!is_real_op && !DECL_P (var))
   1447     var = get_virtual_var (var);
[ ... ]
   1463
   1464   if (is_real_op)
   1465     {
   1466       /* The variable is a GIMPLE register.  Add it to real operands.  */
   1467       if (flags & opf_is_def)
   1468         append_def (var_p);
   1469       else
   1470         append_use (var_p);
   1471     }
   1472   else
   1473     {
[ ... ]
   1497                   append_v_may_def (var);
[ ... ]
   1509                   append_v_must_def (var);
   1510                 }
[ ...]
   1515                   append_v_may_def (var);
[ ... ]
   1520               append_vuse (var);
[ ... ]
   1543                 append_v_may_def (var);
   1544
[ ... ]
   1546                 append_v_may_def (VARRAY_TREE (aliases, i));
[ ... ]
   1556                 append_vuse (var);
   1557
[ ... ]
   1559                 append_vuse (VARRAY_TREE (aliases, i));
[ ... ]
   1565      }

I have elided most of the logic.  But you see that we only ever use
VAR_P if we are given a real operand (i.e., a symbol that
is_gimple_reg).  The only reason we take the address of the variable is
because for GIMPLE regs we want to store a pointer into the operand so
that we can simplify operand replacement.

We could probably split add_stmt_operand in two so that we have a
version for GIMPLE regs and a version for others.  The downside is that
some of the preparatory logic in add_stmt_operands may have to move to
the callers.


> But I shouldn't have to put in that kludge and disable both checks.
> 
No.  Both kludges are papering over the real problem and are wrong.  I
would need a test case.  Do you have a variable that is a GIMPLE reg and
at the same time belongs to an alias set or its address is taken?


Diego.