On 7/14/07, Eric Botcazou <[EMAIL PROTECTED]> wrote:
> First is_gimple_min_invariant in try_to_simplify where it chooses
> DECL_INITIAL should be valid_gimple_expression_p instead.

That's a known problem, see tree-ssa-ccp.c:

I had forgotten about this because i hadn't done IPA in a while.  At
some point, we should just create special initializer functions and
gimplify these into them.

/* The regular is_gimple_min_invariant does a shallow test of the object.
   It assumes that full gimplification has happened, or will happen on the
   object.  For a value coming from DECL_INITIAL, this is not true, so we
   have to be more strict ourselves.  */

static bool
ccp_decl_initial_min_invariant (tree t)
{
  if (!is_gimple_min_invariant (t))
    return false;
  if (TREE_CODE (t) == ADDR_EXPR)
    {
      /* Inline and unroll is_gimple_addressable.  */
      while (1)
        {
          t = TREE_OPERAND (t, 0);
          if (is_gimple_id (t))
            return true;
          if (!handled_component_p (t))
            return false;
        }
    }
  return true;
}

static tree
get_symbol_constant_value (tree sym)
{
  if (TREE_STATIC (sym)
      && TREE_READONLY (sym)
      && !MTAG_P (sym))
    {
      tree val = DECL_INITIAL (sym);
      if (val
          && ccp_decl_initial_min_invariant (val))
        return val;
    }

  return NULL_TREE;
}

You could turn ccp_decl_initial_min_invariant into a global predicate and
invoke it from tree-ssa-sccvn.c on the DECL_INITIAL.

> However, even if i fix this, the testcase still fails because
> valid_gimple_expression says something that is clearly invalid is
> valid.
>
> (gdb) p valid_gimple_expression_p ($2)
> $3 = 1 '\001'
> (gdb) p debug_generic_stmt ($2)
> &((struct RegisterLayout *) (char *) &SimulatedRegisters)->intmask;
>
> This is not valid gimple by a longshot :)

Almost. :-)  The function was extracted unchanged from set_rhs but it looks
like it needs to be beefed up a bit if it is to become of general use.

> If you fix this part, i'll happily fix the bug report with the first part.

The problem is again the ADDR_EXPR case, because it drags lvalues into the
game.  This would be something like:

I'll try this and submit it as part of my patch if it works.

Thanks a ton!

Reply via email to