[Bug tree-optimization/34448] [4.3 Regression] ICE in declare_inline_vars, at tree-inline.c:3243

[jakub at gcc dot gnu dot org]

------- Comment #5 from jakub at gcc dot gnu dot org  2007-12-13 00:08 -------
ret = GS_OK; } ret = GS_UNHANDLED; looks suspicious, but
@@ -3492,6 +3492,7 @@ gimplify_modify_expr_rhs (tree *expr_p, 
          {
            *from_p = DECL_INITIAL (*from_p);
            ret = GS_OK;
+           break;
          }
        ret = GS_UNHANDLED;
        break;
doesn't cure this.  But if you prevent this optimization whenever
gimplify_init_constructor would:
        if (valid_const_initializer
            && num_nonzero_elements > 1
            && TREE_READONLY (object)
            && TREE_CODE (object) == VAR_DECL)
or
if (size > 0 && !can_move_by_pieces (size, align))
then this bug will go away.

BTW, in addition to the volatile check suggested by Mark, shouldn't it test
also
!TREE_SIDE_EFFECTS (DECL_INITIAL (...)))?  Otherwise:
struct A { int i, j; };
static struct A a, b;
int foo (int x, int y)
{
  const struct A i = { x++, y++ };
  a = i;
  b = i;
  return x + y;
}
could evaluate multiple times.  Actually this is a wrong example, because i
here
was already gimplified and while for constant initializer
gimplify_init_constructor puts DECL_INITIAL back (valid_const_initializer &&
&& num_nonzero_elements > 1 && TREE_READONLY (object) && TREE_CODE (object) ==
VAR_DECL case), for non-constant DECL_INITIAL will be NULL.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=34448