------- Comment #7 from dberlin at gcc dot gnu dot org 2007-07-07 20:07 -------
Subject: Re: [4.3 regression]: revision 126369 went into an infinite loop
On 7 Jul 2007 19:35:01 -0000, hjl at lucon dot org
<[EMAIL PROTECTED]> wrote:
>
>
> ------- Comment #6 from hjl at lucon dot org 2007-07-07 19:35 -------
> If I revert
>
> - if (lhsval)
> + if (lhsval && vuse_equiv (lhsval, stmt))
>
> the regression is gone. I suspected that the original code:
>
> if (lhsval)
> {
> set_value_handle (newt, lhsval);
> if (!is_gimple_min_invariant (lhsval))
> add_to_value (lhsval, newt);
> }
> else
> {
> tree val = vn_lookup_or_add_with_stmt (newt, stmt);
> vn_add (lhs, val);
> }
>
> doesnt't handle the case of
>
> if (lhsval && !vuse_equiv (lhsval, stmt))
>
Uh, as it shouldn't.
> If I change it to
>
> --- tree-ssa-pre.c.bad 2007-07-07 08:18:31.000000000 -0700
> +++ tree-ssa-pre.c 2007-07-07 12:33:51.000000000 -0700
> @@ -3362,7 +3362,8 @@ make_values_for_stmt (tree stmt, basic_b
> else
> {
> tree val = vn_lookup_or_add_with_stmt (newt, stmt);
> - vn_add (lhs, val);
> + if (!lhsval)
> + vn_add (lhs, val);
> }
>
> add_to_exp_gen (block, newt);
>
> my testcase works.
Only because you are now assigning different value numbers to things
we know to be the same.
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=32663
