On Fri, Jan 24, 2014 at 09:35:10AM -0700, Jeff Law wrote:
> On 01/24/14 03:52, Jakub Jelinek wrote:
> >On Thu, Jan 23, 2014 at 02:22:35PM -0700, Jeff Law wrote:
> >>--- a/gcc/tree-vrp.c
> >>+++ b/gcc/tree-vrp.c
> >>@@ -5891,8 +5891,13 @@ find_assert_locations_1 (basic_block bb, sbitmap
> >>live)
> >> }
> >> }
> >>
> >>- register_new_assert_for (op, op, comp_code, value, bb, NULL, si);
> >>- need_assert = true;
> >>+ /* Do not register any assertions for a non-returning call. */
> >>+ if (!is_gimple_call (stmt) || !gimple_call_noreturn_p (stmt))
> >>+ {
> >>+ register_new_assert_for (op, op, comp_code,
> >>+ value, bb, NULL, si);
> >>+ need_assert = true;
> >>+ }
> >> }
> >> }
> >
> >I'd say this belongs into infer_value_range instead.
> Seems reasonable. I'll spin that.
>
> The only downside is we lose the ability to backward propagate
> through a typecast which feeds an argument in such a call. But
> that's probably not a big deal.
I couldn't actually reproduce any backwards propagation,
since now all pointer casts are useless, so the code to walk the pointer
casts is likely just dead.
Jakub
