Re: [PATCH] Pass correct type to irange::contains_p() in ipa-cp.cc.

[Richard Biener <richard.guenther@gmail.com>]

On Mon, Apr 24, 2023 at 2:32 PM Aldy Hernandez <aldyh@redhat.com> wrote:
>
>
>
> On 4/24/23 14:10, Richard Biener wrote:
> > On Mon, Apr 24, 2023 at 1:51 PM Aldy Hernandez <aldyh@redhat.com> wrote:
> >>
> >>
> >>
> >> On 4/24/23 10:30, Richard Biener wrote:
> >>> On Mon, Apr 24, 2023 at 9:44 AM Aldy Hernandez via Gcc-patches
> >>> <gcc-patches@gcc.gnu.org> wrote:
> >>>>
> >>>> There is a call to contains_p() in ipa-cp.cc which passes incompatible
> >>>> types.  This currently works because deep in the call chain, the legacy
> >>>> code uses tree_int_cst_lt which performs the operation with
> >>>> widest_int.  With the upcoming removal of legacy, contains_p() will be
> >>>> stricter.
> >>>>
> >>>> OK pending tests?
> >>>>
> >>>> gcc/ChangeLog:
> >>>>
> >>>>           * ipa-cp.cc (ipa_range_contains_p): New.
> >>>>           (decide_whether_version_node): Use it.
> >>>> ---
> >>>>    gcc/ipa-cp.cc | 16 +++++++++++++++-
> >>>>    1 file changed, 15 insertions(+), 1 deletion(-)
> >>>>
> >>>> diff --git a/gcc/ipa-cp.cc b/gcc/ipa-cp.cc
> >>>> index b3e0f62e400..c8013563796 100644
> >>>> --- a/gcc/ipa-cp.cc
> >>>> +++ b/gcc/ipa-cp.cc
> >>>> @@ -6180,6 +6180,19 @@ decide_about_value (struct cgraph_node *node, int index, HOST_WIDE_INT offset,
> >>>>      return true;
> >>>>    }
> >>>>
> >>>> +/* Like irange::contains_p(), but convert VAL to the range of R if
> >>>> +   necessary.  */
> >>>> +
> >>>> +static inline bool
> >>>> +ipa_range_contains_p (const irange &r, tree val)
> >>>> +{
> >>>> +  if (r.undefined_p ())
> >>>> +    return false;
> >>>> +
> >>>> +  val = fold_convert (r.type (), val);
> >>>
> >>> I think that's wrong, it might truncate 'val'.  I think we'd want
> >>>
> >>>      if (r.undefined_p () || !int_fits_type_p (val, r.type ()))
> >>>        return false;
> >>
> >> This won't work for pointers.  Is there a suitable version that handles
> >> pointers as well?
> >
> > Where does it not work?  And when do you get pointer values/types
> > where they mismatch sufficiently (how?) to make ranger unhappy?
> >
>
> It's not ranger that's unhappy, this is just irange::contains_p().
>
> IPA works on both integers and pointers, and pointers don't have
> TYPE_MIN/MAX_VALUE defined, so we ICE in int_fits_type_p:
>
>    type_low_bound = TYPE_MIN_VALUE (type);
>    type_high_bound = TYPE_MAX_VALUE (type);

Ah.  The code handles NULL TYPE_MIN/MAX_VALUE just fine so
I wonder if it works to change the above to

   type_low_bound = INTEGRAL_TYPE_P (type) ? TYPE_MIN_VALUE (type) : NULL_TREE;
...

alternatively you could use wi::fits_to_tree_p (wi::to_wide (val),
type) from the IPA code.

Richard.