This is the mail archive of the libstdc++@gcc.gnu.org mailing list for the libstdc++ project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: [PATCH] Use __builtin_is_constant_evaluated in std::less etc. (PR tree-optimization/88775)


On Mon, 14 Jan 2019 at 09:17, Richard Biener <richard.guenther@gmail.com> wrote:
>
> On Mon, Jan 14, 2019 at 9:42 AM Jakub Jelinek <jakub@redhat.com> wrote:
> >
> > On Mon, Jan 14, 2019 at 09:29:03AM +0100, Richard Biener wrote:
> > > So why is this not just
> > >
> > >   return (__UINTPTR_TYPE__)__x > (__UINTPTR_TYPE__)__y;
> > >
> > > or with the casts elided?  Does the C++ standard say pointers are
> > > to be compared unsigned here?  Or do all targets GCC support
> > > lay out the address space in a way that this is correct for pointers
> > > into distinct objects?
> >
> > See PR78420 for details on why it is done that way.
>
> I see.  So the __builtin_is_constant_evaluated thing makes it
> "correct" (but then eventually exposing the non-total order issue again).

No, because comparing unrelated pointers isn't allowed in constexpr
contexts, so it just gets rejected at compile time.


> And if I read the PR correctly we'd really like to be able to write
>
>  if (__builtin_constant_p (<expr>, &result))
>    return result;
>
> to make sure whatever undesired-in-the-IL things of <expr>
> do not leak there.
>
> Btw, wouldn't sth like
>
>   if (__builtin_is_constant_evaluated())
>     {
>        union U { __UINTPTR_TYPE__ u; _Tp *p } __ux, __uy;
>        __ux.p = __x;
>        __uy.p = __y;
>        return __ux.u < __uy.u;
>     }
>
> be more correct and consistent?  Well, or any other way of
> evading that reinterpret-cast "issue"?
>
> Richard.
>
>
> >
> >         Jakub


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]