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

[Ville Voutilainen <ville.voutilainen@gmail.com>]

On Mon, 14 Jan 2019 at 10:29, Richard Biener <richard.guenther@gmail.com> wrote:
> >        _GLIBCXX14_CONSTEXPR bool
> >        operator()(_Tp* __x, _Tp* __y) const _GLIBCXX_NOTHROW
> >        {
> > +#if __cplusplus >= 201402L
> > +#ifdef _GLIBCXX_HAVE_BUILTIN_IS_CONSTANT_EVALUATED
> > +       if (__builtin_is_constant_evaluated())
> > +#else
> >         if (__builtin_constant_p (__x > __y))
> > +#endif
> >           return __x > __y;
> > +#endif
> >         return (__UINTPTR_TYPE__)__x > (__UINTPTR_TYPE__)__y;
>
> I wonder what the idea behind this is.  It smells like trying to avoid
> undefined behavior (relational compare of pointers to different objects?)
> but then executing that nevertheless when "constant"?
>
> I think this just doesn't work since the compiler, when evaluating
> __x > __y [for constant folding] is exploiting the fact that doing
> non-equality compares on pointers into different objects invoke
> undefined behavior.

When that happens, the function is ill-formed when constant-evaluated,
which is fine.
When the comparison is not UB, it should constant-evaluate without problems.

> So why is this not just
>   return (__UINTPTR_TYPE__)__x > (__UINTPTR_TYPE__)__y;
> or with the casts elided?

Those casts are reinterpret_casts, so the function could never be
constant-evaluated.
The casts need to be there to avoid UB for the run-time cases.