Re: [committed] libstdc++: Allow Lemire's algorithm to be used in more cases

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

On Wed, 4 Nov 2020 at 10:46, Stephan Bergmann via Libstdc++
<libstdc++@gcc.gnu.org> wrote:
> To me it looks like it boils down to disagreement between g++ and
> clang++ over
>
> > struct S { static constexpr int f() { return 0; } };
> > void f(S & s) { static_assert(s.f(), ""); }
>
> where I think Clang might be right in rejecting it based on [expr.const]
> "An expression e is a core constant expression unless [...] an
> id-expression that refers to a variable or data member of reference type
> unless the reference has a preceding initialization [...]"

There's more to it than that. It's a disagreement over [expr.ref]/1.
For a static
member call, gcc just plain doesn't evaluate the s in s.f(). But [expr.ref]/1
says it's evaluated, and since it's not a constant expression, clang rejects
it, and gcc accepts it. That's why your fix works; it removes the use
of the otherwise-mostly-ignored
object expression for a call to a static member function.

So, I think gcc is accepting-invalid here, and we should just apply
the fix you suggested.