[PATCH] D40259: [libcxx] LWG2993: reference_wrapper<T> conversion from T&&

[Eric Fiselier via Phabricator via cfe-commits]

EricWF added inline comments.


================
Comment at: include/__functional_base:396
+        !is_same<__uncvref_t<_Up>, reference_wrapper>::value
+    >::type, bool _IsNothrow = noexcept(__bind(_VSTD::declval<_Up>()))>
+    _LIBCPP_INLINE_VISIBILITY reference_wrapper(_Up&& __u) _NOEXCEPT_(_IsNothrow)
----------------
tcanens wrote:
> EricWF wrote:
> > tcanens wrote:
> > > Is it safe to do this when we are using `_NOEXCEPT_` in the next line?
> > It should be. The noexcept condition should only be evaluated *as needed* for functions selected by overload resolution. i.e. The noexcept condition is only considered on well-formed functions. And this function is only well-formed if `_IsNothrow` is well formed.
> > 
> > If `IsNothrow` is ill-formed, it will prevent the functions noexcept specifier from ever being evaluated.
> My point is that the use of macro-ized `_NOEXCEPT_` suggests that we are supporting compilers that doesn't have noexcept-specifications. Given that, can we then use the `noexcept` operator here?
Ah, I see!

The `_NOEXCEPT_` macro is a holdover from the C++03/C++11 era. This code expects a C++11 conformant compiler, and hence expects `noexcept` to work as a non-buggy feature-complete C++11 construct.

Much in the same way the current (and previous) code expects C++11 template aliases, default function template arguments, `decltype`, ect. to work.


https://reviews.llvm.org/D40259