[libcxx-commits] [PATCH] D128281: [libc++] fix views::all hard error on lvalue move only views instead of SFINAE

Eric Fiselier via Phabricator via libcxx-commits libcxx-commits at lists.llvm.org
Tue Jun 21 13:44:49 PDT 2022

EricWF added inline comments.

Comment at: libcxx/include/__ranges/all.h:42
+      -> decltype(_LIBCPP_AUTO_CAST(std::forward<_Tp>(__t)))
huixie90 wrote:
> EricWF wrote:
> > var-const wrote:
> > > I wonder if we have any other cases where we don't write the return type 3 times and consequently don't get SFINAE.
> > Is this a case of "not getting SFINAE" or "auto" return types not deducing to reference types? Does this work if we use `decltype(auto)` instead?
> > 
> > Also this changes the return type from being by-value to being by-reference, which is the intended change I assume?
> This function always return by value because `_LIBCPP_AUTO_CAST` always yield the decayed type
> ```
> #define _LIBCPP_AUTO_CAST(expr) static_cast<typename decay<decltype((expr))>::type>(expr)
> ```
> In the spec, it says expression equivalent to `decay-copy(E)`, and  `_LIBCPP_AUTO_CAST` is the libc++ way of implementing `decay-copy`
> In any case, the return type should be the same as `decay-copy` (expression equivalent), and this is why `decltype(_LIBCPP_AUTO_CAST(...))` is correct
> `decltype(auto)` won't work because `decltype(auto)` doesn't have any effect on SFINAE
Oh, based on the name I had assumed it did something entirely different. Like forward<T&&>(t), but without the need to write the type.
Maybe it needs a better name.

Also, this is a lesson in using `auto` return types. I think it's clear that:

1. We shouldn't use them in any place the standard doesn't mandate.
1.  We really shouldn't use them in any place we also use `noexcept(noexcept(...))`, because it can poison an entire overload set as your discovering.

  rG LLVM Github Monorepo



More information about the libcxx-commits mailing list