Re: [PATCH v5] c++: implement [[gnu::non_owning]] [PR110358]

[Marek Polacek <polacek@redhat.com>]

On Mon, Mar 04, 2024 at 11:00:18AM +0000, Jonathan Wakely wrote:
> On 01/03/24 15:38 -0500, Jason Merrill wrote:
> > On 3/1/24 14:24, Marek Polacek wrote:
> > > +@smallexample
> > > +template <typename T>
> > > +[[gnu::no_dangling(std::is_reference_v<T>)]] int foo (T& t) @{
> > 
> > I think this function should return a reference.
> 
> The condition in the attribute can only ever be true if you call this
> function with an explicit template argument list: foo<int&>(i). Is
> that intentional?

Not intentional.  I just wanted to make it clear that the user
can use something like std::is_reference as the attribute argument,
but I didn't think about it very long.
 
> And if T is non-const it can't be called with a temporary and so
> dangling seems less of a problem for this function anyway, right?

Right.
 
> Would it make more sense as something like this?
> 
> template <typename T>
> [[gnu::no_dangling(std::is_lvalue_reference_v<T>)]]
> decltype(auto) foo(T&& t) {
>   ...
> }
> 
> Or is this getting too complex/subtle for a simple example?

I like your example; it's only slightly more complex than the
original one and most likely more realistic.  I'm pushing the
following patch.  Thanks!

[pushed] doc: update [[gnu::no_dangling]]

...to offer a more realistic example.

gcc/ChangeLog:

	* doc/extend.texi: Update [[gnu::no_dangling]].
---
 gcc/doc/extend.texi | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/gcc/doc/extend.texi b/gcc/doc/extend.texi
index f679c81acf2..df0982fdfda 100644
--- a/gcc/doc/extend.texi
+++ b/gcc/doc/extend.texi
@@ -29370,7 +29370,8 @@ Or:
 
 @smallexample
 template <typename T>
-[[gnu::no_dangling(std::is_reference_v<T>)]] int& foo (T& t) @{
+[[gnu::no_dangling(std::is_lvalue_reference_v<T>)]]
+decltype(auto) foo(T&& t) @{
   @dots{}
 @};
 @end smallexample

base-commit: 77eb86be8841989651b3150a020dd1a95910cc00
-- 
2.44.0