Re: [wwwdocs] porting_to: Two-stage overload resolution for implicit move removed

[Marek Polacek <polacek@redhat.com>]

On Wed, Oct 12, 2022 at 09:50:36PM +0100, Jonathan Wakely wrote:
> On Wed, 12 Oct 2022 at 20:39, Marek Polacek <polacek@redhat.com> wrote:
> >
> > As I promised in
> > <https://gcc.gnu.org/pipermail/gcc-patches/2022-October/603189.html>,
> > I'd like to update our GCC 13 porting_to.html with the following note.
> >
> > Does this look OK to commit?  Thanks,
> >
> > diff --git a/htdocs/gcc-13/porting_to.html b/htdocs/gcc-13/porting_to.html
> > index 84a00f21..243ed29d 100644
> > --- a/htdocs/gcc-13/porting_to.html
> > +++ b/htdocs/gcc-13/porting_to.html
> > @@ -42,5 +42,57 @@ be included explicitly when compiled with GCC 13:
> >  </li>
> >  </ul>
> >
> > +<h3 id="two-stage-or">Two-stage overload resolution for implicit move removed</h3>
> > +<p>
> > +GCC 13 removed the two-stage overload resolution when performing
> > +implicit move, whereby the compiler does two separate overload resolutions:
> > +one treating the operand as an rvalue, and then (if that resolution fails)
> > +another one treating the operand as an lvalue.  In the standard this was
> > +introduced in C++11 and implemented in gcc in
> > +<a href="https://gcc.gnu.org/git/?p=gcc.git;a=commitdiff;h=4ce8c5dea53d80736b9c0ba6faa7430ed65ed365">
> > +r251035</a>.  In
> > +<a href="https://gcc.gnu.org/git/?p=gcc.git;a=commitdiff;h=1722e2013f05f1f1f99379dbaa0c0df356da731f">
> > +r11-2412</a>, the fallback overload resolution was disabled in C++20 (but
> > +not in C++17).  Then C++23 <a href="https://wg21.link/p2266">P2266</a>
> > +removed the fallback overload resolution, and changed the implicit move
> > +rules once again.
> > +</p>
> > +<p>
> > +The two overload resolutions approach was complicated and quirky, so users
> > +should transition to the newer model.  This change means that code that
> > +previously didn't compile in C++17 will now compile, for example:</p>
> > +
> > +<pre><code>
> > +   struct S1 { S1(S1 &&); };
> > +   struct S2 : S1 {};
> > +
> > +   S1
> > +   f (S2 s)
> > +   {
> > +     return s; // OK, derived-to-base, use S1::S1(S1&&)
> > +   }
> > +</code></pre>
> > +
> > +<p>
> > +And conversely, code that used to work in C++17 may not compile anymore:
> > +</p>
> > +
> > +<pre><code>
> > +   struct W {
> > +     W();
> > +   };
> > +
> > +   struct F {
> > +     F(W&);
> > +     F(W&&) = delete;
> > +   };
> > +
> > +   F fn ()
> > +   {
> > +     W w;
> > +     return w; // use w as rvalue -> use of deleted function F::F(W&&)
> 
> Deleted move constructors are an abomination, and should never occur
> in real code. I'm not sure using one even in an example like this
> should be encouraged. The example added by P2266 to Annex D is more
> realistic (and actually broke a libstdc++ test):
> 
> X& foo(X&& x) { return x; }

Right, but this code still compiles in C++17, it only fails to compile
in C++23.  The previous example now doesn't compile even in C++17.  So
how about this improved patch which makes it clear that code with
deleted move constructors should never occur in practice, and adds a new
note, specifically about P2266 and the code you showed?

Thanks for taking a look,

diff --git a/htdocs/gcc-13/porting_to.html b/htdocs/gcc-13/porting_to.html
index 84a00f21..a9991e8b 100644
--- a/htdocs/gcc-13/porting_to.html
+++ b/htdocs/gcc-13/porting_to.html
@@ -42,5 +42,71 @@ be included explicitly when compiled with GCC 13:
 </li>
 </ul>
 
+<h3 id="P2266">Implicit move rules change</h3>
+<p>
+GCC 13 implements C++23 <a href="https://wg21.link/p2266">P2266</a> which
+simplified the rules for implicit move.  As a consequence, valid C++20
+code that relies on a returned <em>id-expression</em>'s being an lvalue
+may change behavior or fail to compile in C++23.  For example:</p>
+
+<pre><code>
+   decltype(auto) f(int&& x) { return (x); }  // returns int&&; previously returned int&
+   int& g(int&& x) { return x; }  // ill-formed; previously well-formed
+</code></pre>
+
+<h3 id="two-stage-or">Two-stage overload resolution for implicit move removed</h3>
+<p>GCC 13 removed the two-stage overload resolution when performing
+implicit move, whereby the compiler does two separate overload resolutions:
+one treating the operand as an rvalue, and then (if that resolution fails)
+another one treating the operand as an lvalue.  In the standard this was
+introduced in C++11 and implemented in gcc in
+<a href="https://gcc.gnu.org/git/?p=gcc.git;a=commitdiff;h=4ce8c5dea53d80736b9c0ba6faa7430ed65ed365">
+r251035</a>.  In
+<a href="https://gcc.gnu.org/git/?p=gcc.git;a=commitdiff;h=1722e2013f05f1f1f99379dbaa0c0df356da731f">
+r11-2412</a>, the fallback overload resolution was disabled in C++20 (but
+not in C++17).  Then C++23 <a href="https://wg21.link/p2266">P2266</a>
+removed the fallback overload resolution, and changed the implicit move
+rules once again.</p>
+
+<p>The two overload resolutions approach was complicated and quirky, so users
+should transition to the newer model.  This change means that code that
+previously didn't compile in C++17 will now compile, for example:</p>
+
+<pre><code>
+   struct S1 { S1(S1 &&); };
+   struct S2 : S1 {};
+
+   S1
+   f (S2 s)
+   {
+     return s; // OK, derived-to-base, use S1::S1(S1&&)
+   }
+</code></pre>
+
+<p>Conversely, code that used to work in C++17 may not compile anymore.
+For example, the following example used to compile in C++11...17 because
+we performed two separate overload resolutions: one treating the operand
+as an rvalue, and then (if that resolution failed) another one treating
+the operand as an lvalue.<br>
+<strong>NB:</strong> this example is contrived because deleted move
+constructors should not occur in real code.</p>
+
+<pre><code>
+   struct W {
+     W();
+   };
+
+   struct F {
+     F(W&);
+     F(W&&) = delete;
+   };
+
+   F fn ()
+   {
+     W w;
+     return w; // use w as rvalue -> use of deleted function F::F(W&&)
+   }
+</code></pre>
+
 </body>
 </html>