[Bug libstdc++/97876] New: stop_token header doesn't compile on clang-8 with -std=c++20

[Bug libstdc++/97876] New: stop_token header doesn't compile on clang-8 with -std=c++20

[eric.niebler at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97876

            Bug ID: 97876
           Summary: stop_token header doesn't compile on clang-8 with
                    -std=c++20
           Product: gcc
           Version: unknown
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: libstdc++
          Assignee: unassigned at gcc dot gnu.org
          Reporter: eric.niebler at gmail dot com
  Target Milestone: ---

Compiling <stop_token> header with clang-8 in C++20 mode results in this:

In file included from
/usr/bin/../lib/gcc/x86_64-linux-gnu/10/../../../../include/c++/10/condition_variable:50:
/usr/bin/../lib/gcc/x86_64-linux-gnu/10/../../../../include/c++/10/stop_token:404:7:
error: default member initializer for '_M_ptr' needed within definition of
enclosing class 'stop_token' outside of member functions
      _Stop_state_ref() = default;
      ^
/usr/bin/../lib/gcc/x86_64-linux-gnu/10/../../../../include/c++/10/stop_token:478:21:
note: in evaluation of exception specification for
'std::stop_token::_Stop_state_ref::_Stop_state_ref' needed here
    _Stop_state_ref _M_state;
                    ^
/usr/bin/../lib/gcc/x86_64-linux-gnu/10/../../../../include/c++/10/stop_token:58:5:
note: in evaluation of exception specification for
'std::stop_token::stop_token' needed here
    stop_token() noexcept = default;
    ^
/usr/bin/../lib/gcc/x86_64-linux-gnu/10/../../../../include/c++/10/stop_token:475:22:
note: default member initializer declared here
      _Stop_state_t* _M_ptr = nullptr;
                     ^
1 error generated.

[Bug libstdc++/97876] stop_token header doesn't compile on clang-8 with -std=c++20

[redi at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97876

--- Comment #1 from Jonathan Wakely <redi at gcc dot gnu.org> ---
It works fine in Clang 9.0.0 though. We've never really set a hard limit on how
many old versions of Clang we support, but given that 9 and 10 (and trunk)
work, I'm inclined to think that three versions seems enough for new C++20
features.

[Bug libstdc++/97876] stop_token header doesn't compile on clang-8 with -std=c++20

[redi at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97876

--- Comment #2 from Jonathan Wakely <redi at gcc dot gnu.org> ---
Changing the _Stop_state_ref default constructor to:

    _Stop_state_ref() noexcept {}

allows it to compile. It shouldn't be needed, but I suppose since the
constructor is already non-trivial (due to the default member initializer that
is the cause of the error) we could do that.

[Bug libstdc++/97876] stop_token header doesn't compile on clang-8 with -std=c++20

[eric.niebler at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97876

--- Comment #3 from Eric Niebler <eric.niebler at gmail dot com> ---
It seems like GitHub actions uses the latest libstdc++ by default when testing
even with old (e.g., clang-4) toolsets. That seems busted, but regardless, this
is now breaking _all_ my Linux clang tests for anything less than clang-9,
which is unfortunate.

[Bug libstdc++/97876] stop_token header doesn't compile on clang-8 with -std=c++20

[redi at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97876

--- Comment #4 from Jonathan Wakely <redi at gcc dot gnu.org> ---
Github's poor life choices should not be our problem ;-)

If clang-8 doesn't ship <stop_token> and doesn't work with GCC's <stop_token>,
I would interpret that as you can't test <stop_token> with clang-8.

[Bug libstdc++/97876] stop_token header doesn't compile on clang-8 with -std=c++20

[eric.niebler at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97876

--- Comment #5 from Eric Niebler <eric.niebler at gmail dot com> ---
> Github's poor life choices should not be our problem ;-)

I thought you'd say that. :-)

> If clang-8 doesn't ship <stop_token> and doesn't work with GCC's <stop_token>, I would interpret that as you can't test <stop_token> with clang-8.

Yeah, except I'm not trying to test <stop_token>. I'm just trying to include
<condition_variable>, which includes <stop_token> and hard-errors. In fact,
what I'm _really_ trying to do is include <regex>, which includes
<memory_resource>, which includes <shared_mutex>, which includes
<condition_variable>. All roads, it seems, lead to <stop_token>.

[Bug libstdc++/97876] stop_token header doesn't compile on clang-8 with -std=c++20

[redi at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97876

Jonathan Wakely <redi at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
     Ever confirmed|0                           |1
   Last reconfirmed|                            |2020-11-19

--- Comment #6 from Jonathan Wakely <redi at gcc dot gnu.org> ---
Ugh, OK, that's more problematic then.

[Bug libstdc++/97876] stop_token header doesn't compile on clang-8 with -std=c++20

[redi at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97876

--- Comment #7 from Jonathan Wakely <redi at gcc dot gnu.org> ---
g:f5d9bc8ae81abe46640477bc9e655aa093947f5f removed the dependency from
<shared_mutex> on <condition_variable> for most targets.

[Bug libstdc++/97876] stop_token header doesn't compile on clang-8 with -std=c++20

[redi at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97876

Jonathan Wakely <redi at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|NEW                         |ASSIGNED
   Target Milestone|---                         |10.3
           Assignee|unassigned at gcc dot gnu.org      |redi at gcc dot gnu.org

--- Comment #8 from Jonathan Wakely <redi at gcc dot gnu.org> ---
Backporting that to gcc-10 would break any code implicitly relying on
<condition_variable> being included by <regex>. I generally only do such header
dependency changes for major releases. Maybe it's OK, as it's only included for
C++17 mode, which isn't considered stable for gcc-10.

Separately, <regex> shouldn't need <memory_resource> when it already declares
polymorphic allocator (which is sufficient for <string> and <vector> and the
other headers with alias templates in std::pmr). It needs it because
pmr::string::const_iterator and pmr::wstring::const_iterator require string and
wstring to be complete, which requires polymorphic_allocator to be complete.
That's avoidable though:

--- a/libstdc++-v3/include/std/regex
+++ b/libstdc++-v3/include/std/regex
@@ -64,7 +64,6 @@
 #include <bits/regex_executor.h>

 #if __cplusplus >= 201703L && _GLIBCXX_USE_CXX11_ABI
-#include <memory_resource>
 namespace std _GLIBCXX_VISIBILITY(default)
 {
 _GLIBCXX_BEGIN_NAMESPACE_VERSION
@@ -74,11 +73,13 @@ _GLIBCXX_BEGIN_NAMESPACE_VERSION
       using match_results
        = std::match_results<_BidirectionalIterator, polymorphic_allocator<
                                sub_match<_BidirectionalIterator>>>;
-    using cmatch  = match_results<const char*>;
-    using smatch  = match_results<string::const_iterator>;
+    using cmatch = match_results<const char*>;
+    using smatch
+      = match_results<__gnu_cxx::__normal_iterator<const char*, string>>;
 #ifdef _GLIBCXX_USE_WCHAR_T
     using wcmatch = match_results<const wchar_t*>;
-    using wsmatch = match_results<wstring::const_iterator>;
+    using wsmatch
+      = match_results<__gnu_cxx::__normal_iterator<const wchar_t*, wstring>>;
 #endif
   } // namespace pmr
 _GLIBCXX_END_NAMESPACE_VERSION


I think removing that dependency in the gcc-10 branch should be OK too.

[Bug libstdc++/97876] stop_token header doesn't compile on clang-8 with -std=c++20

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97876

--- Comment #9 from CVS Commits <cvs-commit at gcc dot gnu.org> ---
The master branch has been updated by Jonathan Wakely <redi@gcc.gnu.org>:

https://gcc.gnu.org/g:640ebeb336050887cb57417b7568279c588088f0

commit r11-5199-g640ebeb336050887cb57417b7568279c588088f0
Author: Jonathan Wakely <jwakely@redhat.com>
Date:   Fri Nov 20 11:30:33 2020 +0000

    libstdc++: Remove <memory_resource> dependency from <regex> [PR 92546]

    Unlike the other headers that declare alias templates in namespace pmr,
    <regex> includes <memory_resource>. That was done because the
    pmr::string::const_iterator typedef requires pmr::string to be complete,
    which requires pmr::polymorphic_allocator<char> to be complete.

    By using __normal_iterator<const char*, pmr::string> instead of the
    const_iterator typedef we can avoid the completeness requirement.

    This makes <regex> smaller, by not requiring <memory_resource> and its
    <shared_mutex> dependency, which depends on <chrono>.  Backporting this
    will also help with PR 97876, where <stop_token> ends up being needed by
    <regex> via <memory_resource>.

    libstdc++-v3/ChangeLog:

            PR libstdc++/92546
            * include/std/regex (pmr::smatch, pmr::wsmatch): Declare using
            underlying __normal_iterator type, not nested typedef
            basic_string::const_iterator.

[Bug libstdc++/97876] stop_token header doesn't compile on clang-8 with -std=c++20

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97876

--- Comment #10 from CVS Commits <cvs-commit at gcc dot gnu.org> ---
The releases/gcc-10 branch has been updated by Jonathan Wakely
<redi@gcc.gnu.org>:

https://gcc.gnu.org/g:a186d72afd6cfb13efd4a0ec82049d79892334fd

commit r10-9056-ga186d72afd6cfb13efd4a0ec82049d79892334fd
Author: Jonathan Wakely <jwakely@redhat.com>
Date:   Thu Nov 19 22:32:54 2020 +0000

    libstdc++: Fix compilation error with clang-8 [PR 97876]

    This fixes a compilation error with clang-8 and earlier. This change is
    only on the gcc-10 branch, not master, because the <stop_token> header
    is included indirectly in more places on the branch than on master.

            PR libstdc++/97876
            * include/std/stop_token (_Stop_state_t): Define default
            constructor as user-provided not defaulted.

[Bug libstdc++/97876] stop_token header doesn't compile on clang-8 with -std=c++20

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97876

--- Comment #11 from CVS Commits <cvs-commit at gcc dot gnu.org> ---
The releases/gcc-10 branch has been updated by Jonathan Wakely
<redi@gcc.gnu.org>:

https://gcc.gnu.org/g:cbbc28706164873d0323d1a6c7988be3f4d971c9

commit r10-9057-gcbbc28706164873d0323d1a6c7988be3f4d971c9
Author: Jonathan Wakely <jwakely@redhat.com>
Date:   Fri Nov 20 11:30:33 2020 +0000

    libstdc++: Remove <memory_resource> dependency from <regex> [PR 92546]

    Unlike the other headers that declare alias templates in namespace pmr,
    <regex> includes <memory_resource>. That was done because the
    pmr::string::const_iterator typedef requires pmr::string to be complete,
    which requires pmr::polymorphic_allocator<char> to be complete.

    By using __normal_iterator<const char*, pmr::string> instead of the
    const_iterator typedef we can avoid the completeness requirement.

    This makes <regex> smaller, by not requiring <memory_resource> and its
    <shared_mutex> dependency, which depends on <chrono>.  Backporting this
    will also help with PR 97876, where <stop_token> ends up being needed by
    <regex> via <memory_resource>.

    libstdc++-v3/ChangeLog:

            PR libstdc++/92546
            * include/std/regex (pmr::smatch, pmr::wsmatch): Declare using
            underlying __normal_iterator type, not nested typedef
            basic_string::const_iterator.

    (cherry picked from commit 640ebeb336050887cb57417b7568279c588088f0)

[Bug libstdc++/97876] stop_token header doesn't compile on clang-8 with -std=c++20

[redi at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97876

Jonathan Wakely <redi at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|ASSIGNED                    |RESOLVED
         Resolution|---                         |FIXED

--- Comment #12 from Jonathan Wakely <redi at gcc dot gnu.org> ---
(In reply to Jonathan Wakely from comment #2)
> Changing the _Stop_state_ref default constructor to:
> 
>     _Stop_state_ref() noexcept {}
> 
> allows it to compile. It shouldn't be needed, but I suppose since the
> constructor is already non-trivial (due to the default member initializer
> that is the cause of the error) we could do that.

I've done this on the gcc-10 branch, but not trunk. That means the same problem
will recur with clang-8 and <stop_token> in gcc-11.

However, I've also removed the <memory_resource> dependency from <regex> on
gcc-10 and trunk, and <shared_mutex> already stopped including <stop_token> on
trunk. So there are fewer routes to <stop_token> now.

Now you just need to wait for Ubuntu to update gcc-10 and for GitHub to deploy
that update, which I can't help with.