v1.11.0 fails to build on F44

[junghans]

From https://bugzilla.redhat.com/show_bug.cgi?id=2434647:

template argument deduction/substitution failed:
                • error: type/value mismatch at argument 1 in template parameter list for ‘template<class _Tp, class ... _Types> constexpr _Tp&& std::get(tuple<_Elements ...>&&)’

See build.log.zip

[hkaiser]

That's a new issue related to gcc 16. I'll investigate.

My hunch is that gcc 16 breaks our workaround for making zip_iterator appear to be truely random access when being used with sort (zip_iterator doesn't return a true reference as formally required by the standard, but it returns a tuple of true references).

Now, the sort algorithm swaps elements from the sequence to be sorted. The sort_by_key algorithm invokes sort on two zipped sequences, causing sort to attempt to swap two tuple<T1&, T2&> instances, which we explicitly allow by providing a corresponding overload here: https://github.com/STEllAR-GROUP/hpx/blob/1cce0aeb6c95a670c600bf6c8d35472ede5ce0a0/libs/core/datastructures/include/hpx/datastructures/tuple.hpp#L1127-L1143.

[hkaiser]

A closer look reveals that the the operator[]() implemented by our zip_iterator didn't properly interact with with the changed implementation of std::sort in gcc's new standard library.

[hkaiser]

@junghans Would you be able to test applying this diff to see if that fixes the issue? I don't have access to gcc 16 right now.

diff --git a/libs/core/iterator_support/include/hpx/iterator_support/zip_iterator.hpp b/libs/core/iterator_support/include/hpx/iterator_support/zip_iterator.hpp
index 2e68520635b..3c06cd79145 100644
--- a/libs/core/iterator_support/include/hpx/iterator_support/zip_iterator.hpp
+++ b/libs/core/iterator_support/include/hpx/iterator_support/zip_iterator.hpp
@@ -361,10 +373,10 @@ namespace hpx::util {
             detail::zip_iterator_base<hpx::tuple<Ts...>, zip_iterator<Ts...>>;
 
     public:
-        HPX_HOST_DEVICE constexpr zip_iterator() noexcept
-          : base_type()
-        {
-        }
+        // prevent the operator_brackets_proxy from being used
+        using use_brackets_proxy = std::false_type;
+
+        HPX_HOST_DEVICE zip_iterator() = default;
 
         HPX_HOST_DEVICE explicit constexpr zip_iterator(
             Ts const&... vs) noexcept
@@ -442,10 +456,10 @@ namespace hpx::util {
             zip_iterator<hpx::tuple<Ts...>>>;
 
     public:
-        HPX_HOST_DEVICE constexpr zip_iterator() noexcept
-          : base_type()
-        {
-        }
+        // prevent the operator_brackets_proxy from being used
+        using use_brackets_proxy = std::false_type;
+
+        HPX_HOST_DEVICE zip_iterator() = default;
 
         HPX_HOST_DEVICE explicit constexpr zip_iterator(
             Ts const&... vs) noexcept
diff --git a/libs/core/iterator_support/include/hpx/iterator_support/iterator_facade.hpp b/libs/core/iterator_support/include/hpx/iterator_support/iterator_facade.hpp
index 94bf8fb36ba..2f791997c36 100644
--- a/libs/core/iterator_support/include/hpx/iterator_support/iterator_facade.hpp
+++ b/libs/core/iterator_support/include/hpx/iterator_support/iterator_facade.hpp
@@ -307,8 +308,8 @@ namespace hpx::util {
         }
 
         template <typename Iterator>
-        HPX_HOST_DEVICE typename Iterator::value_type
-        make_operator_brackets_result(Iterator const& iter, std::false_type)
+        HPX_HOST_DEVICE decltype(auto) make_operator_brackets_result(
+            Iterator const& iter, std::false_type)
         {
             return *iter;
         }
@@ -333,7 +334,8 @@ namespace hpx::util {
 
             HPX_HOST_DEVICE iterator_facade_base() = default;
 
-            HPX_HOST_DEVICE constexpr auto operator[](difference_type n) const
+            HPX_HOST_DEVICE constexpr decltype(auto) operator[](
+                difference_type n) const
             {
                 using use_proxy = use_operator_brackets_proxy<Derived, T>;

[junghans]

Testing it here: https://koji.fedoraproject.org/koji/taskinfo?taskID=141749396

[hkaiser]

Testing it here: https://koji.fedoraproject.org/koji/taskinfo?taskID=141749396

That's using gcc 15.2.1, I doubt you will be able to see the problem on that platform.

[junghans]

Testing it here: https://koji.fedoraproject.org/koji/taskinfo?taskID=141749396

That's using gcc 15.2.1, I doubt you will be able to see the problem on that platform.

Sorry, wrong distro, here is F44: https://koji.fedoraproject.org/koji/taskinfo?taskID=141779086

[hkaiser]

Testing it here: https://koji.fedoraproject.org/koji/taskinfo?taskID=141749396

That's using gcc 15.2.1, I doubt you will be able to see the problem on that platform.

Sorry, wrong distro, here is F44: https://koji.fedoraproject.org/koji/taskinfo?taskID=141779086

Thanks! Looks like the patch fixed the issue.

[junghans]

Let's wait for the arm64 build and then I will add that patch!

[hkaiser]

I'm still working on making older versions of gcc agree to this patch as well. I'll let you know.

[junghans]

Let's wait for the arm64 build and then I will add that patch!

That build passed as well, so I merged the patch into the rpm: https://koji.fedoraproject.org/koji/taskinfo?taskID=141787402

Hopefully, you can include a similar fix in the next release.

[hkaiser]

@junghans many thanks!

[hkaiser]

@junghans how much effort would it be for you to test this PR (#6868) to verify that the problem will be fixed in the next HPX release?