This is my first time submitting a patch, so my apologies if I'm
submitting incorrectly or missing something.
Clang is unable to compile parts of libstdc++'s ranges implementation
due to LLVM-61763, a Clang frontend compiler bug in handling the
declarations of constrained friends. The problem areas are zip_view,
zip_transform_view, and adjacent_transform_view.
A simple ranges program like the following fails to compile using
Clang trunk and libstdc++.
std::vector v = {1, 2, 3, 4};
int sum = 0;
for (auto&& [i, j] : std::ranges::views::zip(v, v))
sum += i * j;
In file included from <source>:1:
/opt/compiler-explorer/gcc-snapshot/lib/gcc/x86_64-linux-gnu/14.0.0/../../../../include/c++/14.0.0/ranges:4655:14:
error: type constraint differs in template redeclaration
4655 | template<move_constructible _Fp, input_range... _Ws>
| ^
. . . . . .
Godbolt: https://godbolt.org/z/Ynbs15aGh
This patch adds a small workaround that avoids declaring constrained
friends when compiling with Clang, instead making some members public.
MSVC's standard library has implemented a similar workaround.
Scanning through libstdc++, there do appear to be other workarounds
for Clang, e.g. in complex and experimental/simd. Hopefully this kind
of workaround is acceptable---while the core issue is a Clang compiler
bug, it may take a while to fix, and it would be very useful for
libstdc++ ranges to work with Clang in the meantime.
2023-10-13 Benjamin Brock <br...@cs.berkeley.edu>
libstdc++-v3/ChangeLog:
* include/std/ranges: implement workaround for LLVM-61763 in
zip_view and adjacency_view
---
diff --git a/libstdc++-v3/include/std/ranges b/libstdc++-v3/include/std/ranges
index 1d529a886be..7893e3a84c9 100644
--- a/libstdc++-v3/include/std/ranges
+++ b/libstdc++-v3/include/std/ranges
@@ -4632,6 +4632,9 @@ namespace views::__adaptor
class zip_view<_Vs...>::_Iterator
: public __detail::__zip_view_iter_cat<_Const, _Vs...>
{
+#ifdef __clang__ // LLVM-61763 workaround
+ public:
+#endif
__detail::__tuple_or_pair_t<iterator_t<__detail::__maybe_const_t<_Const,
_Vs>>...> _M_current;
constexpr explicit
@@ -4652,11 +4655,13 @@ namespace views::__adaptor
return input_iterator_tag{};
}
+#ifndef __clang__ // LLVM-61763 workaround
template<move_constructible _Fp, input_range... _Ws>
requires (view<_Ws> && ...) && (sizeof...(_Ws) > 0) && is_object_v<_Fp>
&& regular_invocable<_Fp&, range_reference_t<_Ws>...>
&& std::__detail::__can_reference<invoke_result_t<_Fp&,
range_reference_t<_Ws>...>>
friend class zip_transform_view;
+#endif
public:
// iterator_category defined in __zip_view_iter_cat
@@ -5327,6 +5332,9 @@ namespace views::__adaptor
template<bool _Const>
class adjacent_view<_Vp, _Nm>::_Iterator
{
+#ifdef __clang__ // LLVM-61763 workaround
+ public:
+#endif
using _Base = __detail::__maybe_const_t<_Const, _Vp>;
array<iterator_t<_Base>, _Nm> _M_current = array<iterator_t<_Base>, _Nm>();
@@ -5367,12 +5375,14 @@ namespace views::__adaptor
friend class adjacent_view;
+#ifndef __clang__ // LLVM-61763 workaround
template<forward_range _Wp, move_constructible _Fp, size_t _Mm>
requires view<_Wp> && (_Mm > 0) && is_object_v<_Fp>
&& regular_invocable<__detail::__unarize<_Fp&, _Mm>,
range_reference_t<_Wp>>
&&
std::__detail::__can_reference<invoke_result_t<__detail::__unarize<_Fp&,
_Mm>,
range_reference_t<_Wp>>>
friend class adjacent_transform_view;
+#endif
public:
using iterator_category = input_iterator_tag;
