offset_based_getter produces wrong result with clang on non-X86 platform

[varlax77]

The attached mini reproducer fails if compiled with Clang-16 on non-X86 platform:

#include <boost/pfr/detail/offset_based_getter.hpp>
#include <boost/pfr/detail/sequence_tuple.hpp>

#include <boost/core/lightweight_test.hpp>

struct user_type {
    char c;
    double d;
};

int main() {
   using pfr_tuple = boost::pfr::detail::sequence_tuple::tuple<char, double>;
   using getter = boost::pfr::detail::offset_based_getter<user_type, pfr_tuple>;
   using boost::pfr::detail::size_t_;
   using boost::pfr::detail::sequence_tuple::get;

   user_type value{};
   auto begin = reinterpret_cast<char*>(&value);
   auto native_offset = reinterpret_cast<char*>(&value.d) - begin;

   auto getter_offset = reinterpret_cast<char*>(&getter{}.get(value, size_t_<1>{})) - begin;
   if (native_offset != getter_offset) {
        printf("native_offset != getter_offset ('%li' != '%li')\n", native_offset, getter_offset);
	printf("FAIL\n");
   } else {
	printf("PASS\n");
   }   
}

Test output is:

native_offset != getter_offset ('8' != '4')
FAIL

[varlax77]

@apolukhin Apparently this problem is caused by the following workaround in detail/offset_based_getter.hpp:

template <typename... Ts>
struct tuple_of_aligned_storage<sequence_tuple::tuple<Ts...>> {
  using type = sequence_tuple::tuple<internal_aligned_storage<sizeof(Ts),
#if defined(__GNUC__) && __GNUC__ < 8 && !defined(__x86_64__) && !defined(__CYGWIN__)
      // Before GCC-8 the `alignof` was returning the optimal alignment rather than the minimal one.
      // We have to adjust the alignment because otherwise we get the wrong offset.
      (alignof(Ts) > 4 ? 4 : alignof(Ts))
#else
      alignof(Ts)
#endif
  >...>;
};

Probably the condition for workaround applicability is too wide and could be fixed with extra clause && !defined(__clang__).

[denzor200]

Guess this snipped is related to the issue: https://godbolt.org/z/7xfaKf45c

[denzor200]

@apolukhin look at these 2 lines: (alignof(Ts) > 4 ? 4 : alignof(Ts)) and alignof(Ts). Here we take alignment of types, but why we can't take alignment of field accessing expression? Sticking up this way, both lines might be simplyfied to one:

template <typename T, typename I>
struct tuple_of_aligned_storage;

template <typename... Ts, std::size_t... Is>
struct tuple_of_aligned_storage<sequence_tuple::tuple<Ts...>, std::integer_sequence<std::size_t, Is...>> {
  using tuple = sequence_tuple::tuple<Ts...>;
  using type = sequence_tuple::tuple<internal_aligned_storage<sizeof(Ts),
      sequence_tuple::get_align<Is, tuple>()
  >...>;
};

And what is the more important - without tricky ifdefs, nice!

All what we need is to extend sequence_tuple adding get_align method:

template <std::size_t N, class T>
constexpr auto get_align_impl(const base_from_member<N, T>& t) noexcept {
    // NOLINTNEXTLINE(clang-analyzer-core.uninitialized.UndefReturn)
    return std::integral_constant<std::size_t, alignof(t.value)>{};
}

template <std::size_t N, class Tuple>
constexpr auto get_align() noexcept {
    static_assert(N < Tuple::size_v, "====================> Boost.PFR: Tuple index out of bounds");
    using A = decltype(sequence_tuple::get_align_impl<N>(std::declval<const Tuple&>()));
    return A::value;
}

The whole working snipped you could see here: https://godbolt.org/z/P9c3WMeha In my opinion it looks more reliable than the current implementation. What do you think about it, Antony?