[Bug libstdc++/103923] is_invocable<const T &, ...> inexplicably fails

["redi at gcc dot gnu.org" <gcc-bugzilla@gcc.gnu.org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=103923

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |fdumont at gcc dot gnu.org

--- Comment #7 from Jonathan Wakely <redi at gcc dot gnu.org> ---
Ugh, this is a big can of worms.

The current code is wrong, because __cache_default<Key, Hash> can silently give
the wrong answer, and cause ODR violations.

But if we change __cache_default to give the right answer, we break existing
tests:

FAIL: 23_containers/unordered_map/48101_neg.cc  (test for errors, line 27)
FAIL: 23_containers/unordered_set/85965.cc (test for excess errors)

We have tests that depend on NOT checking the hash function until all types are
complete. But that means it is impossible for __cache_default to do the right
thing, because it requires completeness.

The reason that __cache_default checks the hash function is to see whether it
throws or not. The code comments say:

                      // Mandatory to have erase not throwing.
                      __is_nothrow_invocable<const _Hash&, const _Tp&>>>;

and elsewhere:

      // Getting a bucket index from a node shall not throw because it is used
      // in methods (erase, swap...) that shall not throw. Need a complete
      // type to check this, so do it in the destructor not at class scope.
      static_assert(noexcept(declval<const __hash_code_base_access&>()
                        ._M_bucket_index(declval<const __node_value_type&>(),
                                         (std::size_t)0)),
                    "Cache the hash code or qualify your functors involved"
                    " in hash code and bucket index computation with
noexcept");

But this is nonsense. erase is not noexcept. [unord.req.except] specifically
says "erase(k) does not throw an exception unless that exception is thrown by
the container’s Hash or Pred object (if any)".

So why are we doing all this work (and causing bugs) to avoid throwing in
erase, when we are allowed to throw in erase?! swap isn't allowed to call a
throwing hash function, which seems to be a problem because we use
_M_update_bbegin() in swap, but maybe we can do that another way that doesn't
require calling _M_bucket_index.

I'm not sure how we can change the definition of __cache_default now, without
causing ABI breaks. But the current caching policy just seems wrong.