[Bug libstdc++/112882] [14 Regression] std::clamp no longer usable in header only mode

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

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=112882

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
     Ever confirmed|0                           |1
   Last reconfirmed|                            |2023-12-06
             Status|UNCONFIRMED                 |NEW

--- Comment #1 from Jonathan Wakely <redi at gcc dot gnu.org> ---
(In reply to Tamar Christina from comment #0)
> I know this is not a supported scenario, but I'm wondering if it's still
> easy to support.
> 
> We have some libraries that use C++ mostly as an abstraction layer and try
> to ensure that it needs no runtime support from libstdc++.
> 
> A recent commit: g:5e8a30d8b8f4d7ea0a8340b46c1e0c865dbde781 changed how
> `__glibcxx_assert` is defined and now always calls
> `std::__glibcxx_assert_fail`.
> 
> This means that now you always need libstdc++ even in contex where things
> would have been folded away before.  Similarly we're getting the same thing
> through usage of `std::unique_ptr`.

Ugh, that's undesirable. It should only depend on that symbol when
_GLIBCXX_ASSERTIONS is defined, but the change means we also use that symbol
for constexpr checks.

I'm surprised it doesn't get folded away though, since the code looks like:

  if (std::__is_constant_evaluated())
    if (__builtin_expect(COND, false))
      __glibcxx_assert_fail(...);

Since __is_constant_evaluated is always false at runtime, that function can
never be called. Either it's needed during compile-time and makes the program
ill-formed, or it's not needed at all.

Ah, __is_constant_evaluated() is not marked always_inline, so at -O0 it just
looks like a normal function call.

> It seems that undefining `_GLIBCXX_VERBOSE_ASSERT` gets it to go to
> `__builtin_abort()` which makes it work again.
> 
> If this change was intentional, would it be possible to make
> `_GLIBCXX_VERBOSE_ASSERT` user configurable?

You can use --disable-libstdcxx-verbose for that, or do you need to control it
during preprocessing?