[Bug libstdc++/69116] [4.9/5/6 Regression] compile error when including valarray

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

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=69116

--- Comment #7 from Jonathan Wakely <redi at gcc dot gnu.org> ---
The problem overload is this one:

template<typename _Tp>
  inline _Expr<_BinClos<__shift_left, _Constant, _ValArray, _Tp, _Tp>, typename
__fun<__shift_left, _Tp>::result_type>
  operator <<(const _Tp& __t, const valarray<_Tp>& __v)
  {
    typedef _BinClos<__shift_left, _Constant, _ValArray, _Tp, _Tp> _Closure;
    typedef typename __fun<__shift_left, _Tp>::result_type _Rt;
    return _Expr<_Closure, _Rt>(_Closure(__t, __v));
  }

That deduces _Tp as x and so tries to instantiate valarray<x> to see if __v can
be initialized from std::endl, which causes the "invalid abstract return type"
outside the immediate context.

It can be fixed using enable_if (or __gnu_cxx::__enable_if since this is C++98
code) to reject invalid types for valarray, maybe using __is_abstract. I just
need to figure out the best place to put it, and how many of the other
overloaded operators are affected.

It might make sense to make the first argument a non-deduced context, so we
only use this operator when the RHS is really a valarray, rather than matching
for anything on the LHS and then trying to convert the RHS to a valarray. I
don't know if that would break real code using valarray.