[Bug c++/102063] New: Default arguments should not be a SFINAE context

[Bug c++/102063] New: Default arguments should not be a SFINAE context

[redi at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=102063

            Bug ID: 102063
           Summary: Default arguments should not be a SFINAE context
           Product: gcc
           Version: 12.0
            Status: UNCONFIRMED
          Keywords: accepts-invalid
          Severity: normal
          Priority: P3
         Component: c++
          Assignee: unassigned at gcc dot gnu.org
          Reporter: redi at gcc dot gnu.org
  Target Milestone: ---

G++ compiles this invalid code:

struct base { };

template<typename T>
   void
   can_haz_base(base* = (T*)nullptr);  // XXX

template<typename T, typename = void>
struct haz_base
{
  static constexpr bool value = false;
};

template<typename T>
struct haz_base<T, decltype(can_haz_base<T>())>
{
  static constexpr bool value = true;
};

struct X { };

static_assert( ! haz_base<X>::value, "" );


The default argument at XXX is not in "the immediate context of the function
type, its template parameter types, and its explicit-specifier", and so should
not result in substitution failure. I think it should be an error, although
there is implementation divergence. GCC and MSVC accept it.


Clang says:

def.C:5:23: error: cannot initialize a parameter of type 'base *' with an
rvalue of type 'X *'
   can_haz_base(base* = (T*)nullptr);
                      ^ ~~~~~~~~~~~
def.C:14:29: note: in instantiation of default function argument expression for
'can_haz_base<X>' required here
struct haz_base<T, decltype(can_haz_base<T>())>
                            ^
def.C:21:18: note: during template argument deduction for class template
partial specialization 'haz_base<T, decltype(can_haz_base<T>())>' [with T = X]
static_assert( ! haz_base<X>::value, "" );
                 ^
def.C:21:18: note: in instantiation of template class 'haz_base<X>' requested
here
def.C:5:23: note: passing argument to parameter here
   can_haz_base(base* = (T*)nullptr);
                      ^
1 error generated.


EDG says:

"def.C", line 5: error: default argument of type "X *" is incompatible with
          parameter of type "base *"
     can_haz_base(base* = (T*)nullptr);
                          ^
          detected during instantiation of "void can_haz_base<T>(base *) [with
                    T=X]" at line 21

"def.C", line 21: error: static assertion failed with ""
  static_assert( ! haz_base<X>::value, "" );
  ^

2 errors detected in the compilation of "def.C".

[Bug c++/102063] Default arguments should not be a SFINAE context

[ppalka at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=102063

Patrick Palka <ppalka at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |ppalka at gcc dot gnu.org
           See Also|                            |https://gcc.gnu.org/bugzill
                   |                            |a/show_bug.cgi?id=95607

--- Comment #1 from Patrick Palka <ppalka at gcc dot gnu.org> ---
FWIW there's the EWG paper P2285 about default arguments and SFINAE, which IIUC
argues for something closer to GCC/MSVC's behavior than Clang/EDG's

[Bug c++/102063] Default arguments should not be a SFINAE context

[redi at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=102063

--- Comment #2 from Jonathan Wakely <redi at gcc dot gnu.org> ---
Interesting, thanks!

[Bug c++/102063] Default arguments should not be a SFINAE context

[redi at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=102063

--- Comment #3 from Jonathan Wakely <redi at gcc dot gnu.org> ---
Also Jason's https://wg21.link/cwg2296