Question about _GLIBCXX_HAVE_BUILTIN_IS_SAME

Question about _GLIBCXX_HAVE_BUILTIN_IS_SAME

[Ken Matsui]

Hi,

I found we are using _GLIBCXX_HAVE_BUILTIN_IS_SAME in type_traits, but I
think we can use _GLIBCXX_USE_BUILTIN_TRAIT(__is_same) instead.  I feel
this is a bit more readable and consistent with other traits.  With this
change, AFAIK, _GLIBCXX_HAVE_BUILTIN_IS_SAME is not used anywhere, but
can we completely remove it from gcc/libstdc++-v3/include/bits/c++config
or is there any reason to keep it?

   /// is_same
+#if _GLIBCXX_USE_BUILTIN_TRAIT(__is_same)
   template<typename _Tp, typename _Up>
     struct is_same
-#ifdef _GLIBCXX_HAVE_BUILTIN_IS_SAME
     : public __bool_constant<__is_same(_Tp, _Up)>
+    { };
 #else
+  template<typename _Tp, typename _Up>
+    struct is_same
     : public false_type
-#endif
     { };
 
-#ifndef _GLIBCXX_HAVE_BUILTIN_IS_SAME
   template<typename _Tp>
     struct is_same<_Tp, _Tp>
     : public true_type
    { };
#endif

I am also wondering if we could replace other _GLIBCXX_HAVE_BUILTIN_*
macros with _GLIBCXX_USE_BUILTIN_TRAIT(*):

* _GLIBCXX_HAVE_BUILTIN_HAS_UNIQ_OBJ_REP
* _GLIBCXX_HAVE_BUILTIN_IS_AGGREGATE
* _GLIBCXX_HAVE_BUILTIN_LAUNDER

-- 
Ken Matsui

Re: Question about _GLIBCXX_HAVE_BUILTIN_IS_SAME

[Jonathan Wakely]

[-- Attachment #1: Type: text/plain, Size: 2343 bytes --]

On Wed, 7 Feb 2024 at 21:59, Ken Matsui <kmatsui@cs.washington.edu> wrote:

> Hi,
>
> I found we are using _GLIBCXX_HAVE_BUILTIN_IS_SAME in type_traits, but I
> think we can use _GLIBCXX_USE_BUILTIN_TRAIT(__is_same) instead.  I feel
> this is a bit more readable and consistent with other traits.  With this
>

Agreed.


> change, AFAIK, _GLIBCXX_HAVE_BUILTIN_IS_SAME is not used anywhere, but
> can we completely remove it from gcc/libstdc++-v3/include/bits/c++config
> or is there any reason to keep it?
>

No, it can go.

The reason we have that macro is historical. Originally, gcc did not define
__is_same, only __is_same_as. Clang defined __is_same. So we needed to use
a different built-in depending which compiler we were using.

Since https://gcc.gnu.org/g:73ae6eb57251 we just use __is_same for both GCC
and Clang (and Intel). We can simplify it as you suggest.



>
>    /// is_same
> +#if _GLIBCXX_USE_BUILTIN_TRAIT(__is_same)
>    template<typename _Tp, typename _Up>
>      struct is_same
> -#ifdef _GLIBCXX_HAVE_BUILTIN_IS_SAME
>      : public __bool_constant<__is_same(_Tp, _Up)>
> +    { };
>  #else
> +  template<typename _Tp, typename _Up>
> +    struct is_same
>      : public false_type
> -#endif
>      { };
>
> -#ifndef _GLIBCXX_HAVE_BUILTIN_IS_SAME
>    template<typename _Tp>
>      struct is_same<_Tp, _Tp>
>      : public true_type
>     { };
> #endif
>
> I am also wondering if we could replace other _GLIBCXX_HAVE_BUILTIN_*
> macros with _GLIBCXX_USE_BUILTIN_TRAIT(*):
>
> * _GLIBCXX_HAVE_BUILTIN_HAS_UNIQ_OBJ_REP
> * _GLIBCXX_HAVE_BUILTIN_IS_AGGREGATE
> * _GLIBCXX_HAVE_BUILTIN_LAUNDER
>

Yes, let's normalize all of these to use the new form.

At some point we should also simplify this:

#ifdef __has_builtin
# ifdef __is_identifier
// Intel and older Clang require !__is_identifier for some built-ins:
#  define _GLIBCXX_HAS_BUILTIN(B) __has_builtin(B) || ! __is_identifier(B)
# else
#  define _GLIBCXX_HAS_BUILTIN(B) __has_builtin(B)
# endif
#endif

I think we can stop supporting old versions of Clang where
__is_builtin(__is_same) is false, but we should verify which version of
Clang stopped requiring !__is_identifier for those built-ins. If it's older
than Clang 14 we can remove _GLIBCXX_HAS_BUILTIN and just use __has_builtin
directly in _GLIBCXX_USE_BUILTIN_TRAIT. That should wait for GCC 15 though.

Re: Question about _GLIBCXX_HAVE_BUILTIN_IS_SAME

[Ken Matsui]

On Wed, 07 Feb 2024 at 16:13, Jonathan Wakely <jwakely@redhat.com> wrote:
> 
> 
> On Wed, 7 Feb 2024 at 21:59, Ken Matsui <kmatsui@cs.washington.edu> wrote:
> 
>     Hi,
> 
>     I found we are using _GLIBCXX_HAVE_BUILTIN_IS_SAME in type_traits, but I
>     think we can use _GLIBCXX_USE_BUILTIN_TRAIT(__is_same) instead.  I feel
>     this is a bit more readable and consistent with other traits.  With this
> 
> 
> Agreed.
>  
> 
>     change, AFAIK, _GLIBCXX_HAVE_BUILTIN_IS_SAME is not used anywhere, but
>     can we completely remove it from gcc/libstdc++-v3/include/bits/c++config
>     or is there any reason to keep it?
> 
> 
> No, it can go.
> 
> The reason we have that macro is historical. Originally, gcc did not define
> __is_same, only __is_same_as. Clang defined __is_same. So we needed to use a
> different built-in depending which compiler we were using.
> 
> Since https://gcc.gnu.org/g:73ae6eb57251 we just use __is_same for both GCC and
> Clang (and Intel). We can simplify it as you suggest.
> 
>  
> 
> 
>        /// is_same
>     +#if _GLIBCXX_USE_BUILTIN_TRAIT(__is_same)
>        template<typename _Tp, typename _Up>
>          struct is_same
>     -#ifdef _GLIBCXX_HAVE_BUILTIN_IS_SAME
>          : public __bool_constant<__is_same(_Tp, _Up)>
>     +    { };
>      #else
>     +  template<typename _Tp, typename _Up>
>     +    struct is_same
>          : public false_type
>     -#endif
>          { };
> 
>     -#ifndef _GLIBCXX_HAVE_BUILTIN_IS_SAME
>        template<typename _Tp>
>          struct is_same<_Tp, _Tp>
>          : public true_type
>         { };
>     #endif
> 
>     I am also wondering if we could replace other _GLIBCXX_HAVE_BUILTIN_*
>     macros with _GLIBCXX_USE_BUILTIN_TRAIT(*):
> 
>     * _GLIBCXX_HAVE_BUILTIN_HAS_UNIQ_OBJ_REP
>     * _GLIBCXX_HAVE_BUILTIN_IS_AGGREGATE
>     * _GLIBCXX_HAVE_BUILTIN_LAUNDER
> 
> 
> Yes, let's normalize all of these to use the new form.
> 

It seems that these macros are used for feature test macros.  I think
it does not make sense to use _GLIBCXX_USE_BUILTIN_TRAIT for them IIUC.
Or do you think we can replace them with _GLIBCXX_USE_BUILTIN_TRAIT?

#if !defined(__cpp_lib_launder)
# if (__cplusplus >= 201703L) && (defined(_GLIBCXX_HAVE_BUILTIN_LAUNDER))
#  define __glibcxx_launder 201606L
#  if defined(__glibcxx_want_all) || defined(__glibcxx_want_launder)
#   define __cpp_lib_launder 201606L
#  endif
# endif
#endif /* !defined(__cpp_lib_launder) && defined(__glibcxx_want_launder) */
#undef __glibcxx_want_launder

> At some point we should also simplify this:
> 
> #ifdef __has_builtin
> # ifdef __is_identifier
> // Intel and older Clang require !__is_identifier for some built-ins:
> #  define _GLIBCXX_HAS_BUILTIN(B) __has_builtin(B) || ! __is_identifier(B)
> # else
> #  define _GLIBCXX_HAS_BUILTIN(B) __has_builtin(B)
> # endif
> #endif
> 
> I think we can stop supporting old versions of Clang where __is_builtin
> (__is_same) is false, but we should verify which version of Clang stopped
> requiring !__is_identifier for those built-ins. If it's older than Clang 14 we
> can remove _GLIBCXX_HAS_BUILTIN and just use __has_builtin directly in
> _GLIBCXX_USE_BUILTIN_TRAIT. That should wait for GCC 15 though.
> 
> 
> 
> 
>  

-- 
Ken Matsui

Re: Question about _GLIBCXX_HAVE_BUILTIN_IS_SAME

[Ken Matsui]

On Fri, Feb 9, 2024 at 11:35 AM Ken Matsui <kmatsui@cs.washington.edu> wrote:
>
> On Wed, 07 Feb 2024 at 16:13, Jonathan Wakely <jwakely@redhat.com> wrote:
> >
> >
> > On Wed, 7 Feb 2024 at 21:59, Ken Matsui <kmatsui@cs.washington.edu> wrote:
> >
> >     Hi,
> >
> >     I found we are using _GLIBCXX_HAVE_BUILTIN_IS_SAME in type_traits, but I
> >     think we can use _GLIBCXX_USE_BUILTIN_TRAIT(__is_same) instead.  I feel
> >     this is a bit more readable and consistent with other traits.  With this
> >
> >
> > Agreed.
> >
> >
> >     change, AFAIK, _GLIBCXX_HAVE_BUILTIN_IS_SAME is not used anywhere, but
> >     can we completely remove it from gcc/libstdc++-v3/include/bits/c++config
> >     or is there any reason to keep it?
> >
> >
> > No, it can go.
> >
> > The reason we have that macro is historical. Originally, gcc did not define
> > __is_same, only __is_same_as. Clang defined __is_same. So we needed to use a
> > different built-in depending which compiler we were using.
> >
> > Since https://gcc.gnu.org/g:73ae6eb57251 we just use __is_same for both GCC and
> > Clang (and Intel). We can simplify it as you suggest.
> >
> >
> >
> >
> >        /// is_same
> >     +#if _GLIBCXX_USE_BUILTIN_TRAIT(__is_same)
> >        template<typename _Tp, typename _Up>
> >          struct is_same
> >     -#ifdef _GLIBCXX_HAVE_BUILTIN_IS_SAME
> >          : public __bool_constant<__is_same(_Tp, _Up)>
> >     +    { };
> >      #else
> >     +  template<typename _Tp, typename _Up>
> >     +    struct is_same
> >          : public false_type
> >     -#endif
> >          { };
> >
> >     -#ifndef _GLIBCXX_HAVE_BUILTIN_IS_SAME
> >        template<typename _Tp>
> >          struct is_same<_Tp, _Tp>
> >          : public true_type
> >         { };
> >     #endif
> >
> >     I am also wondering if we could replace other _GLIBCXX_HAVE_BUILTIN_*
> >     macros with _GLIBCXX_USE_BUILTIN_TRAIT(*):
> >
> >     * _GLIBCXX_HAVE_BUILTIN_HAS_UNIQ_OBJ_REP
> >     * _GLIBCXX_HAVE_BUILTIN_IS_AGGREGATE
> >     * _GLIBCXX_HAVE_BUILTIN_LAUNDER
> >
> >
> > Yes, let's normalize all of these to use the new form.
> >
>
> It seems that these macros are used for feature test macros.  I think
> it does not make sense to use _GLIBCXX_USE_BUILTIN_TRAIT for them IIUC.
> Or do you think we can replace them with _GLIBCXX_USE_BUILTIN_TRAIT?
>
> #if !defined(__cpp_lib_launder)
> # if (__cplusplus >= 201703L) && (defined(_GLIBCXX_HAVE_BUILTIN_LAUNDER))
> #  define __glibcxx_launder 201606L
> #  if defined(__glibcxx_want_all) || defined(__glibcxx_want_launder)
> #   define __cpp_lib_launder 201606L
> #  endif
> # endif
> #endif /* !defined(__cpp_lib_launder) && defined(__glibcxx_want_launder) */
> #undef __glibcxx_want_launder
>

I talked with Patrick about this.  Since we don't have a fallback
implementation for the following traits, it would be better to keep
the current implementation instead of using
_GLIBCXX_USE_BUILTIN_TRAIT.  For _GLIBCXX_HAS_BUILTIN, if it works
with Clang 14, I will submit a patch.  Thank you!

> >     * _GLIBCXX_HAVE_BUILTIN_HAS_UNIQ_OBJ_REP
> >     * _GLIBCXX_HAVE_BUILTIN_IS_AGGREGATE
> >     * _GLIBCXX_HAVE_BUILTIN_LAUNDER

> > At some point we should also simplify this:
> >
> > #ifdef __has_builtin
> > # ifdef __is_identifier
> > // Intel and older Clang require !__is_identifier for some built-ins:
> > #  define _GLIBCXX_HAS_BUILTIN(B) __has_builtin(B) || ! __is_identifier(B)
> > # else
> > #  define _GLIBCXX_HAS_BUILTIN(B) __has_builtin(B)
> > # endif
> > #endif
> >
> > I think we can stop supporting old versions of Clang where __is_builtin
> > (__is_same) is false, but we should verify which version of Clang stopped
> > requiring !__is_identifier for those built-ins. If it's older than Clang 14 we
> > can remove _GLIBCXX_HAS_BUILTIN and just use __has_builtin directly in
> > _GLIBCXX_USE_BUILTIN_TRAIT. That should wait for GCC 15 though.
> >
> >
> >
> >
> >
>
> --
> Ken Matsui