Rework builtin handling (NVIDIA#2461)

* Move builtin detection to its own file * Try to reenable more builtins * Address review comments
fbusato · Oct 2, 2024 · caff9bc · caff9bc
1 parent 2b8de16
commit caff9bc
Show file tree

Hide file tree

Showing 212 changed files with 1,204 additions and 1,103 deletions.
diff --git a/cudax/include/cuda/experimental/__async/meta.cuh b/cudax/include/cuda/experimental/__async/meta.cuh
@@ -534,13 +534,13 @@ using __m_at = _Ts...[__v<_Np>];
 template <size_t _Np, class... _Ts>
 using __m_at_c = _Ts...[_Np];
 
-#elif __has_builtin(__type_pack_element)
+#elif defined(_CCCL_BUILTIN_TYPE_PACK_ELEMENT)
 
 template <bool>
 struct __m_at_
 {
   template <class _Np, class... _Ts>
-  using __f = __type_pack_element<__v<_Np>, _Ts...>;
+  using __f = _CCCL_BUILTIN_TYPE_PACK_ELEMENT(__v<_Np>, _Ts...);
 };
 
 template <class _Np, class... _Ts>
@@ -675,18 +675,9 @@ struct __mconcat_into_q
 };
 
 // The following must be super-fast to compile, so use an intrinsic directly if it is available
-#if defined(_LIBCUDACXX_IS_BASE_OF) && !defined(_LIBCUDACXX_USE_IS_BASE_OF_FALLBACK)
-
-template <class _Set, class... _Ty>
-_CCCL_INLINE_VAR constexpr bool __mset_contains = (_LIBCUDACXX_IS_BASE_OF(__mtype<_Ty>, _Set) && ...);
-
-#else
-
 template <class _Set, class... _Ty>
 _CCCL_INLINE_VAR constexpr bool __mset_contains = (_CUDA_VSTD::is_base_of_v<__mtype<_Ty>, _Set> && ...);
 
-#endif
-
 namespace __set
 {
 template <class... _Ts>

diff --git a/cudax/include/cuda/experimental/__async/type_traits.cuh b/cudax/include/cuda/experimental/__async/type_traits.cuh
@@ -30,26 +30,13 @@
 
 namespace cuda::experimental::__async
 {
-#if __has_builtin(__remove_reference)
 
 template <class _Ty>
-using __remove_ref_t = __remove_reference(_Ty);
-
-#elif __has_builtin(__remove_reference_t)
-
-template <class _Ty>
-using __remove_ref_t = __remove_reference_t(_Ty);
-
-#else
-
-template <class _Ty>
-using __remove_ref_t = _CUDA_VSTD::remove_reference_t<_Ty>;
-
-#endif
+using __remove_ref_t = _CUDA_VSTD::__libcpp_remove_reference_t<_Ty>;
 
 //////////////////////////////////////////////////////////////////////////////////////////////////
 // __decay_t: An efficient implementation for ::std::decay
-#if __has_builtin(__decay)
+#if defined(_CCCL_BUILTIN_DECAY)
 
 template <class _Ty>
 using __decay_t = __decay(_Ty);
@@ -59,7 +46,7 @@ using __decay_t = __decay(_Ty);
 //   template <class _Ty>
 //   using __decay_t = _CUDA_VSTD::decay_t<_Ty>;
 
-#else
+#else // ^^^ _CCCL_BUILTIN_DECAY ^^^ / vvv !_CCCL_BUILTIN_DECAY vvv
 
 struct __decay_object
 {
@@ -141,7 +128,7 @@ inline __decay_void __mdecay<void const>;
 template <class _Ty>
 using __decay_t = typename decltype(__mdecay<_Ty>)::template __f<_Ty>;
 
-#endif
+#endif // _CCCL_BUILTIN_DECAY
 
 //////////////////////////////////////////////////////////////////////////////////////////////////
 // __copy_cvref_t: For copying cvref from one type to another

diff --git a/libcudacxx/include/cuda/__cccl_config b/libcudacxx/include/cuda/__cccl_config
@@ -12,6 +12,7 @@
 #define _CUDA__CCCL_CONFIG
 
 #include <cuda/std/__cccl/attributes.h> // IWYU pragma: export
+#include <cuda/std/__cccl/builtin.h> // IWYU pragma: export
 #include <cuda/std/__cccl/compiler.h> // IWYU pragma: export
 #include <cuda/std/__cccl/diagnostic.h> // IWYU pragma: export
 #include <cuda/std/__cccl/dialect.h> // IWYU pragma: export

diff --git a/libcudacxx/include/cuda/std/__algorithm/copy.h b/libcudacxx/include/cuda/std/__algorithm/copy.h
@@ -91,13 +91,13 @@ template <class _Tp, class _Up>
 _LIBCUDACXX_HIDE_FROM_ABI _CCCL_CONSTEXPR_CXX14 bool __constexpr_tail_overlap(_Tp* __first, _Up* __needle, _Tp* __last)
 {
   _LIBCUDACXX_UNUSED_VAR(__last);
-#if __has_builtin(__builtin_constant_p) || defined(_CCCL_COMPILER_GCC)
+#if defined(_CCCL_BUILTIN_CONSTANT_P)
   NV_IF_ELSE_TARGET(NV_IS_HOST,
-                    (return __builtin_constant_p(__first < __needle) && __first < __needle;),
+                    (return _CCCL_BUILTIN_CONSTANT_P(__first < __needle) && __first < __needle;),
                     (return __constexpr_tail_overlap_fallback(__first, __needle, __last);))
-#else
+#else // ^^^ _CCCL_BUILTIN_CONSTANT_P ^^^ / vvv !_CCCL_BUILTIN_CONSTANT_P vvv
   return __constexpr_tail_overlap_fallback(__first, __needle, __last);
-#endif
+#endif // !_CCCL_BUILTIN_CONSTANT_P
 }
 
 template <class _AlgPolicy,

diff --git a/libcudacxx/include/cuda/std/__bit/bit_cast.h b/libcudacxx/include/cuda/std/__bit/bit_cast.h
@@ -29,16 +29,16 @@
 
 _LIBCUDACXX_BEGIN_NAMESPACE_STD
 
-#if defined(_LIBCUDACXX_BIT_CAST)
+#if defined(_CCCL_BUILTIN_BIT_CAST)
 #  define _LIBCUDACXX_CONSTEXPR_BIT_CAST constexpr
-#else // ^^^ _LIBCUDACXX_BIT_CAST ^^^ / vvv !_LIBCUDACXX_BIT_CAST vvv
+#else // ^^^ _CCCL_BUILTIN_BIT_CAST ^^^ / vvv !_CCCL_BUILTIN_BIT_CAST vvv
 #  define _LIBCUDACXX_CONSTEXPR_BIT_CAST
 #  if defined(_CCCL_COMPILER_GCC) && __GNUC__ >= 8
 // GCC starting with GCC8 warns about our extended floating point types having protected data members
 _CCCL_DIAG_PUSH
 _CCCL_DIAG_SUPPRESS_GCC("-Wclass-memaccess")
 #  endif // _CCCL_COMPILER_GCC >= 8
-#endif // !_LIBCUDACXX_BIT_CAST
+#endif // !_CCCL_BUILTIN_BIT_CAST
 
 template <
   class _To,
@@ -48,23 +48,23 @@ template <
   __enable_if_t<_CCCL_TRAIT(is_trivially_copyable, _From) || _CCCL_TRAIT(__is_extended_floating_point, _From), int> = 0>
 _CCCL_NODISCARD _LIBCUDACXX_HIDE_FROM_ABI _LIBCUDACXX_CONSTEXPR_BIT_CAST _To bit_cast(const _From& __from) noexcept
 {
-#if defined(_LIBCUDACXX_BIT_CAST)
-  return _LIBCUDACXX_BIT_CAST(_To, __from);
-#else // ^^^ _LIBCUDACXX_BIT_CAST ^^^ / vvv !_LIBCUDACXX_BIT_CAST vvv
+#if defined(_CCCL_BUILTIN_BIT_CAST)
+  return _CCCL_BUILTIN_BIT_CAST(_To, __from);
+#else // ^^^ _CCCL_BUILTIN_BIT_CAST ^^^ / vvv !_CCCL_BUILTIN_BIT_CAST vvv
   static_assert(_CCCL_TRAIT(is_trivially_default_constructible, _To),
                 "The compiler does not support __builtin_bit_cast, so bit_cast additionally requires the destination "
                 "type to be trivially constructible");
   _To __temp;
   _CUDA_VSTD::memcpy(&__temp, &__from, sizeof(_To));
   return __temp;
-#endif // !_LIBCUDACXX_BIT_CAST
+#endif // !_CCCL_BUILTIN_BIT_CAST
 }
 
-#if !defined(_LIBCUDACXX_BIT_CAST)
+#if !defined(_CCCL_BUILTIN_BIT_CAST)
 #  if defined(_CCCL_COMPILER_GCC) && __GNUC__ >= 8
 _CCCL_DIAG_POP
 #  endif // _CCCL_COMPILER_GCC >= 8
-#endif // !_LIBCUDACXX_BIT_CAST
+#endif // !_CCCL_BUILTIN_BIT_CAST
 
 _LIBCUDACXX_END_NAMESPACE_STD
 

diff --git a/libcudacxx/include/cuda/std/__bit/clz.h b/libcudacxx/include/cuda/std/__bit/clz.h
@@ -112,7 +112,7 @@ _LIBCUDACXX_HIDE_FROM_ABI constexpr int __libcpp_clz(uint32_t __x)
     }
     return 32; // Undefined Behavior.
   }
-#  endif // _LIBCUDACXX_IS_CONSTANT_EVALUATED && !defined(__CUDA_ARCH__)
+#  endif // _CCCL_BUILTIN_IS_CONSTANT_EVALUATED && !defined(__CUDA_ARCH__)
 
   return __binary_clz32(static_cast<uint64_t>(__x), 0);
 }
@@ -141,7 +141,7 @@ _LIBCUDACXX_HIDE_FROM_ABI constexpr int __libcpp_clz(uint64_t __x)
 #    endif
     return 64; // Undefined Behavior.
   }
-#  endif // _LIBCUDACXX_IS_CONSTANT_EVALUATED && !defined(__CUDA_ARCH__)
+#  endif // _CCCL_BUILTIN_IS_CONSTANT_EVALUATED && !defined(__CUDA_ARCH__)
 
   return __binary_clz64(static_cast<uint64_t>(__x));
 }

diff --git a/libcudacxx/include/cuda/std/__bit/ctz.h b/libcudacxx/include/cuda/std/__bit/ctz.h
@@ -114,7 +114,7 @@ _LIBCUDACXX_HIDE_FROM_ABI constexpr int __libcpp_ctz(uint32_t __x)
     }
     return 32;
   }
-#  endif // _LIBCUDACXX_IS_CONSTANT_EVALUATED && !defined(__CUDA_ARCH__)
+#  endif // _CCCL_BUILTIN_IS_CONSTANT_EVALUATED && !defined(__CUDA_ARCH__)
 
   return __binary_ctz32(static_cast<uint64_t>(__x), 0);
 }
@@ -143,7 +143,7 @@ _LIBCUDACXX_HIDE_FROM_ABI constexpr int __libcpp_ctz(uint64_t __x)
 #    endif
     return 64;
   }
-#  endif // _LIBCUDACXX_IS_CONSTANT_EVALUATED && !defined(__CUDA_ARCH__)
+#  endif // _CCCL_BUILTIN_IS_CONSTANT_EVALUATED && !defined(__CUDA_ARCH__)
 
   return __binary_ctz64(__x);
 }