Merge pull request #985 from NVIDIA/intrinsic-traits

use compiler intrinsics for some expensive type traits

Author: ericniebler

include/exec/any_sender_of.hpp

@@ -362,7 +362,7 @@ namespace exec {
     class __storage<_Vtable, _Allocator, _Copyable, _Alignment, _InlineSize>::__t
       : __if_c<_Copyable, __, __move_only> {
       static_assert(
-        std::is_convertible_v<typename std::allocator_traits<_Allocator>::void_pointer, void*>);
+        STDEXEC_IS_CONVERTIBLE_TO(typename std::allocator_traits<_Allocator>::void_pointer, void*));
 
       static constexpr std::size_t __buffer_size = std::max(_InlineSize, sizeof(void*));
       static constexpr std::size_t __alignment = std::max(_Alignment, alignof(void*));

include/nvexec/stream/common.cuh

@@ -357,12 +357,12 @@ namespace nvexec {
     template <class S>
     concept stream_sender = //
       sender<S> &&          //
-      std::is_base_of_v<stream_sender_base, __decay_t<S>>;
+      STDEXEC_IS_BASE_OF(stream_sender_base, __decay_t<S>);
 
     template <class R>
     concept stream_receiver = //
       receiver<R> &&          //
-      std::is_base_of_v<stream_receiver_base, __decay_t<R>>;
+      STDEXEC_IS_BASE_OF(stream_receiver_base, __decay_t<R>);
 
     struct stream_op_state_base { };
 

include/nvexec/stream/when_all.cuh

@@ -235,7 +235,7 @@ namespace nvexec::STDEXEC_STREAM_DETAIL_NS {
 
         template <class OpT>
         static void sync(OpT& op) noexcept {
-          if constexpr (std::is_base_of_v<stream_op_state_base, OpT>) {
+          if constexpr (STDEXEC_IS_BASE_OF(stream_op_state_base, OpT)) {
             if (op.stream_provider_.status_ == cudaSuccess) {
               op.stream_provider_.status_ = STDEXEC_DBG_ERR(cudaStreamSynchronize(op.get_stream()));
             }

include/stdexec/__detail/__config.hpp

@@ -95,6 +95,14 @@
 #define STDEXEC_IS_BASE_OF(...) std::is_base_of_v<__VA_ARGS__>
 #endif
 
+#if STDEXEC_HAS_BUILTIN(__is_convertible_to) || STDEXEC_MSVC()
+#define STDEXEC_IS_CONVERTIBLE_TO(...) __is_convertible_to(__VA_ARGS__)
+#elif STDEXEC_HAS_BUILTIN(__is_convertible)
+#define STDEXEC_IS_CONVERTIBLE_TO(...) __is_convertible(__VA_ARGS__)
+#else
+#define STDEXEC_IS_CONVERTIBLE_TO(...) std::is_convertible_v<__VA_ARGS__>
+#endif
+
 // Before gcc-12, gcc really didn't like tuples or variants of immovable types
 #if STDEXEC_GCC() && (__GNUC__ < 12)
 #define STDEXEC_IMMOVABLE(_XP) _XP(_XP&&)

include/stdexec/concepts.hpp

@@ -55,13 +55,13 @@ namespace stdexec::__std_concepts {
   concept integral = std::is_integral_v<T>;
 
   template <class _Ap, class _Bp>
-  concept derived_from =           //
-    std::is_base_of_v<_Bp, _Ap> && //
-    std::is_convertible_v<const volatile _Ap*, const volatile _Bp*>;
+  concept derived_from =            //
+    STDEXEC_IS_BASE_OF(_Bp, _Ap) && //
+    STDEXEC_IS_CONVERTIBLE_TO(const volatile _Ap*, const volatile _Bp*);
 
   template <class _From, class _To>
-  concept convertible_to =               //
-    std::is_convertible_v<_From, _To> && //
+  concept convertible_to =                   //
+    STDEXEC_IS_CONVERTIBLE_TO(_From, _To) && //
     requires(_From (&__fun)()) { static_cast<_To>(__fun()); };
 
   template <class _Ty>

include/stdexec/execution.hpp

@@ -2533,7 +2533,7 @@ namespace stdexec {
       requires __decays_to<_Tp, _Tp>
     {
       static_assert(std::is_reference_v<__copy_cvref_t<_Up&&, _Tp>>);
-      static_assert(std::is_base_of_v<_Tp, __decay_t<_Up>>);
+      static_assert(STDEXEC_IS_BASE_OF(_Tp, __decay_t<_Up>));
       return (__copy_cvref_t<_Up&&, _Tp>) (_Up&&) u;
     }