NVIDIA · ericniebler · Aug 4, 2025 · Aug 2, 2025 · Aug 4, 2025 · Aug 4, 2025
diff --git a/include/exec/__detail/__bit_cast.hpp b/include/exec/__detail/__bit_cast.hpp
@@ -39,7 +39,7 @@ namespace exec {
     requires(sizeof(_To) == sizeof(_From))
   [[nodiscard]]
   constexpr _To bit_cast(const _From& __from) noexcept {
-#  if STDEXEC_HAS_BUILTIN(__builtin_bit_cast) || (_MSC_VER >= 1926)
+#  if STDEXEC_HAS_BUILTIN(__builtin_bit_cast) || (STDEXEC_MSVC() && STDEXEC_MSVC_VERSION >= 19'26)
     return __builtin_bit_cast(_To, __from);
 #  else
     _To __to;

diff --git a/include/exec/__detail/__numa.hpp b/include/exec/__detail/__numa.hpp
@@ -17,7 +17,6 @@
 #pragma once
 
 #include "../../stdexec/__detail/__config.hpp"
-#include "../../stdexec/__detail/__meta.hpp"
 #include "../scope.hpp" // IWYU pragma: keep
 
 #include <algorithm> // IWYU pragma: keep
@@ -27,14 +26,10 @@
 #include <thread>
 #include <utility>
 
-// Work around a bug in the NVHPC compilers prior to version 24.03
-#if STDEXEC_NVHPC()
-#  if STDEXEC_NVHPC_VERSION() <= 2403
-#    define STDEXEC_NUMA_VTABLE_INLINE
-#  endif
-#endif
-
-#ifndef STDEXEC_NUMA_VTABLE_INLINE
+// Work around a bug in the NVHPC compilers prior to version 24.3
+#if STDEXEC_NVHPC() && STDEXEC_NVHPC_VERSION < 24'03
+#  define STDEXEC_NUMA_VTABLE_INLINE
+#else
 #  define STDEXEC_NUMA_VTABLE_INLINE inline
 #endif
 

diff --git a/include/exec/task.hpp b/include/exec/task.hpp
@@ -88,8 +88,6 @@ namespace exec {
       __sticky
     };
 
-    struct __parent_promise_t { };
-
     template <__scheduler_affinity _SchedulerAffinity = __scheduler_affinity::__sticky>
     class __default_task_context_impl {
       template <class _ParentPromise>
@@ -103,7 +101,7 @@ namespace exec {
 
      public:
       template <class _ParentPromise>
-      explicit __default_task_context_impl(__parent_promise_t, _ParentPromise& __parent) noexcept {
+      explicit __default_task_context_impl(_ParentPromise& __parent) noexcept {
         if constexpr (_SchedulerAffinity == __scheduler_affinity::__sticky) {
           if constexpr (__check_parent_promise_has_scheduler<_ParentPromise>()) {
             __scheduler_ = get_scheduler(get_env(__parent));
@@ -281,6 +279,25 @@ namespace exec {
       __variant_for<__void, std::exception_ptr> __data_{};
     };
 
+    template <class _Sch>
+    struct __just_void {
+      using sender_concept = sender_t;
+      using completion_signatures = stdexec::completion_signatures<set_value_t()>;
+
+      template <class _Rcvr>
+      [[nodiscard]]
+      static constexpr auto connect(_Rcvr __rcvr) noexcept {
+        return stdexec::connect(just(), static_cast<_Rcvr&&>(__rcvr));
+      }
+
+      [[nodiscard]]
+      constexpr auto get_env() const noexcept {
+        return prop{get_completion_scheduler<set_value_t>, __sch_};
+      }
+
+      _Sch __sch_;
+    };
+
     enum class disposition : unsigned {
       stopped,
       succeeded,
@@ -319,6 +336,8 @@ namespace exec {
       }
 
      private:
+      using __scheduler_t = __query_result_or_t<get_scheduler_t, _Context, inline_scheduler>;
+
       struct __final_awaitable {
         static constexpr auto await_ready() noexcept -> bool {
           return false;
@@ -399,6 +418,11 @@ namespace exec {
         }
 #endif
 
+        template <__sender_adaptor_closure_for<__just_void<__scheduler_t>> _Closure>
+        auto await_transform(_Closure&& __closure) noexcept -> decltype(auto) {
+          return await_transform(static_cast<_Closure&&>(__closure)(__just_void<__scheduler_t>()));
+        }
+
         template <class _Awaitable>
         auto await_transform(_Awaitable&& __awaitable) noexcept -> decltype(auto) {
           return with_awaitable_senders<__promise>::await_transform(
@@ -431,7 +455,7 @@ namespace exec {
         auto await_suspend(__coro::coroutine_handle<_ParentPromise2> __parent) noexcept
           -> __coro::coroutine_handle<> {
           static_assert(__one_of<_ParentPromise, _ParentPromise2, void>);
-          __coro_.promise().__context_.emplace(__parent_promise_t(), __parent.promise());
+          __coro_.promise().__context_.emplace(__parent.promise());
           __context_.emplace(*__coro_.promise().__context_, __parent.promise());
           __coro_.promise().set_continuation(__parent);
           if constexpr (requires { __coro_.promise().stop_requested() ? 0 : 1; }) {

diff --git a/include/exec/timed_scheduler.hpp b/include/exec/timed_scheduler.hpp
@@ -126,7 +126,7 @@ namespace exec {
     };
 
     struct schedule_after_t : __schedule_after_base_t {
-#if !STDEXEC_CLANG() || (__clang_major__ >= 16)
+#if !STDEXEC_CLANG() || STDEXEC_CLANG_VERSION >= 16'00
       using __schedule_after_base_t::operator();
 #else
       // clang prior to 16 is not able to find the correct overload in the
@@ -198,7 +198,7 @@ namespace exec {
     };
 
     struct schedule_at_t : __schedule_at_base_t {
-#if !STDEXEC_CLANG() || (__clang_major__ >= 16)
+#if !STDEXEC_CLANG() || STDEXEC_CLANG_VERSION >= 16'00
       using __schedule_at_base_t::operator();
 #else
       // clang prior to 16 is not able to find the correct overload in the

diff --git a/include/stdexec/__detail/__config.hpp b/include/stdexec/__detail/__config.hpp
@@ -48,28 +48,37 @@
 // macro name; nothing, otherwise.
 #if defined(__NVCC__)
 #  define STDEXEC_NVCC(...) STDEXEC_HEAD_OR_TAIL(1, __VA_ARGS__)
+#  define STDEXEC_NVCC_VERSION (__CUDACC_VER_MAJOR__ * 100 + __CUDACC_VER_MINOR__)
 #elif defined(__EDG__)
 #  define STDEXEC_EDG(...) STDEXEC_HEAD_OR_TAIL(1, __VA_ARGS__)
+#  define STDEXEC_EDG_VERSION __EDG_VERSION__
 #  if defined(__NVCOMPILER)
 #    define STDEXEC_NVHPC(...) STDEXEC_HEAD_OR_TAIL(1, __VA_ARGS__)
+#    define STDEXEC_NVHPC_VERSION (__NVCOMPILER_MAJOR__ * 100 + __NVCOMPILER_MINOR__)
 #  endif
 #  if defined(__INTELLISENSE__)
 #    define STDEXEC_INTELLISENSE(...) STDEXEC_HEAD_OR_TAIL(1, __VA_ARGS__)
 #    define STDEXEC_MSVC_HEADERS(...) STDEXEC_HEAD_OR_TAIL(1, __VA_ARGS__)
 #  endif
 #elif defined(__clang__)
 #  define STDEXEC_CLANG(...) STDEXEC_HEAD_OR_TAIL(1, __VA_ARGS__)
+#  define STDEXEC_CLANG_VERSION (__clang_major__ * 100 + __clang_minor__)
 #  if defined(_MSC_VER)
 #    define STDEXEC_CLANG_CL(...) STDEXEC_HEAD_OR_TAIL(1, __VA_ARGS__)
 #  endif
 #  if defined(__apple_build_version__)
 #    define STDEXEC_APPLE_CLANG(...) STDEXEC_HEAD_OR_TAIL(1, __VA_ARGS__)
+// Apple clang version is encoded as major * 1000000 + minor * 1000 + patch. We ignore the patch
+// version here, as it is not relevant for the purposes of this library.
+#    define STDEXEC_APPLE_CLANG_VERSION (__apple_build_version__ / 1000)
 #  endif
 #elif defined(__GNUC__)
 #  define STDEXEC_GCC(...) STDEXEC_HEAD_OR_TAIL(1, __VA_ARGS__)
+#  define STDEXEC_GCC_VERSION (__GNUC__ * 100 + __GNUC_MINOR__)
 #elif defined(_MSC_VER)
 #  define STDEXEC_MSVC(...)         STDEXEC_HEAD_OR_TAIL(1, __VA_ARGS__)
 #  define STDEXEC_MSVC_HEADERS(...) STDEXEC_HEAD_OR_TAIL(1, __VA_ARGS__)
+#  define STDEXEC_MSVC_VERSION _MSC_VER
 #endif
 
 #ifndef STDEXEC_NVCC
@@ -103,10 +112,6 @@
 #  define STDEXEC_INTELLISENSE(...) STDEXEC_HEAD_OR_NULL(0, __VA_ARGS__)
 #endif
 
-#if STDEXEC_NVHPC()
-#  define STDEXEC_NVHPC_VERSION() (__NVCOMPILER_MAJOR__ * 100 + __NVCOMPILER_MINOR__)
-#endif
-
 ////////////////////////////////////////////////////////////////////////////////////////////////////
 #if defined(__CUDACC__) || STDEXEC_NVHPC()
 #  define STDEXEC_CUDA_COMPILATION(...) STDEXEC_HEAD_OR_TAIL(1, __VA_ARGS__)
@@ -173,13 +178,25 @@ namespace __coro = std::experimental;
 
 #if STDEXEC_NVHPC()
 // NVBUG #4067067: NVHPC does not fully support [[no_unique_address]]
-#  define STDEXEC_ATTR_WHICH_3(_ATTR) /*nothing*/
+#  if STDEXEC_NVHPC_VERSION < 23'05
+#    define STDEXEC_ATTR_WHICH_3(_ATTR) /*nothing*/
+#  else
+#    define STDEXEC_ATTR_WHICH_3(_ATTR) [[no_unique_address]]
+#  endif
+#elif STDEXEC_CLANG_CL()
+// clang-cl does not support [[no_unique_address]]: https://reviews.llvm.org/D110485
+#  if STDEXEC_CLANG_VERSION < 18'01 // TODO: Find the version that started supporting [[msvc::no_unique_address]]
+#    define STDEXEC_ATTR_WHICH_3(_ATTR) /*nothing*/
+#  else
+#    define STDEXEC_ATTR_WHICH_3(_ATTR) [[msvc::no_unique_address]]
+#  endif
 #elif STDEXEC_MSVC()
 // MSVCBUG https://developercommunity.visualstudio.com/t/Incorrect-codegen-when-using-msvc::no_/10452874
-#  define STDEXEC_ATTR_WHICH_3(_ATTR) // [[msvc::no_unique_address]]
-#elif STDEXEC_CLANG_CL()
-// clang-cl does not support: https://reviews.llvm.org/D110485
-#  define STDEXEC_ATTR_WHICH_3(_ATTR) // [[msvc::no_unique_address]]
+#  if STDEXEC_MSVC_VERSION < 19'43
+#    define STDEXEC_ATTR_WHICH_3(_ATTR) /*nothing*/
+#  else
+#    define STDEXEC_ATTR_WHICH_3(_ATTR) [[msvc::no_unique_address]]
+#  endif
 #else
 #  define STDEXEC_ATTR_WHICH_3(_ATTR) [[no_unique_address]]
 #endif
@@ -190,7 +207,7 @@ namespace __coro = std::experimental;
 #elif STDEXEC_CLANG()
 #  define STDEXEC_ATTR_WHICH_4(_ATTR)                                                              \
     __attribute__((__always_inline__, __artificial__, __nodebug__)) inline
-#elif defined(__GNUC__)
+#elif STDEXEC_GCC()
 #  define STDEXEC_ATTR_WHICH_4(_ATTR) __attribute__((__always_inline__, __artificial__)) inline
 #else
 #  define STDEXEC_ATTR_WHICH_4(_ATTR) /*nothing*/
@@ -272,7 +289,7 @@ namespace __coro = std::experimental;
 #  define STDEXEC_IS_TRIVIALLY_COPYABLE(...) std::is_trivially_copyable_v<__VA_ARGS__>
 #endif
 
-#if STDEXEC_HAS_BUILTIN(__is_base_of) || (_MSC_VER >= 1914)
+#if STDEXEC_HAS_BUILTIN(__is_base_of) || (STDEXEC_MSVC_VERSION >= 19'14)
 #  define STDEXEC_IS_BASE_OF(...) __is_base_of(__VA_ARGS__)
 #else
 #  define STDEXEC_IS_BASE_OF(...) std::is_base_of_v<__VA_ARGS__>
@@ -371,17 +388,19 @@ namespace stdexec {
 #endif
 
 // Before gcc-12, gcc really didn't like tuples or variants of immovable types
-#if STDEXEC_GCC() && (__GNUC__ < 12)
+#if STDEXEC_GCC() && (STDEXEC_GCC_VERSION < 12'00)
 #  define STDEXEC_IMMOVABLE(_XP) _XP(_XP&&)
 #else
 #  define STDEXEC_IMMOVABLE(_XP) _XP(_XP&&) = delete
 #endif
 
+#if STDEXEC_GCC()
 // BUG (gcc#98995): copy elision fails when initializing a [[no_unique_address]] field
 // from a function returning an object of class type by value.
-//
 // See: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98995
-#if STDEXEC_GCC()
+#  define STDEXEC_IMMOVABLE_NO_UNIQUE_ADDRESS
+#elif STDEXEC_CLANG() && (__clang_major__ >= 15 && __clang_major__ < 19)
+// See https://github.com/llvm/llvm-project/issues/93563
 #  define STDEXEC_IMMOVABLE_NO_UNIQUE_ADDRESS
 #else
 #  define STDEXEC_IMMOVABLE_NO_UNIQUE_ADDRESS STDEXEC_ATTRIBUTE(no_unique_address)
@@ -401,7 +420,7 @@ namespace stdexec {
 // Some compilers turn on pack indexing in pre-C++26 code. We want to use it if it is
 // available. Pack indexing is disabled for clang < 20 because of:
 // https://github.com/llvm/llvm-project/issues/116105
-#if defined(__cpp_pack_indexing) && !STDEXEC_NVCC() && !(STDEXEC_CLANG() && __clang_major__ < 20)
+#if defined(__cpp_pack_indexing) && !STDEXEC_NVCC() && !(STDEXEC_CLANG() && STDEXEC_CLANG_VERSION < 20'00)
 #  define STDEXEC_HAS_PACK_INDEXING() 1
 #else // ^^^ has pack indexing ^^^ / vvv no pack indexing vvv
 #  define STDEXEC_HAS_PACK_INDEXING() 0
@@ -416,7 +435,7 @@ namespace stdexec {
 // Before clang-16, clang did not like libstdc++'s ranges implementation
 #if __has_include(<ranges>) && \
   (defined(__cpp_lib_ranges) && __cpp_lib_ranges >= 201911L) && \
-  (!STDEXEC_CLANG() || __clang_major__ >= 16 || defined(_LIBCPP_VERSION))
+  (!STDEXEC_CLANG() || STDEXEC_CLANG_VERSION >= 16'00 || defined(_LIBCPP_VERSION))
 #  define STDEXEC_HAS_STD_RANGES() 1
 #else
 #  define STDEXEC_HAS_STD_RANGES() 0
@@ -461,7 +480,7 @@ namespace stdexec {
   }
 
 // GCC 13 implements lexical friendship, but it is incomplete. See https://gcc.gnu.org/bugzilla/show_bug.cgi?id=111018
-#if STDEXEC_CLANG() // || (STDEXEC_GCC() && __GNUC__ >= 13)
+#if STDEXEC_CLANG() // || (STDEXEC_GCC() && STDEXEC_GCC_VERSION >= 13'00)
 #  define STDEXEC_FRIENDSHIP_IS_LEXICAL() 1
 #else
 #  define STDEXEC_FRIENDSHIP_IS_LEXICAL() 0
@@ -473,25 +492,7 @@ namespace stdexec {
 #  define STDEXEC_EXPLICIT_THIS(...) STDEXEC_HEAD_OR_NULL(0, __VA_ARGS__)
 #endif
 
-// Configure extra type checking
-#define STDEXEC_TYPE_CHECKING_ZERO()                                   0
-#define STDEXEC_TYPE_CHECKING_ONE()                                    1
-#define STDEXEC_TYPE_CHECKING_TWO()                                    2
-
-#define STDEXEC_PROBE_TYPE_CHECKING_                                   STDEXEC_TYPE_CHECKING_ONE
-#define STDEXEC_PROBE_TYPE_CHECKING_0                                  STDEXEC_TYPE_CHECKING_ZERO
-#define STDEXEC_PROBE_TYPE_CHECKING_1                                  STDEXEC_TYPE_CHECKING_ONE
-#define STDEXEC_PROBE_TYPE_CHECKING_STDEXEC_ENABLE_EXTRA_TYPE_CHECKING STDEXEC_TYPE_CHECKING_TWO
-
-#define STDEXEC_TYPE_CHECKING_WHICH3(...)                              STDEXEC_PROBE_TYPE_CHECKING_##__VA_ARGS__
-#define STDEXEC_TYPE_CHECKING_WHICH2(...)                              STDEXEC_TYPE_CHECKING_WHICH3(__VA_ARGS__)
-#define STDEXEC_TYPE_CHECKING_WHICH                                    STDEXEC_TYPE_CHECKING_WHICH2(STDEXEC_ENABLE_EXTRA_TYPE_CHECKING)
-
-#ifndef STDEXEC_ENABLE_EXTRA_TYPE_CHECKING
-#  define STDEXEC_ENABLE_EXTRA_TYPE_CHECKING() 0
-#elif STDEXEC_TYPE_CHECKING_WHICH() == 2
-// do nothing
-#elif STDEXEC_TYPE_CHECKING_WHICH() == 0
+#if STDEXEC_ENABLE_EXTRA_TYPE_CHECKING == 0
 #  undef STDEXEC_ENABLE_EXTRA_TYPE_CHECKING
 #  define STDEXEC_ENABLE_EXTRA_TYPE_CHECKING() 0
 #else

diff --git a/include/stdexec/__detail/__connect_awaitable.hpp b/include/stdexec/__detail/__connect_awaitable.hpp
@@ -173,7 +173,7 @@ namespace stdexec {
       }
 
       template <class _Awaitable, class _Receiver>
-#  if STDEXEC_GCC() && (__GNUC__ > 11)
+#  if STDEXEC_GCC() && (STDEXEC_GCC_VERSION >= 12'00)
       __attribute__((__used__))
 #  endif
       static auto __co_impl(_Awaitable __awaitable, _Receiver __rcvr) -> __operation_t<_Receiver> {

diff --git a/include/stdexec/__detail/__cpo.hpp b/include/stdexec/__detail/__cpo.hpp
@@ -82,15 +82,15 @@
 #  pragma deprecated(STDEXEC_CUSTOM)
 #endif
 
-#if STDEXEC_GCC() || (STDEXEC_CLANG() && __clang_major__ < 14)
+#if STDEXEC_GCC() || (STDEXEC_CLANG() && STDEXEC_CLANG_VERSION < 14'00)
 #  define STDEXEC_CUSTOM                                                                           \
     _Pragma("GCC warning \"STDEXEC_CUSTOM is deprecated; use STDEXEC_MEMFN_DECL instead.\"")       \
       STDEXEC_MEMFN_DECL
 #else
 #  define STDEXEC_CUSTOM STDEXEC_MEMFN_DECL
 #endif
 
-#if STDEXEC_CLANG() && __clang_major__ >= 14
+#if STDEXEC_CLANG() && STDEXEC_CLANG_VERSION >= 14'00
 #  pragma clang deprecated(STDEXEC_CUSTOM, "use STDEXEC_MEMFN_DECL instead.")
 #endif