update the start scheduler before connecting scheduler sender in `continues_on` by ericniebler · Pull Request #2277 · NVIDIA/stdexec

ericniebler · 2026-09-20T01:49:37Z

copy-pr-bot · 2026-09-20T01:49:40Z

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

maartenarnst · 2026-09-20T07:44:01Z

+      template <class _Env>
+      using __env2_t = __secondary_env_t<_Sender, _Env, set_value_t>;
+
+      template <class _Env>
+      constexpr auto __mk_env2(_Env&& __env) const noexcept -> __env2_t<_Env>
+      {
+        return __mk_secondary_env_t<set_value_t>()(__sndr_, static_cast<_Env&&>(__env));
+      }


Noting that the tag is hardcoded here to set_value_t, whereas the call sites are templated on _SetTag? Could this lead to issues e.g. when the predecessor completes with error/stop?

I think this issue is still open and may be relevant.

The __complete lambda stores all completions with their tag, along with a comment that states that the goal is to forward them from within the scheduler's context.

https://github.com/ericniebler/stdexec/blob/233768b59e2d92c65c6652a6d18864db15e8f1ac/include/stdexec/__detail/__continues_on.hpp#L362-L369

This is indeed what __receiver2 does:

https://github.com/ericniebler/stdexec/blob/233768b59e2d92c65c6652a6d18864db15e8f1ac/include/stdexec/__detail/__continues_on.hpp#L79-L82

If the predecessor completes with an error, we can thus expect that error to be re-emitted from the scheduler's context, and it seems that's a case that the queries don't capture yet.

Incidentally, the first sentence of the docstring of __receiver2 appears too strong because its set_error and set_stopped would be invoked from the predecessor's context.

maartenarnst · 2026-09-20T07:56:23Z

+    auto sndr = ex::schedule(pool.get_scheduler())  //
+              | ::then_store_thread_id(pool_tid)    //
+              | ex::let_value(
+                  [&]() noexcept


Just curious why you added noexcept?

ericniebler · 2026-09-20T17:44:14Z

/ok to test d5e63d7

ericniebler · 2026-09-21T16:03:01Z

/ok to test 4c50519

ericniebler · 2026-09-21T18:58:20Z

/ok to test 3e2bddf

ericniebler · 2026-09-21T19:39:12Z

/ok to test 660d14d

…inues_on fixes NVIDIA#2268

Co-authored-by: Maarten Arnst <maarten.arnst@ulg.ac.be>

ericniebler · 2026-10-05T13:51:06Z

/ok to test c127035

…2294) * fix(exec): make write_env transparent to sequence sender semantics A write_env sender wrapping a sequence sender was not itself a sequence sender: get_item_types has no branch that looks through stdexec sender adaptors, so sequence-aware algorithms like exec::transform_each treated the wrapper as a sequence of one and collapsed the item types. Make write_env senders transparent to sequence semantics: - specialize enable_sequence_sender for __write_env_t sexprs - compute item types of a write_env sender from its child in the joined environment (recursing through stacked wrappers) - subscribe to the child through a receiver that joins the written environment with the wrapped receiver's environment First half of #2053. The starts_on half goes through the continues_on machinery that #2277 is reworking, so it is left for a follow-up. * fix(exec): make sequence-adaptor transparency an open customization point Review feedback on #2294: get_item_types and subscribe must not hard-code individual sender adaptors, since the set of adaptors is open. Replace the write_env-specific branches with a generic mechanism driven by a new __sequence_adaptor_traits customization point: an adaptor that forwards set_next and all completions to a single child declares itself transparent for its tag, optionally naming the environment its child observes. write_env now only specializes that trait, and the sequence machinery handles every transparent adaptor the same way. The enable_sequence_sender specialization also goes through tag_of_t rather than digging into the sender's descriptor. Also merge upstream/main, and extend the regression tests with a user-defined adaptor that opts in via the new trait, to show that the extension point is open. * test: fix clang-format-21 violations in test_write_env_sequence.cpp Empty struct bodies written as `{ };` fail the style CI job, which runs clang-format-21 --dry-run --Werror over all tracked sources. 🤖 Generated with Codebuff Co-Authored-By: Codebuff <noreply@codebuff.com> --------- Co-authored-by: alwaysprince05 <alwaysprince05@users.noreply.github.com> Co-authored-by: Codebuff <noreply@codebuff.com>

ericniebler mentioned this pull request Sep 20, 2026

continues_on: hop opstate not connected with secondary environment #2268

Open

ericniebler force-pushed the fix-continues-on-attributes branch 2 times, most recently from c79da61 to cd3d86b Compare September 20, 2026 01:56