In D145183#4456591, @ldionne wrote:

@lewissbaker are you aware of any benchmarks that would be representative of the way stop_token intends to be used?

Sorry for the belated feedback - I only just noticed that the stop_token implementation had been published.

So am I correct in understanding that the main issue with the chicken/egg problem for updating both the compiler to use the new stdlib facilities and updating the stdlib facilities is that we don't want to issue warnings about using <experimental/coroutine> and telling users to use <coroutine> instead if there is no <coroutine> header.

General comment regarding stdlib/compiler compatibility.

There are still room for improvements, in particular, GCC has a 4K frame for this function.

LGTM

Gentle ping.

This updated diff should address most of @CaseyCarter's review comments.

Thanks very much for the detailed review @CaseyCarter! Very much appreciated :)

@EricWF This implementation will currently not work with MSVC as MSVC does not yet support the symmetric-transfer capability added to the coroutines specification in P0913R0.

Added missing 'require coroutines' for experimental/task entry in module.modulemap
Added missing #include in experimental/task header that was failing module builds

Reopening as the commit for this diff was reverted due to it breaking the buildbot module builds.

Fix race condition in __oneshot_event::set() method used by sync_wait().

• Don't use __allocator as variable name (it is #defined to NASTY_MACRO).
• Remove workaround in test for https://bugs.llvm.org/show_bug.cgi?id=37265
• Add support for custom allocators on member functions.

• Make task<T> constructor from coroutine_handle private.
• Run clang-format over <experimental/task>.

Many improvements:

• Remove move-assignment operator
• task<T>::operator co_await() now returns by value when awaiting task<T> rvalue.
• Explicitly scope calls to std-library functions that live in _VSTD.
• Make __task_promise ABI stable under exception/no-exception compilation modes.
• Moved __aligned_allocation_size to <experimental/__memory>
• Make task<T> have a default template parameter of void, allowing task<> to mean task<void>
• Allow co_return { ... } by adding overload of return_value() taking T&& and another taking initializer_list<U>.
• Implement optimisation that avoids storing the allocator in the coroutine frame if allocator_traits<A>::is_always_equal is true and the allocator is default constructible.
• Added unit-tests (currently failing some custom allocator tests under debug builds)

• Explicitly delete __task_promise_base special member functions.
• Declare _task_promise_base::operator delete() as _NOEXCEPT.
• Explicitly scope _VSTD::move() calls for user-defined types.
• Explicitly call global-scope placement operator new for user-defined types.
• Tweak task::operator=() to remove ambiguity with deleted operator=(const task&).

Addressed some feedback from initial version.

• factored out allocator padding calculations
• use anonymous unions instead of char-buffers and reinterpret_cast
• make co_await'ing an invalid task have undefined behaviour
• run clang-format over the file to fix some formatting

I have posted an updated version of the code here: https://gist.github.com/lewissbaker/38ba1d8a13e4fb0906559b7aa1c413d3

The coroutine_handle<noop_coroutine_promise> type does not have a from_address() or a from_promise() static functions in the same way that the coroutine_handle<P> implementation does.
Is this intentional or an oversight in the TS wording?

All looks fine to me.