Did they really get such a performance boost (x6) over standard coroutine implementation, or was it because of the single-threaded implementation?

I’ve carefully analyzed these benchmarks. The first improvement is achieved by narrowing responsibility scope: the channel becomes single-producer/single consumer,

close

and

select

support is dropped. In general, it’s a good idea to write very specialized solutions for high-performance code, but such solutions are rarely useful outside of specific use-case. x6 boost is achieved only because solution became single-threaded. Note: not SP/SC, but single-threaded, producer and consumer should be executed on the same thread. No

volatile

, no barriers, no memory synchronization.

That's what I expected. Thanks!