I think you're misunderstanding the benefit of stackless coroutines. Here's the core of the issue: Stacks, by definition, are dynamically growable objects. But sometimes (in most cases?) the state that needs to be stored across yield points can be statically bounded up front. In that case, stackful coroutines (really, threads) add unnecessary overhead, because you pay for that ability to grow. Golang, for instance, has to start stacks small and grow them by copying, because it can't reliably statically analyze the goroutine to determine the needed stack space. A properly-implemented stackless coroutine system (or even the classic C approach of writing state machines by hand), by contrast, can determine the amount of space necessary for each coroutine and allocate it precisely. There are even slab allocation tricks for many coroutines of the same size. Fundamentally, this is why manually-written C code using epoll can outperform any stackful coroutine system: optimal allocation of stackful coroutines in a language with recursion and first-class functions requires higher-order control flow analysis, which is generally considered infeasible.
Of course, I've only focused on the performance issues here, and obviously there are also important issues of ergonomics at play. But my point is that there are benefits to stackless coroutines that can't be simply dismissed.
Comments
I think you're misunderstanding the benefit of stackless coroutines. Here's the core of the issue: Stacks, by definition, are dynamically growable objects. But sometimes (in most cases?) the state that needs to be stored across yield points can be statically bounded up front. In that case, stackful coroutines (really, threads) add unnecessary overhead, because you pay for that ability to grow. Golang, for instance, has to start stacks small and grow them by copying, because it can't reliably statically analyze the goroutine to determine the needed stack space. A properly-implemented stackless coroutine system (or even the classic C approach of writing state machines by hand), by contrast, can determine the amount of space necessary for each coroutine and allocate it precisely. There are even slab allocation tricks for many coroutines of the same size. Fundamentally, this is why manually-written C code using epoll can outperform any stackful coroutine system: optimal allocation of stackful coroutines in a language with recursion and first-class functions requires higher-order control flow analysis, which is generally considered infeasible.
Of course, I've only focused on the performance issues here, and obviously there are also important issues of ergonomics at play. But my point is that there are benefits to stackless coroutines that can't be simply dismissed.