[clang] await_suspend() -> std::noop_coroutine() = missed optimization

[tzcnt]

When returning std::noop_coroutine() from await_suspend, this always seems to result in:

lea          std::__n4861::coroutine_handle<std::__n4861::noop_coroutine_promise,%rcx
jmp          *(%rcx)

This delegates to a function that is simply ret (<std::__n4861::coroutine_handlestd::__n4861::noop_coroutine_promise::__frame::__dummy_resume_destroy()>)

This seems like a missed optimization. Can we simply inline the ret and skip the indirect jump?

[llvmbot]

@llvm/issue-subscribers-coroutines

Author: None (tzcnt)

When returning `std::noop_coroutine()` from `await_suspend`, this always seems to result in: ``` lea std::__n4861::coroutine_handle<std::__n4861::noop_coroutine_promise,%rcx jmp *(%rcx) ```

This delegates to a function that is simply ret (<std::__n4861::coroutine_handle<std::__n4861::noop_coroutine_promise>::__frame::__dummy_resume_destroy()>)

This seems like a missed optimization. Can we simply inline the ret and skip the indirect jump?

[ChuanqiXu9]

Yeah, it sounds like a missed optimization. Currently the noop_coroutine was partially implemented in libraries and the middle end. But the compiler didn't make treat noop_coroutine specially. The compiler just treat noop_coroutine as any other coroutine. But actually noop_coroutine is implemented in the compiler. I think it makes sense to identify noop_coroutine specially.