llvm dev - Feb 2021 - [Coroutine] Should we normalize the ramp function for whose initial_suspend isn't always_suspend?

Hi all,

I find that the ramp function of coroutine whose initial_suspend is
always_suspend (Let me call it "init-suspend-coroutine")
would be very simple. Normally it would only contain one basic block. And all of
the codes in the coroutine body would be
in the resume function of coroutine.
For the ramp function of coroutine whose initial_suspend isn't
always_suspend (Let me call it "init-no-suspend-coroutine"),
it may be large and it may contain the codes in the coroutine body. It may cause
the size of codes be multiple of the original
size. And the large ramp function is harder to inline and harder to be elided.

I want to make the ramp function for init-no-suspend-coroutine keep the same
formal with the init-suspend-coroutine.
My plan is to insert a fake suspend in the front of initial_suspend in the
CodeGen phase of clang and in the llvm part, we can
insert a tail call to the resume function so the control flow would remain with
the original.

Does any one have any suggestion?

Thanks,
Chuanqi
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20210223/ad1d5f66/attachment.html>