On Wed, Oct 14, 2015 at 9:27 PM, Rich Felker <dalias@libc.org> wrote:

> On Wed, Oct 14, 2015 at 09:23:59PM +0200, Alex wrote:
> > On Wed, Oct 14, 2015 at 9:14 PM, Rich Felker <dalias@libc.org> wrote:
> >
> > > On Wed, Oct 14, 2015 at 12:21:05PM +0200, Alex wrote:
> > > > This has been an interesting exercise so far. Is there any other arch
> > > which
> > > > you think it would be worthwhile to develop a CFI generation script
> for?
> > > It
> > > > should be something which has enough users to avoid problems with
> bitrot.
> > >
> > > CFI is probably a lot less interesting on archs where you have a
> > > plenty registers not to need to manipulate stack frames in asm
> > > functions, since in that case the debugger mostly works fine without
> > > CFI. I don't know right off which of the other archs have significant
> > > amounts of asm that adjusts the stack pointer, but you could go
> > > through and check them. Having ABI info for them all would be helpful;
> > > I'm pasting my draft ABI reference (which might have errors) below.
> >
> > Fair enough. If it's not likely to help anyone, I'll leave the CFI
> > generation here.
> >
> > Another idea: are you interested in an implementation of POSIX AIO which
> > uses the native AIO syscalls? Bad idea?
>
> Those syscalls have nothing at all to do with POSIX AIO. They're
> completely different. :(


The interface presented by the raw syscalls is not the POSIX AIO interface,
but I haven't seen any reason yet why io_setup, io_submit, io_getevents,
etc. couldn't be used to implement the POSIX AIO interface. Is there such a
reason?

The payoff would be that you wouldn't have to spawn a thread for each AIO
operation. Instead, you could potentially spawn just one extra thread,
which would run in a loop, calling io_getevents() and doing some
recordkeeping each time an AIO operation completed.

AD