Re: [RFA] choosing __platform_wait_t on targets without lock-free 64 atomics

[Thomas Rodgers <trodgers@redhat.com>]

[-- Attachment #1: Type: text/plain, Size: 5467 bytes --]

On Sun, Jan 1, 2023 at 11:47 PM Iain Sandoe <iain@sandoe.co.uk> wrote:

>
>
> > On 2 Jan 2023, at 00:53, Thomas Rodgers <trodgers@redhat.com> wrote:
> >
> > __platform_wait_t should be whatever the platform supports lock free
> natively. The use of a 64 bit int there in the fall through was copied from
> Olivier's original implementation for libc++, which uses __ulock_wait/wake
> on Darwin which takes a unit64_t, because I had intended to add support
> Darwin, but haven't done so yet.
>
> In general, libc++ supports fewer versions of Darwin than GCC does (I
> don’t know right now which versions/archs we support have the
> __ulock_wait/wake).  Of course, I would very much like to see an efficient
> solution for Darwin - so please let me know/share patches with me when you
> get to it - I can test on older supported versions.
>

I *think* it's supported for 10.12 onward, for anything older the
mutex/condvar implementation would have to be used.


>
> However, the issue here is not really Darwin-specific - it will effect any
> target that does not have either a futex or a 64b lock-free atomic.
>
>
I agree. I was just relating the back-story of why it was 64b in the
not-Linux case.


> thanks
> Iain
>
> >
> > On Fri, Dec 30, 2022 at 2:51 AM Iain Sandoe <iain@sandoe.co.uk> wrote:
> >
> >
> > > On 29 Dec 2022, at 17:02, Jonathan Wakely <jwakely.gcc@gmail.com>
> wrote:
> > >
> > >
> > >
> > > On Thu, 29 Dec 2022, 16:22 Iain Sandoe, <iain@sandoe.co.uk> wrote:
> > >
> > >
> > > > On 29 Dec 2022, at 15:44, Jonathan Wakely <jwakely.gcc@gmail.com>
> wrote:
> > > >
> > > >
> > > >
> > > > On Thu, 29 Dec 2022, 15:30 Iain Sandoe, <iain@sandoe.co.uk> wrote:
> > > > Hi Folks,
> > > >
> > > > > On 29 Dec 2022, at 12:09, Jonathan Wakely <jwakely.gcc@gmail.com>
> wrote:
> > > >
> > > > > On Thu, 29 Dec 2022, 11:29 Iain Sandoe, <iain@sandoe.co.uk> wrote:
> > > >
> > > > >> The recent addition of the tz handling has pulled in a dependency
> on </bits/atomic_wait.h>
> > > > >>
> > > > >> This currently specifies __platform_wait_t as a 64bit quatity on
> platforms without _GLIBCXX_HAVE_LINUX_FUTEX.
> > > > >>
> > > > >> PowerPC does not have a 64b atomic without library support - so
> that this causes a bootstrap
> > > > >> fail on powerpc-darwin (and I guess any other 32b powerpc
> non-futex target).
> > > > >>
> > > > >> Rather than contrive to build and add libatomic (which is not at
> present available at the point
> > > > >> that libstdc++ is built), I wonder if there is any specific
> reason that __platform_wait_t needs
> > > > >> to be 64 bits on these platforms? (Especially since the futex
> case uses an int.)
> > > > >>
> > > > > I think we do want the generic case's _M_wait atomic variable to
> be lock free, otherwise we use two locks for every operation, the one in
> libatomic and the waiter mutex. That's more important than it being any
> specific width.
> > > >
> > > > Definitely, that’s probably a recipe for some subtle race condition
> .. nested locks etc.
> > > >
> > > > I didn't see any nested cases from a quick look, but it would still
> be better to avoid two locks.
> > > >
> > > >
> > > > >> Advice on the right way to fix this welcome — as a work-around to
> allow bootstrap to complete
> > > > >> I applied the patch below - but that seems unlikely to be the
> right thing generically .
> > > > >>
> > > > > Rather than __lp64__ I think we should check the
> ATOMIC_LONG_LOCK_FREE macro and use long if it's lock free and int
> otherwise. But Tom needs to confirm that. That would be approximately the
> same as your patch in practice.
> > > >
> > > > OK.. that makes sense here’s a proposed patch (pending subsequent
> input from Tom).
> > > >
> > > > I am using “lock free always” as the criterion, “sometimes” does not
> seem useful here.
> > > >
> > > > Agreed.
> > > >
> > > >
> > > > Although we normally build libstdc++ with the just-built GCC...
> > > > .. AFAIK the __SIZEOF_* are available from any version of GCC or
> clang that would
> > > > be capable of building the sources.
> > > >
> > > > Yep, but do we need the size checks at all?
> > > >
> > > > I was thinking we could just use 'unsigned long' or 'unsigned int'
> directly, instead of a uintN_t typedef. Using the typedef just seems to
> complicate things.
> > >
> > > That’s fine by me - I was just copying what was there :)
> > >
> > > In this patch I made it so that a target without a ‘suitable'
> lock-free size would fail to
> > > compile the source, which seems better than a link fail later — I
> could make it more
> > > specific (e.g. # fail clause) or we could test for smaller lock-free
> entities…
> > >
> > > I think we can just eschew atomics altogether in that case, and just
> use the mutex for all accesses. I can do that after the break when I'm back
> online.
> >
> > Great, thanks!
> > cheers
> > Iain
> >
> > I’m using this locally in the meantime:
> >
> > # if  ATOMIC_LONG_LOCK_FREE == 2
> >     using __platform_wait_t = long;
> > # elif ATOMIC_INT_LOCK_FREE == 2
> >     using __platform_wait_t = int;
> > # elif ATOMIC_SHORT_LOCK_FREE == 2
> >     using __platform_wait_t = short;
> > # elif ATOMIC_CHAR_LOCK_FREE == 2
> >     using __platform_wait_t = char;
> > # else
> > # fail No suitable lock-free type found.
> > # endif
> >
>
>