Re: atomic operations for shared_ptr ?

[Marc Glisse <marc dot glisse at inria dot fr>]

On Mon, 26 Aug 2013, Florian Weimer wrote:

> On 08/26/2013 01:04 PM, Marc Glisse wrote:
> > On Mon, 26 Aug 2013, Florian Weimer wrote:
> >
> > > On 08/25/2013 01:52 PM, Oleg Endo wrote:
> > >
> > > > Just for my understanding, on targets that use 32 bit pointers and have
> > > > a 64 bit atomic (compare and) exchange, that would be a no-brainer,
> > > > wouldn't it?
> > >
> > > Unfortunately, no.  At least i686 requires 8-byte alignment for the
> > > pair, while we currently only provide 4-byte alignment.  Similarly for
> > > x86_64, where it's 16-byte alignment vs 8-byte.
> >
> > If changing the alignment in the library would help, please make sure
> > there is a PR open with the ABI keyword. There may be some form of ABI
> > change in 4.10, though I don't know if it will be large enough to allow
> > this...
>
> I'm not sure if we can actually reliably allocate objects with 16 byte 
> alignment on the heap (for x86_64, that is).  If not, achieving this 
> alignment would be rather invasive.

On x86_64, 16 bytes is the alignment used by malloc (on all OS I know of), 
so it should work. I believe it is only over-aligned types that we can't 
handle, like SSE (16) on x86 or AVX (32) on x86_64.

> > (I hate binary compatibility taken too far)
>
> This seems to require pretty far-reaching layout changes, and shared_ptr has 
> been around as std::tr1::shared_ptr for a long time.

We don't have the same expectations wrt binary compatibility, that's ok.

--
Marc Glisse