Re: [Gc] Fwd: Re: AArch64: Shareability domain for dmb st overly conservative (#11)

Hans Boehm <[email protected]> Wed, 5 Nov 2014 10:16:26 -0800
Newsgroups gmane.comp.programming.garbage-collection.boehmgc
Message-ID <CAPUmR1Zv12yzsbmtFWcSAd_vxs4s=jgZWdTeBqpsoyKfVP134w@mail.gmail.com>
I don't think a relaxed fence does anything.

Memory_order_relaxed loads and stores require single-location coherence.
The aarch64 spec guarantees that for ordinary memory operations.  (So does
arm32, though there is a common arm32 bug that violates it.  It is unclear
whether arm32 relaxed loads should be treated as acquire for that reason.
But I don't think that concern applies to aarch64.)

Hans

On Wed, Nov 5, 2014 at 9:06 AM, Andrew Haley <[email protected]> wrote:

> On 11/04/2014 09:28 AM, Yvan Roux wrote:
> > To give a full picture of the __atomic_thread_fence builtin for
> > AArch64, here is the generated code (by trunk) for each model:
> >
> > __ATOMIC_RELAXED -> nothing
>
> I'm pretty sure that is wrong.  ATOMIC_RELAXED does imply single-
> word coherence, AIUI.
>
> Andrew.
>
>

_______________________________________________
bdwgc mailing list
[email protected]
https://lists.opendylan.org/mailman/listinfo/bdwgc