[PATCH][AArch64] Testcase fix for __ATOMIC_CONSUME

[James Greenhalgh]

On Wed, Jan 28, 2015 at 05:51:27PM +0000, Marcus Shawcroft wrote:
> On 28 January 2015 at 17:41, Mike Stump <mikestump@comcast.net> wrote:
> > On Jan 27, 2015, at 8:24 AM, Alex Velenko <Alex.Velenko@arm.com> wrote:
> >> This patch fixes aarch64/atomic-op-consume.c test to expect safe "LDAXR"
> >> instruction to be generated when __ATOMIC_CONSUME semantics is requested.
> >
> > Did you see:
> >
> >   /* Workaround for Bugzilla 59448. GCC doesn't track consume properly, so
> >      be conservative and promote consume to acquire.  */
> >   if (val == MEMMODEL_CONSUME)
> >     val = MEMMODEL_ACQUIRE;
> >
> > in builtins.c?  Feels like if gcc isn’t going to support it for you, then testing for it would be, hard?
> 
> The original test was written pre 59448 and expects GCC to implement
> consume behaviour.  The workaround for 59448 changes GCC behaviour but
> did not update this test case.  Going forward we can either remove the
> test case completely, xfail the test case pending a proper solution to
> 59448 ? or change the test case to expect the current intended
> behaviour of gcc.  This patch implements that latter, which seems
> reasonable to me. Mike do you prefer one of the other two approaches ?

I'll vote for keeping the test case, please. It still fulfills a useful
purpose.

GCC must now promote __ATOMIC_CONSUME to __ATOMIC_ACQUIRE, and the test
case now reflects that - expecting the LDAXR instruction (load acquire
exclusive) over the more relaxed LDXR (load exclusive) that we emitted
prior to PR59448.

If we lose that promotion, or if we start emitting different instructions
for __ATOMIC_ACQUIRE, we have a regression, and this test should spot
that. I also wouldn't want to see the test XFAIL; we know what the correct
expected behaviour is and should update the test to reflect that - as in
Alex' patch.

Thanks,
James