On 01/13/2015 09:59 AM, Richard Biener wrote:
On Tue, Jan 13, 2015 at 3:56 PM, Andrew MacLeod <amacl...@redhat.com> wrote:
Lengthy discussion : https://gcc.gnu.org/bugzilla/show_bug.cgi?id=59448
Basically we can generate incorrect code for an atomic consume operation in
some circumstances. The general feeling seems to be that we should simply
promote all consume operations to an acquire operation until there is a
better definition/understanding of the consume model and how GCC can track
it.
I proposed a simple patch in the PR, and I have not seen or heard of any
dissenting opinion. We should get this in before the end of stage 3 I
think.
The problem with the patch in the PR is the memory model is immediately
promoted from consume to acquire. This happens *before* any of the
memmodel checks are made. If a consume is illegally specified (such as in a
compare_exchange), it gets promoted to acquire and the compiler doesn't
report the error because it never sees the consume.
This new patch simply makes the adjustment after any errors are checked on
the originally specified model. It bootstraps on x86_64-unknown-linux-gnu
and passes all regression testing.
I also built an aarch64 compiler and it appears to issue the LDAR as
specified in the PR, but anyone with a vested interest really ought to check
it out with a real build to be sure.
OK for trunk?
Why not patch get_memmodel? (not sure if that catches all cases)
Richard.
That was the original patch.
The issue is that it promotes consume to acquire before any error
checking gets to look at the model, so then we allow illegal
specification of consume. (It actually triggers a failure in the testsuite)
Andrew
