Hi Richard and Philippe,
I was finally able to test the patch.
As expected, the fences generated by TCG with your patch are the same
ones as in mine.
However, I was not able to reproduce the failure with the ahci-test on
my ARM system (2x 28-core Thunder X2, 4 threads per core). I ran 500
tests
Hi Richard,
On 1/5/22 01:45, Richard Henderson wrote:
When st:ld is not required by the guest but ld:st is, we can
put ld:ld+ld:st barriers after loads, and then st:st barriers
before stores to enforce all required barriers.
The st:st barrier is often special cased by hosts, and that
is expecte
When st:ld is not required by the guest but ld:st is, we can
put ld:ld+ld:st barriers after loads, and then st:st barriers
before stores to enforce all required barriers.
The st:st barrier is often special cased by hosts, and that
is expected to be more efficient than a full barrier.
Signed-off-b
