On 7/5/23 03:05, Boqun Feng wrote:
On Mon, Jul 03, 2023 at 03:20:31PM -0400, Olivier Dion wrote:
[...]
NOTE: On x86-64, we found at least one corner case [7] with Clang where
a RELEASE exchange is optimized to a RELEASE store, when the returned
value of the exchange is unused, breaking the above expectations.
Although this type of optimization respect the standard "as-if"
statement, we question its pertinence since a user should simply do a
RELEASE store instead of an exchange in that case. With the
introduction of these new primitives, these type of optimizations should
be revisited.
FWIW, this is actually a LLVM bug:
https://github.com/llvm/llvm-project/issues/60418
So it was more than a dubious optimization, it's actually broken as well.
I am worried about adding to the compiler's ability to optimize those
atomics because of the subtle corner-cases/bugs that can creep up.
Thanks,
Mathieu
--
Mathieu Desnoyers
EfficiOS Inc.
https://www.efficios.com
