[lttng-dev] [PATCH 02/11] urcu/uatomic: Use atomic builtins if configured

[Paul E. McKenney]

On Thu, Jun 22, 2023 at 03:53:33PM -0400, Olivier Dion wrote:
> On Thu, 22 Jun 2023, "Paul E. McKenney" <paulmck at kernel.org> wrote:
> 
> > I suggest C11 volatile atomic load/store.  Load/store fusing is permitted
> > for non-volatile atomic loads and stores, and such fusing can ruin your
> > code's entire day.  ;-)
> 
> Good catch.  Seems like not a problem on GCC (yet), but Clang is extremely
> aggressive and seems to do store fusing on some corner cases [0].
> 
> However, I do not find any simple reproducer of load/store fusing.  Do
> you have example of such fusing, or is this a precaution?  In the
> meantime, back to reading the standard to be certain :-)
> 
>  [0] https://godbolt.org/z/odKG9a75a

I certainly have heard a number of compiler writers thinking in terms
of doing load/store fusing, some of whom were trying to get rid of the
volatile variants in order to remove an impediment to their mission of
optimizing all programs out of existence.  ;-)

I therefore suggest taking this possibility quite seriously.

							Thanx, Paul