llvm dev - Jun 2019 - Is it possible to mark llvm.masked.store as volatile or to achieve similar functionality?

I'm using llvm.masked.store in our compiler. However, as this is an
intrinsic function, I cannot mark it as volatile like I could with an ordinary
store.

I don't really care if the masked store is broken down into multiple stores
or the ordering of those, I just effectively want those stores to be volatile
w.r.t other volatile stores (and loads) before and after.

Is there any easy way to either mark them as volatile, or add fences around the
intrinsic call  which force similar functionality?

I was thinking of something like adding fences
with syncscope of "singlethreaded" to achieve the same things as
volatile but I can't work out whether that would work.

Code in question which doesn't let me include a volatile parameter:

bldr.CreateMaskedStore(value, addrVal, align, mask);

thanks

Mike Vine

Hi Mike,

On Tue, 4 Jun 2019 at 10:40, Mike Vine (EMUL8R) via llvm-dev
<llvm-dev at lists.llvm.org> wrote:
> I was thinking of something like adding fences with syncscope of
"singlethreaded" to achieve the same things as volatile but I
can't work out whether that would work.
Nothing will convert an access to volatile, even ignoring the
properties you mention you don't care about (they're actually not even
guaranteed by volatile). The implementation might still duplicate or
eliminate the stores for example, which would never happen with
volatile.

But if you're actually trying to enforce ordering (and not with
memory-mapped I/O registers on the host) then volatile is probably the
wrong tool for the job anyway, and fences have a pretty good chance of
being what you should be doing.

But exactly what fences you need will depend on more details of your
use-case. Specifically, on what other resources (CPU/GPU/...) might be
able to see that the operation happened out of order.

Cheers.

Tim.