[llvm] [FunctionAttrs][IR] Fix memory attr inference for volatile mem intrinsics (PR #122926)

[Nikita Popov via llvm-commits]

nikic wrote:

Yes, our modelling of volatile memory intrinsics is generally broken -- but I don't think there's a lot of value in changing the attributes on the declaration. Ultimately we'll still have to fix things up, just in the opposite direction of what we do now.

If we really want to improve things, we should split off separate llvm.memset.volatile etc intrinsics. That way everything gets correct memory modelling and all optimizations will implicitly ignore the volatile variants by default.

https://github.com/llvm/llvm-project/pull/122926