r? @withoutboats

(rust_highfive has picked a reviewer for you, use r? to override)

r? @alexcrichton

I'm not sure we should land these without first establishing a more formal memory model. In particular, afaict a types-as-contracts-based interpretation would make copy_opaque useless, since the value behind the reference should not have mutated during the duration of the borrow.

@cramerj I see &T as guaranteeing that nothing in the same program has modified the value; that doesn't necessarily guarantee that the value wasn't changed. Plus, types like Cell break this guarantee through an API not unlike copy_opaque.

@clarcharr Cell and other !Freeze types are a special exception to this rule that the compiler understands.

I share the doubts about how useful a volatile load through a safe reference is. I don't believe it's likely to be miscompiled -- in practice LLVM will probably let volatile take precedence over any other information we give it, and there might be some way to semantically make sense of volatile accesses to UnsafeCell memory. But without an UnsafeCell involved it's inconsistent to create a &T reference (which implies no mutation of the referent for some duration) despite the referent beig mutated from some other source. I can think of a few other use cases and exculpatory circumstances that make this not UB and just a very hairy affair, but I believe the danger of wrongly implying safe references without interior mutability can be combined with "external" mutation outweight any (dubious, IMO) safety benefit for the niche use cases.

I also have some doubts about the robustness of the functions that are intended to partially replace black_box -- I haven't been able to formulate them precisely, but basically a volatile operation has much less power as an optimization barrier than inline asm with side effects, so it seems much trickier to ensure that it's actually doing its job.

In terms of review I'd personally like to see these functions used to see whether in practice this works or not first as well.

It feels like there's a moderate amount of response to not accept this PR; am I reading the comments correct? Is the team leaning towards closing this PR, perhaps in favor of discussion on IRLO?

@cramertj @alexcrichton @rkruppe Should this PR be closed in favor of an RFC? The author has offered to draft one.

Yeah let's close this for now, I think this may want to have more discussion on an RFC or perhaps on internals before we add to libstd. In the meantime as well this can certainly be added as a crate on crates.io!