[BEAM] &'static mut references to static variables and LLVM noalias

[japaric]

Please read #111 first, if you haven't already.

Rendered discussion document

[comex]

noalias was originally used to implement C restrict and has roughly equivalent semantics. It basically means (slightly oversimplified): "as long as this pointer is in scope, I promise not to read or write the memory behind this pointer using anything but the pointer itself (or pointers derived from it)".

It's not UB to create a noalias pointer to a global variable. It's only UB if you refer to the original global variable a second time while the pointer is still in scope.

I think the example should also be fine under @RalfJung's proposed "stacked borrows" memory model for Rust, though he's the expert on that.

[RalfJung]

Meta-comment: in terms of process, "active_discussion" is for things in our current "area of discussion". The idea is to keep active discussion focused around one area or else we'll never get anywhere. What you posted as a PR here would, I think, make more sense as the body of an issue.

[RalfJung]

For your actual question, I concur with @comex: &'static mut T means "this is a unique pointer and no other pointer (not derived from this one) will be used to access said memory". It's a bit like Box<T> but without the heap/deallocation part. So, that seems to be exactly what you want.

noalias_pointer was derived from x in the implementation of BumpPointerAlloc, so that seems fine as well.

[RalfJung]

As an example for something that would not be okay here: even just reading from x or X will invalidate overlapping parts of noalias_pointer. So if the implementation of BumpPointerAlloc would do that, that would be a problem.

[RalfJung]

Closing due to inactivity, see #169 for tracking.