The offset_of! macro may need to be a compiler intrinsic

[alexcrichton]

This came up a few weeks ago in libs triage specifically related to #48493 where we were wondering if there were other locations in the standard library that needed to be adjusted to handle uninhabited enums being handled in generic locations. While not many locations in the standard library leapt to mind one thing we thought of was the offset_of! macro.

For example if we've got something like:

struct Foo<T> { a: T, b: usize }

do existing offset_of! macros handle this?

offset_of!(Foo<Void>, b)

or is there "technical UB that rustc may paper over today" somewhere in there?

cc @dtolnay, I think you had better ideas about examples here than I!

[kennytm]

Existing crates implementing offset_of:

• https://crates.io/crates/memoffset — Implemented using mem::uninitialized(). Result: returns 0.

• https://crates.io/crates/field-offset — Implemented using mem::zeroed(). Result: panic from its own sanity check.

  thread 'main' panicked at 'assertion failed: offset + mem::size_of::<U>() <= mem::size_of::<T>()'
  

(Also, there should be an offset_of_val! in additional to offset_of! due to DST field.)

[oli-obk]

Why do we even need a memory location for that? Can't we just create a raw pointer from the ether and operate on that?

https://play.rust-lang.org/?gist=79229d888ce819235d5350241a037a78&version=stable

[alexcrichton]

Perhaps! I recall some discussion in the past about how valid the temporary reference is there, but I'd naively assume that such a construction was reasonable!

[steveklabnik]

I guess @rust-lang/lang would have to answer this, but I think the answer is "yes". We've had some things around this area like https://internals.rust-lang.org/t/pre-rfc-add-a-new-offset-of-macro-to-core-mem/9273

[Centril]

(It's currently not possible to implement offset_of! in a sound manner with Rust. The memoffset crate, while being unsound, has been improved to make the UB less exploitable.)

cc #64490
cc @RalfJung

[est31]

Given that we have had rust-lang/rfcs#3308 and now have the offset_of! macro (#106655), I think this can be closed.