bjorn3
commented
1 year ago How would the right alignment to use for loads and stored in backends be determined if the requirement for places to be aligned is dropped? Would it for example the alignment requirement instead apply to typed copies of the value stored behind the place instead of the place itself?
RalfJung
Currently, places (the things place expressions evaluate to, which only exist as intermediate concept during the execution of a statement) have an effective validity invariant: they are dereferenceable and aligned (but their contents can be anything). We do not distinguish "safe" and "unsafe" places or anything like that, this applies to all of them.
That's why
addr_of!(*ptr)is UB if the pointer dangles or is unaligned.There's not a ton of strong motivation for this, so -- is that really the semantics we want? https://github.com/rust-lang/unsafe-code-guidelines/issues/319 discusses this to some extend but then diverged a lot into surface language design questions that boil down to "how can we avoid even using place expressions for raw pointer arithmetic".
The design space I see here has several dimensions:
I don't see a good motivation for the 2nd point: even if we don't impose alignment/dereferenceability on places, we still impose it on references, so we know that places created via
*referenceare aligned and dereferenceable.My inclination would hence be to simply drop the requirement entirely. This does however open a new question which is currently moot: when doing a place projection, which is basically an offset operation, what are the rules for that? The obvious answer is to say "same as
ptr.offset", meaning the question is tracked in #350 and https://github.com/rust-lang/unsafe-code-guidelines/issues/93, but in principle we could also say "same asptr.wrapping_offset". The latter would have the advantage of making place expressions based on raw pointers entirely safe, only the place-to-value coercion would have to be unsafe.