Alignment Woes and False Positives

[cmazakas]

Miri seems to have significant trouble when operating with raw memory and users constructing objects in place.

Simply consider some slightly modified code from the stdlib documentation:
https://play.rust-lang.org/?version=stable&mode=debug&edition=2018&gist=c29624c1b97ae8f0730d5f41816f3143

Miri also seems to needlessly complain about accessing 1 byte aligned data as a type with higher alignment. To me, this is the entire point of what I'm doing.

Is there an idiomatic way of conveying to Miri that I know what I'm doing and it doesn't need to complain (aside from just disabling alignment checking altogether)?

[RalfJung]

The code in that playground has UB, so this is Miri doing its job.

Or is your problem about align_offset (commented out in the playground) not working as it does at run-time? See #1074 for that.

[cmazakas]

What's UB about it?

Fwiw, I ripped that code sample for the stdlib docs so maybe a PR is in order.

[bjorn3]

Let's assume that ptr is 2. Then u16_ptr is 3. This is not aligned to 2 as required by dereferencing a pointer to a u16. If offset was ptr.align_offset(align_of::<u16>()), then the offset would be 0 when the ptr was already correctly aligned for a u16. If ptr wasn't, then offset would be either the minimum value required to correctly align ptr to 2 or usize::max_value(). For miri align_offset() returns usize::max_value() due to #1074.

If you want to read from unaligned memory, you have to use std::ptr::read_unaligned.

[cmazakas]

Oh, that.

Yeah, that's probably UB in most cases. That was the offset I was getting during normal runtime so I just hard-coded it into the playground link. "Real code" would have align_offset just work or manually round up to the next aligned.

Or is your problem about align_offset (commented out in the playground) not working as it does at run-time? See #1074 for that.

I was caught off guard by this, yeah.

[RalfJung]

"Real code" would have align_offset just work or manually round up to the next aligned.

Yeah, Miri's align_offset is strange. It's spec-conforming, but still surprising.

I think resolving #1074 will resolve your confusion though, so closing as a duplicate. Let me know if there's anything beyond #1074 that you ran into here.