Fix misaligned transmute lint.

[devonhollowood]

This clears up some ambiguous wording within the lint's descriptions,
and makes it so that the lint only applies to pointers. This should fix
a number of false positives (when the lint was applied to non-pointer
types) as well as false negatives (when the lint was applied to two
pointers of the same alignment, but whose pointed-to types had differing
alignments). See discussion at #2400.

[oli-obk]

I don't think we should go ahead with this. We should probably remove the lint entirely and create a lint for transmuting references to references, because one should be using raw ptr casts in that case.

[devonhollowood]

This lint also checks raw ptr -> raw ptr casting for alignment issues--are you saying that we should remove that part too, or just the part of the lint that checks ref -> ref casting?

[oli-obk]

No I meant we should forbid transmuting references to references or pointers to pointers and suggest casts instead. And then we can detect casts that go from low to high alignment

[devonhollowood]

Ah, understood. That seems like a pretty reasonable approach to me as well.

[oli-obk]

@devonhollowood do you want to tackle that change?

[devonhollowood]

@oli-obk Yes, although I probably won't get a chance for another couple weeks, so if someone else wants to jump in and make the change they should feel free to do so.

[devonhollowood]

@oli-obk I got some free time so I started working on this tonight, and I was wondering if you could clarify something you said earlier:

I meant we should forbid transmuting references to references or pointers to pointers and suggest casts instead.

As far as I know you can't cast to a ref with as. So in the case that you want to (say) turn an &f32 into a &u32 I think you need to use transmute. Should we still forbid this? If so, what should we recommend instead?

---

Also, what is the policy on deprecating old lints? Since it sounds like you want to deprecate misaligned-transmute. Should I just make the lint a no-op and mark it as deprecated in the docs (with an explanation)? Or should I remove it entirely (which I think is a breaking change for code with #[allow(misaliged_transmute)] etc)?

[oli-obk]

So in the case that you want to (say) turn an &f32 into a &u32 I think you need to use transmute. Should we still forbid this? If so, what should we recommend instead?

You can do &*(&some_u32 as *const u32 as *const f32) which conveys the horror of it much better than transmute(&some_u32)

Also, what is the policy on deprecating old lints? Since it sounds like you want to deprecate misaligned-transmute. Should I just make the lint a no-op and mark it as deprecated in the docs (with an explanation)?

We have a collection of deprecated lints in https://github.com/rust-lang-nursery/rust-clippy/blob/master/clippy_lints/src/deprecated_lints.rs