relax lifetime constraints on AtomicRef

[SrTobi]

fixes and tests the problem described in #24.

I had to set editon to 2018, to make the async test work. If this is undesirable, I can move the test to a different crate.

I'm reasonable confident, that the changes are correct, because the compiler would generate even less restrictive Sync/Send implementation in a previous version.

impl<'b, T: ?Sized> Send for AtomicRefMut<'b, T> where
    T: Send

On the other hand, I have to admit, that I'm not really understanding why this change fixes #24...
pub struct AtomicRef<'b, T: ?Sized + 'b> has T: 'b after all.

Maybe @Imberflur can weigh in?

[Imberflur]

My best guess is that #24 is caused by rust-lang/rust#110338 and the explicit T: 'b on the Send impl adds some constraint that the compiler currently can't solve in this context.

I used the for<'a>: &... hrtb bounds here since it is less error prone than manually writing out the required bounds on T, they should be equivalent, and I couldn't find any cases where they caused issues. However, given that there are cases where they are confusing the compiler, I think it would be best to write the required bounds for T directly.

I think these are the correct bounds, but you can double check them against the previously generated bounds:
T: Sync + Send -> AtomicRef: Send
T: Sync -> AtomicRef: Send
T: Sync -> AtomicRefMut: Sync
T: Send -> AtomicRefMut: Send

[Imberflur]

It's interesting that T: 'b doesn't need to be written here when it is included on the struct definition. I think this is because lifetime bounds can be implied.

[SrTobi]

Yeah, I was surprised as well... and also that it leads to this inference problem... I could have bet that it's because of the hrtb. but no... those are fine

[Imberflur]

Yeah, I was surprised as well... and also that it leads to this inference problem... I could have bet that it's because of the hrtb. but no... those are fine

I would be surprised if the hrtb aren't involved though, i.e. I would guess that it is this combined with the hrtb

[Imberflur]

e.g. I'm guessing this would not cause issues even though it keeps the unnecessary T: 'b bound, although I'm not sure 🙃

unsafe impl<'b, T: ?Sized + 'b> Send for AtomicRefMut<'b, T> where T: Send {}

[SrTobi]

No, I removed them first, because I thought they were the reason... but the error stayed... only removing 'b from T fixed it...

I also think we can leave the for<'a> &'a mut T: * ... that's a good trick to let the compiler do the reasoning, as you are right, AtomicRefMut<T> is equivalent to &mut T

[Imberflur]

No, I removed them first, because I thought they were the reason... but the error stayed... only removing 'b from T fixed it...

Oh!

[Imberflur]

I also think we can leave the for<'a> &'a mut T: * ... that's a good trick to let the compiler do the reasoning, as you are right, AtomicRefMut is equivalent to &mut T

That sounds reasonable to me then.

[bholley]

Thanks!