Use cfg(target_has_atomic) instead of hard coding which targets support atomics

[bjorn3]

This has been stable since 1.60.0: rust-lang/rust#93824 The hard coded list is bound to go out of sync with rustc, especially when new targets get added. As rustc itself depends on crossbeam, this means it isn't possible to build rustc for targets lacking some atomics until a new crossbeam release with those new targets. Using cfg(target_has_atomic) will immediately work. While the MSRV of crossbeam is lower than 1.60.0, it should be possible to use the hard coded list as fallback when the rustc version is too old for stable cfg(target_has_atomic).

[bjorn3]

cc rust-lang/rust#117170 (comment)

[taiki-e]

While the MSRV of crossbeam is lower than 1.60.0, it should be possible to use the hard coded list as fallback when the rustc version is too old for stable cfg(target_has_atomic).

This is an approach I have previously explicitly stated I accept and would accept a PR that implements this.
#937 (comment)

[bjorn3]

Do you know how to handle cfg(not(crossbeam_no_atomic_cas))? Use cfg(target_has_atomic="ptr")?

[taiki-e]

Use cfg(target_has_atomic="ptr")?

Yes.

• not(crossbeam_no_atomic_cas) corresponds to target_has_atomic = "ptr".
• not(crossbeam_no_atomic_64) (almost) corresponds to target_has_atomic = "64". (To be precise, it is closer to unstable target_has_atomic_load_store = "64", but currently, the only time the two would actually have different meanings is when using a custom no-std target.)
• not(crossbeam_no_atomic) (almost) corresponds to unstable target_has_atomic_load_store = "ptr". There is no stable alternative to this, so I think this may require continuing to use our own cfg. This cfg only affects some no-std targets.

[bjorn3]

Is adding autocfg as build dependency to the crates that don't have it yet fine?

[taiki-e]

Yeah, it's fine.

That said, given that #1015 will increase MSRV to 1.59, it would also actually be fine to increase MSRV to 1.60 or 1.61 and remove the need for build scripts. (I had forgotten about that PR when I first commented in this issue.)

[bjorn3]

I think I will wait for that PR.