Change flags for having an atomic u64

[notgull]

Motivation

See this comment.

Solution

The aim of this PR is to use the target_has_atomic feature on Rust 1.60, in order to avoid using a less efficient strategy (e.g. mutexes) on platforms that do actually have atomic U64. It changes the cfg_(not_)has_atomic_u64 to incorporate target_has_atomic = "64" based on a build flag.

[taiki-e]

LGTM

[Darksonn]

If we're going to have a probe, we might as well replace the entire 64-bit atomic detection logic with a build-time probe that just checks whether the AtomicU64 type exists.

[taiki-e]

If we're going to have a probe, we might as well replace the entire 64-bit atomic detection logic with a build-time probe that just checks whether the AtomicU64 type exists.

Unfortunately, probe is not very robust, and detecting everything that way does not work well in some environments or targets. See also rust-lang/futures-rs#2294.

[Darksonn]

We could have it use target_has_atomic when available, and use an autocfg probe for old compilers.

[Darksonn]

As an additional precaution against autocfg being brittle, we can have the build script emit a warning and enable everything if autocfg fails. Users can explicitly disable things themselves by passing --cfg flags if they're on a weird platform where autocfg doesn't work.

[notgull]

I'm fine with that approach, since that particular autocfg probe won't run on newer versions of Rust, so the performance/weirdness penalty won't come up for versions we care about.

[taiki-e]

As an additional precaution against autocfg being brittle, we can have the build script emit a warning and enable everything if autocfg fails. Users can explicitly disable things themselves by passing --cfg flags if they're on a weird platform where autocfg doesn't work.

Well, the warnings from the build script are usually not shown to users, but I'm fine with this approach since only the old rustc users are affected.

[Darksonn]

This variable naming is a bit confusing currently, because target_has_atomic_u64 is always true on the latest compiler even if the target doesn't have atomic u64.

[notgull]

Would something like target_needs_atomic_u64_fallback be better?

[Darksonn]

That's certainly an option.

[Darksonn]

It seems like the boolean is inverted here. Surely we would only disable atomic u64 if we need the fallback?

[notgull]

For some reason, when I do this, it looks like the FreeBSD checks fail?

[taiki-e]

(CI issue is maybe related to https://www.githubstatus.com/incidents/4p7zympgsr86)

[Darksonn]

You can push an empty commit to trigger CI again when the incident is resolved.