CPU feature detection may be racy

[briansmith]

When cpu.rs was originally written, it used std::sync::Once. std::sync::Once explicitly guarantees "that any memory writes performed by the executed closure can be reliably observed by other threads at this point (there is a happens-before relation between the closure and code executing after the return)" and we relied on this to ensure that the writes to GFp_ia32cap_P and GFp_iarmcap_P are seen by all threads.

Later, cpu.rs was modified to use spin-rs directly instead of std::sync::Once However, spin-rs's Once does NOT provide the same guarantee in its documentation. (Neither does lazy_static!.) See https://mvdnes.github.io/rust-docs/spin-rs/spin/struct.Once.html#method.call_once.

I think that because x86 and x86-64 have a strong memory model, only the feature detection for other CPUs (currently ARM) may be racy.