Add a system for explicitly marking const fns as promotable

[oli-obk]

r? @nikomatsakis

addresses the last point of https://gist.github.com/nikomatsakis/eee212d7c4a49ae75e19765c1e06bf49

[rust-highfive]

Warning

• These commits modify submodules.

[gnzlbg]

As mentioned on the stdsimd PR (rust-lang/stdarch#471): the feature has no tracking issue, neither the PR nor the commits explain what it does, nor there are any docs... yet std library APIs need to be marked with the attribute.

I don't know where the right place to document this is, but I think it should be explained somewhere (and maybe the libs team should be pinged here as well), since std maintainers need to know what this is and how to use it.

[oli-obk]

The idea is that you do not need to mark anything as promotable. We just needed an opt-in for preventing the automatic promotion of any const fn call due to #50814 . Without that opt in, further stabilization of const fns is not something we want to do.

[nikomatsakis]

let's extend the "promotable" whitelist to intrinsics too

[nikomatsakis]

I approve of the general approach — but I agree that we should document. I was gonna say the same thing. @oli-obk and I talked and agreed we'd put it in the unstable reference.

[nikomatsakis]

cc @eddyb — do you want to take a look at this PR? (Not strictly necessary, I don't think) It adds a separate attribute to mark const fns as promotable.

---

@gnzlbg as I wrote above, I agree we need docs — but I think the short version is that we would only very rarely (if ever) want to add this attribute, and only after consulting with @oli-obk and/or myself. Basically we want to be conservative about what we will automatically promote to static lifetime (that is, you write let x = &foo and we allow that to escape the fn) until we've fully worked out the rules.

[oli-obk]

cc @eddyb

[eddyb]

I'm not sure we should mark any const fn to be promotable, if it does anything to its arguments other than moving them into the return value (usually some sort of ADT).
There is a risk that any operation may fail when applied to an unknown value.

[emilyalbini]

Ping from triage @nikomatsakis! This PR needs your review.

[oli-obk]

closing in favour of #51570