Bumping MSRV of getrandom 0.2

[josephlr]

A few issues and PRs (#279 #281) have brought up the idea of increasing the MSRV of this crate. Currently the MSRV is 1.34, and I've seen 3 features we might want to use which are only present in later versions:

• MaybeUninit: added in 1.36 (Released 2019-07-04)
• <*mut T>::cast: added in 1.38 (Released 2019-09-26)
• Const Generics: Added in 1.51 (Released 2021-03-25)

When raising the MSRV, we should keep in mind that rand 0.8 and rand_core 0.6 have an MSRV of 1.36, and they are our largest users. Other top crates that directly depend on us include:

• uuid: MSRV 1.57
• ahash: MSRV 1.43.1
• fastrand: MSRV 1.34 (but it's only a dev dependancy)
• os_str_bytes: MSRV 1.57
• redox_users: MSRV at least 1.36
• async-io: MSRV 1.46
• zip: MSRV 1.54

Also, we can always use rustversion to have APIs that require a newer version of Rust without increasing our MSRV, provided we can implement the newer API in terms of one of the older APIs.

I would propose: raising the MSRV to 1.36 so we can use MaybeUninit inside of our crate.

For example,

pub fn getrandom_uninit(buf: &mut [MaybeUninit<u8>]) -> Result<&[u8], Error> {
  // Unconditionally use MaybeUninit
  // ...
}

#[rustversion::since(1.51)]
#[inline]
pub fn getrandom_array<const N: usize>() -> Result<[u8; N], Error> {
  let mut arr: [MaybeUninit<u8>; N] = [MaybeUninit;;uninit(); N];
  let s: &mut [u8; N] = getrandom_uninit(&mut arr)?.try_into().unwrap();
  Ok(*s)
}

[notgull]

I think getrandom_array is a superfluous addition to this crate that is out of its scope. It should belong in crates like rand.

Personally, I don't mind a bump from 1.34 to 1.36. It's only two versions, and they were released long enough ago that it shouldn't really matter. However, I would mind a bump to 1.51. I maintain some crates that aim to target the current distro version of Debian Stable, which is 1.48. If getrandom bumps to 1.51 it would no longer build here.

This is a pretty hotly contended issue in the Rust community, see matklad/once_cell#201

(async-io has an MSRV of 1.46, but it shouldn't matter since it's only a testing dependency anyhow).

[briansmith]

<*mut T>::cast: added in 1.38 (Released 2019-09-26)

This feature is trivial to polyfill, IIUC, so it's not very compelling to bump MSRV for it. Though, I wouldn't object to it.

I would propose: raising the MSRV to 1.36 so we can use MaybeUninit inside of our crate.

I think that's extremely uncontroversial given the data you looked up, especially regarding rand's current MSRV. I suggest bumping MSRV to 1.36 for now so the MaybeUninit stuff can move forward, and decide separately whether to bump it further. In particular, I don't think the decision of "1.36" vs "newer" should block the MaybeUninit stuff. It would probably be smart to get the MaybeUninit stuff landed and shipped in a release before bumping the MSRV further for other features.

The CI/CD configuration and ensuring tests cover all the different code paths gets really messy really fast when there are optional features, so I would recommend avoiding the use of optional features to try to allow people to opt into newer language features like const generics.

[briansmith]

If we make the MSRV 1.37 or higher, instead of 1.36, then we can use MaybeUninit<u8> in FFI type declarations, which would allow us to remove several as *mut u8 casts from the code (possibly to be) added in PR #291; see https://github.com/rust-lang/rust/blob/master/RELEASES.md#libraries-27.

[newpavlov]

If we make the MSRV 1.37 or higher, instead of 1.36, then we can use MaybeUninit in FFI type declarations

I don't think there is enough merit in it, the casts are trivially safe, having *mut MaybeUninit<u8> in FFI signatures of external functions would be somewhat surprising, and it would affect rand's MSRV.

[briansmith]

• 1.50.0: See https://github.com/rust-random/getrandom/pull/291#discussion_r997495101, where we are discussing adding our own unsafe implementation of slice::fill, which was added in 1.50.0.
• 1.68.0 or so: https://github.com/rust-lang/rust/pull/99767 is stabilizing feature(target_feature_11) which, I believe, will allow us to reduce the amount of unsafe in the rdrand implementation.

[briansmith]

• 1.54.0: target_family="wasm" was added, which is relevant to #296, though we can probably live without it.

[newpavlov]

I don't think that target support influences crate's MSRV. You obviously can not build crate for a target with a toolchain which does not anything about it. We already have several targets which were introduced after Rust 1.34 was released.

[josephlr]

I don't think that target support influences crate's MSRV. You obviously can not build crate for a target with a toolchain which does not anything about it. We already have several targets which were introduced after Rust 1.34 was released.

I think @briansmith just meant that we could shorten the code in #303 by using

#[cfg(all(target_family = "wasm", target_os = "unknown"))]

instead of

#[cfg(all(any(target_arch = "wasm32", target_arch = "wasm64"), target_os = "unknown"))]

but I agree, we can live without it.

[josephlr]

With #291 and #305 increasing the MSRV to 1.36, and the fact that rand 0.8 has an MSRV of 1.36 (making it harder to raise the MSRV further), can we close this issue?

[briansmith]

I don't object to closing the issue, as I don't think increasing the MSRV even requires an issue. :)