Move rand_pcg to rngs repository

[vks]

rand_pcg is no longer used for SmallRng, therefore it may be moved to the rust-random/rngs repository. This does not affect the API, only the repository and the issue tracker.

This will result in a bit of churn, but I think it's good to move unused code from the rust-random/rand repository.

We might also want to move rand_hc, which is currently only relevant for emscripten targets, because they lack u128 support. Maybe this was fixed in the meanwhile?

[dhardy]

I agree that rand_pcg should move. A little searching the web doesn't show me any progress on making emscripten support 128-bit integers. I think rand_hc should therefore stay here.

[dhardy]

Just looked at moving rand_pcg and realised it's still used by quite a few tests in rand and rand_distr. So, we have to choose:

• depending on out-of-tree RNGs is problematic when bumping the rand_core version, which isn't 1.0 yet
• PCG is not much code; we could just put an implementation in each library (under #[cfg(test)])
• rand::rngs::SmallRng is not suitable for our value stability tests
• we could just leave rand_pcg where it is (in the source tree)

[vks]

I see. I think our best option is to embed the PCG implementation, or leave it as it is.

[vks]

About emscripten: It looks like they fixed 128-bit integer support according to emscripten-core/emscripten-fastcomp#169?

[dhardy]

I see. I think our best option is to embed the PCG implementation, or leave it as it is.

Yes. Either is fine IMO.

About emscripten ...

So we can test using rand_chacha now? Except, we don't have any CI tests using emscripten since #874, so we need some way of testing whether removing the emscripten special casing in Cargo.toml is okay. I wonder if someone like @est31 could help with that?

[vks]

Now that we no longer have a special case for emscripten, I agree that we have the following options:

1. Inline PCG and remove rand_pcg.
2. Replace use of rand_pcg with rand_chacha.
3. Keep rand_pcg and close this issue.

Option 2 implies significantly more churn than the others, but I don't see any other problems.

[dhardy]

Option (3) is the least effort. Is there a good reason not to leave it here?

[newpavlov]

I think we can follow (3) for now and open a low-priority issue for porting tests from rand_pcg to rand_chacha.

[vks]

Thanks for the input, I opened #1185 in favor of this issue.