Switch from lazy_static to once_cell about ahash

[tkaitchuck]

Ok. c828cc6
Will be included in the next release.

[tkaitchuck]

I prefer once_cell. I actually started with it. However it pulls in a huge number of transitive dependencies (or at least it did when I last checked), and I would like to keep aHash as light as possible given how low level it is.

[jplatte]

Just checked, once_cell has a single optional dependency (parking_lot). Same for version 1.0.1 (1.0.0 is yanked). Maybe a pre-1.0 version had more deps, but it's certainly not the case now 🙂

[tkaitchuck]

I think the trouble is that aHash would need parking_lot. Per the doc:

To implement a sync flavor of OnceCell, this crates uses either a custom re-implementation of std::sync::Once or parking_lot::Mutex. This is controlled by the parking_lot feature, which is enabled by default. Performance is the same for both cases, but the parking_lot based OnceCell is smaller by up to 16 bytes.

Which is sort of a shame, because the whole reason for the need for the lock is to block to make sure the initialization code doesn't get run multiple times. However I don't actually care about that. If there were an "ApproximatelyOnceCell" with zero deps that would be ideal.

[jplatte]

I don't exactly understand. Is it that you wouldn't want to use once_cell with a 16 byte overhead?

[tkaitchuck]

I don't care about the 16 bytes. I care about pulling in a dependency on parking_lot. If I can avoid having all those deps and get a working solution I'd be happy. I created matklad/once_cell#133 to track the issue. Either of matklad/once_cell#61 or matklad/once_cell#53 would also be acceptable.

[jplatte]

Oh, so the issue is no_std support. That wasn't clear before. Thanks for following up with once_cell, I'm curious to see where that goes; it would be relevant for tracing too (tokio-rs/tracing#1165).