breaking: directly depend on crossbeam-channel instead of pulling in all of crossbeam

[m4rch3n1ng]

Warning

this is a breaking change, since you have previously been re-exporting crossbeam.

termimad does not actually need most of the features of crossbeam, it only needs its channel. so you can save on a few unneeded dependencies if you just depend on the crossbeam-channel directly.

at this point it might be worth it to consider just using the std mpsc channel (or putting crossbeam behind a feature flag), as since rust 1.67 the std mpsc and the crossbeam channel are mostly the same performance wise and the extra functionality that crossbeam provides is, as far as i can tell, not used here.

[Canop]

What you say is reasonnable.

I have still to try estimate the consequences for applications using the reexported crossbeam. Many of them use the select! macro, for exemple, and mpmc capabilities.

Did you measure a gain (binary size, perhaps) ?

[m4rch3n1ng]

I have still to try to estimate the consequences for applications using the reexported crossbeam.

i think it's fair to expect people that need crossbeam to be able to import it themselves. i personally am against exporting entire crates entirely (with a few exceptions like for macros, but i think those should be marked as doc(hidden)), as that can cause a lot more breaking changes than necessary. if crossbeam has a breaking 0.9 release, then you have no way of cleanly upgrading to that version without having to release a breaking change yourself. that's why i would actually advocate to completely remove the re-export of crossbeam-channel and just ask users to add the dependency themselves.

i would also still advocate for putting the crossbeam dependency behind an optional feature and to just rely on the std mpsc by default.

Did you measure a gain (binary size, perhaps)

i don't really care about binary size (unless it gets concerningly big, which i highly doubt would be the fault of crossbeam), but i mostly did this for compile time. on a clean project i can consistently get ~1s of speedup (~11s -> ~10s) for a clean build in --release and ~.5s of speedup for a debug build (~6.5s -> ~6.0). on a larger project like steel the speedup is almost entirely negligible and probably just noise (although i still think it is unnecessary ).

[m4rch3n1ng]

i also wanted to say that this is not time-sensitive at all. if you want to wait for something else that warrants a breaking change anyway - something like a new crossterm version - and only then merge this, then that would be fine by me too, as this is not actually a change in any functionality.

[Canop]

i think it's fair to expect people that need crossbeam to be able to import it themselves

The point of reexporting the crate is to prevent version conflicts. That's why this reexport and the one of crossterm were required. I'm aware there are also downsides and that it's a balance to find, but it's not something that can be changed lightly.