This makes sense. I'll be officially discontinuing and freezing my crate since rand turned out to be its only direct user (besides what I originally wrote it for, which is defunct). There's some simplification I plan to do to the ppv-lite86 API eventually, mostly consolidating traits. I'll probably do that when I add the next backend (avx512, NEON, WASM, and packed_simd are planned eventually, mostly waiting on std::arch implementing the intrinsics), so after I do that I can PR the corresponding update here.

You might want to annex some of my test cases too; they're closer to the implementation and able to cover some edge cases the rand_chacha tests probably don't.

I've also just noticed that I SIMD-accelerated a few top-level functions that aren't performance-sensitive, so I've written a patch to just do those bits with portable code. It's a minor change but to keep things simple I'll PR it separately after this lands.

I see no reason not to merge this now. All happy?

(Whether we replace this crate is another issue that doesn't need to influence this PR — I'll make a new issue.)

In the end you merged c2-chacha into rand_chacha, so the chacha used here no longer exists stand alone? I suppose doing this sounds unavoidable because rand supplies its own BlockRng which differs from that supplied by stream ciphers.

I guess the chacha crate https://github.com/RustCrypto/stream-ciphers/tree/master/chacha20 is a dependency zoo until const generics land, and still significantly slower.

Less 'unavoidable' than 'an unnecessary extra step'.

This crate (c2-chacha) already had some work put into it to reduce dependencies. The same could be done for chacha20. The primary reason not to do so at the moment is that the latter is slower and doesn't auto-detect CPU features (#934).

I guess the chacha crate is a dependency zoo until const generics land, and still significantly slower.

It only has two mandatory dependencies: cfg-if and cpufeatures. We could probably get rid of the former if it's a big concern.

Regarding performance: RustCrypto/stream-ciphers#267 (comment)

[chacha20] doesn't auto-detect CPU features

It does now!

Looks like the performance gap is closed, that's very nice! Migrating to chacha20 would require bumping the MSRV to 1.51.

It looks like chacha20 always tries to uses SSE2 on x86 and x86_64 even if it isn't available. It only checks if AVX2 exists. This is will cause issues on certain CPUs. While the doc comments says "the sse2 target feature is enabled-by-default on all x86(_64) CPUs", this is not true for the i586 targets. Only for the i686 and x86_64 targets.

@bjorn3 I opened RustCrypto/stream-ciphers#270 for that

Thanks!