While the Go compiler currently will not eliminate dead memset
equivalents, attempt to future-pr...

* Change the license to a strong Copyleft.
* Use go modules.
* Change the canonical upstream ...

I seem to recall the compiler doing the right thing regardless even
prior to the intrinsics, sug...

Turns out the Go compiler doesn't align the state vector in certain
conditions. Sigh.

Oops, that bug's been there forever, since the time I added back the
ChaCha20 constant to the st...

Apparently needed to avoid false positives, see the discusson at:
https://bugzilla.mozilla.org/s...

Not a huge improvement since we spend a good amont of time fucking
around with registers or hitt...

Having them split didn't meaningfully benefit performance, and this is
easier to maintain.

LOL what the fuck is the crypto package doing...

SSE2:
BenchmarkChaCha20_16-4 20000000 61.7 ns/op 259.32 MB/s
BenchmarkChaCha20_64-4 1...

Yes, I used to have these macros and I got rid of them, but in hindsight
that was a dumb idea.

* Add a codepath that does 2 blocks at a time to handle 128 <= n < 256,
which is probably fa...

Bleah.

Everyone else's optimized implementation, except Ted Krovetz's does
this, so I should too. The ...

Just cleanups, on the off chance I decide to do AVX2 one day.

* It seems safe to assume that x is 16 byte aligned. I was doing so
anyway by the use of PA...

This is the IETF RFC 7539 style ChaCha20 with a 96 bit nonce and 32 bit
counter, used for constr...

Technically, I can do this in `ReKey()` and `HChaCha` as well, but it's
less straight forward, a...