Separated the complex buffering logic from key stream generation more
clearly, added plenty of comments and generally refactored the Go
implementation for readability. Made the interface with the
generic/assembly cores smaller and more consistent, according to
golang.org/wiki/TargetSpecific.
We will recover the lost performance on unaligned calls by caching 3/4
of the first round across XORKeyStream invocations, which we now have
complexity budget for.
name old speed new speed delta
ChaCha20/64-4 435MB/s ± 2% 429MB/s ± 2% -1.47% (p=0.013 n=10+9)
ChaCha20/256-4 496MB/s ± 1% 493MB/s ± 2% ~ (p=0.280 n=10+10)
ChaCha20/10x25-4 283MB/s ± 1% 274MB/s ± 2% -3.13% (p=0.000 n=10+10)
ChaCha20/4096-4 494MB/s ± 1% 493MB/s ± 5% ~ (p=0.631 n=10+10)
ChaCha20/100x40-4 421MB/s ± 3% 408MB/s ± 1% -3.14% (p=0.003 n=9+9)
ChaCha20/65536-4 515MB/s ± 1% 519MB/s ± 3% ~ (p=0.161 n=7+10)
ChaCha20/1000x65-4 501MB/s ± 2% 501MB/s ± 3% ~ (p=0.497 n=9+10)
Also applied a fix for a lingering bug in the ppc64le assembly written
by Lynn Boger <laboger@linux.vnet.ibm.com>.
Updates golang/go#24485
Change-Id: I10cf24a7f10359b1b4ae63c9bb1946735b98ac9b
Reviewed-on: https://go-review.googlesource.com/c/crypto/+/185439
Reviewed-by: Michael Munday <mike.munday@ibm.com>
Filippo Valsorda
85e5e33df3