Re: cgd and aes instruction set

[Robert Swindells <rjs%fdy2.co.uk@localhost>]

Alexander Nasonov <alnsn%yandex.ru@localhost> wrote:
>I'd like to bring up two things on this list. The first one is
>trivial, the second will take some time to implement but it should
>be pretty straightforward.
>
>2. Our current impementation is a bit slow and it's not constant-time.
>   I'd like to work on improving it. Crypto code itself is
>   straightforward but it uses simd registers (or vfp, I'm not very
>   sure on the terminology), and they should be saved/restored.
>   ryo@ pointed me out at matt@'s work in cpu_in_csum_neon.c but
>   it's for arm32 and it's incomplete. Is there anything similar
>   for aarch64?

Do we know what is incomplete in the NEON cksum code ? It doesn't seem
to be enabled for any kernel config.

We do seem to build some of the aarch64 assembler sources for openssl,
some others seem to be disabled. The aes sources seem to use byte values
for the ARMv8-a crypto instructions, I'm guessing some versions of
binutils don't know about the crypto mnemonics.