Re: [PATCH] lib/crypto: x86/nh: Optimize nh_avx2()

[Eric Biggers <ebiggers@kernel.org>]

On Wed, Nov 04, 2026 at 08:24:42PM -0800, Eric Biggers wrote:
> Various microoptimizations for nh_avx2().  Nothing too exciting:
> 
> - Instead of adding the message words to the key words and then
>   permuting both, do the following: permute both when loading, add the
>   permuted words, and multiply together adjacent pairs.  This saves one
>   vpshufd instruction per 32 bytes processed, and the ones that remain
>   are now better separated in the flow.
> 
> - Reassign register aliases to put the most frequently used variables in
>   %ymm0-%ymm7, which take fewer bytes to encode in many instructions.
> 
> - Update the registers used in _nh_2xstride() to match nh_sse2() and
>   eliminate the need for T7.
> 
> - Replace the initial 'add $0x20, KEY' with base+displacement
>   addressing, which was already being used anyway in most cases.
> 
> - Update the horizontal sum algorithm at the end of the function to save
>   a couple instructions.
> 
> nh_avx2() shrinks from 664 bytes to 518 bytes, and a quick benchmark
> shows 1% faster Adiantum encryption on x86_64.
> 
> Signed-off-by: Eric Biggers <ebiggers@kernel.org>
> ---
>  lib/crypto/x86/nh-avx2.S | 116 +++++++++++++++++++--------------------
>  1 file changed, 57 insertions(+), 59 deletions(-)

Applied to https://git.kernel.org/pub/scm/linux/kernel/git/ebiggers/linux.git/log/?h=libcrypto-next

- Eric