On Mon, 8 Apr 2024 at 16:18, Stefan Kanthak <[email protected]> wrote:
>
> Get rid of 16 byte constant, use shift instead of mask
>
This is missing a signoff
> --- -/arch/x86/crypto/sha1_ni_asm.S
> +++ +/arch/x86/crypto/sha1_ni_asm.S
> @@ -104,9 +104,9 @@
> add DATA_PTR, NUM_BLKS /* pointer to end of data */
>
> /* load initial hash values */
> - pinsrd $3, 1*16(DIGEST_PTR), E0
> + pinsrd $0, 1*16(DIGEST_PTR), E0
> movdqu 0*16(DIGEST_PTR), ABCD
> - pand UPPER_WORD_MASK(%rip), E0
> + pslldq $12, E0
> pshufd $0x1B, ABCD, ABCD
>
Wouldn't it be simpler and therefore better to simply zero register E0
before loading E into the top lane?
> movdqa PSHUFFLE_BYTE_FLIP_MASK(%rip), SHUF_MASK
> @@ -297,8 +297,3 @@
> .align 16
> PSHUFFLE_BYTE_FLIP_MASK:
> .octa 0x000102030405060708090a0b0c0d0e0f
> -
> -.section .rodata.cst16.UPPER_WORD_MASK, "aM", @progbits, 16
> -.align 16
> -UPPER_WORD_MASK:
> - .octa 0xFFFFFFFF000000000000000000000000
>