[PATCH] lib/crypto: x86/chacha: Drop redundant moves from AVX2 code

[PATCH] lib/crypto: x86/chacha: Drop redundant moves from AVX2 code

[Eric Biggers]

The AVX2 optimized ChaCha code has many instances of moving one register
to another, then left-shifting the destination register with 'vpslld'.
This pattern likely was copied from the SSSE3 code, which uses 'pslld'.
However, like all the other AVX instructions, 'vpslld' has separate
source and destination registers.  Just utilize that instead.

Signed-off-by: Eric Biggers <ebiggers@kernel.org>
---
 lib/crypto/x86/chacha-avx2-x86_64.S | 36 ++++++++++-------------------
 1 file changed, 12 insertions(+), 24 deletions(-)

diff --git a/lib/crypto/x86/chacha-avx2-x86_64.S b/lib/crypto/x86/chacha-avx2-x86_64.S
index 18ae03e05074f..8aa3227f7302b 100644
--- a/lib/crypto/x86/chacha-avx2-x86_64.S
+++ b/lib/crypto/x86/chacha-avx2-x86_64.S
@@ -74,8 +74,7 @@ SYM_FUNC_START(chacha_1to2blocks_avx2)
 	# x2 += x3, x1 = rotl32(x1 ^ x2, 12)
 	vpaddd		%ymm3,%ymm2,%ymm2
 	vpxor		%ymm2,%ymm1,%ymm1
-	vmovdqa		%ymm1,%ymm6
-	vpslld		$12,%ymm6,%ymm6
+	vpslld		$12,%ymm1,%ymm6
 	vpsrld		$20,%ymm1,%ymm1
 	vpor		%ymm6,%ymm1,%ymm1
 
@@ -87,8 +86,7 @@ SYM_FUNC_START(chacha_1to2blocks_avx2)
 	# x2 += x3, x1 = rotl32(x1 ^ x2, 7)
 	vpaddd		%ymm3,%ymm2,%ymm2
 	vpxor		%ymm2,%ymm1,%ymm1
-	vmovdqa		%ymm1,%ymm7
-	vpslld		$7,%ymm7,%ymm7
+	vpslld		$7,%ymm1,%ymm7
 	vpsrld		$25,%ymm1,%ymm1
 	vpor		%ymm7,%ymm1,%ymm1
 
@@ -107,8 +105,7 @@ SYM_FUNC_START(chacha_1to2blocks_avx2)
 	# x2 += x3, x1 = rotl32(x1 ^ x2, 12)
 	vpaddd		%ymm3,%ymm2,%ymm2
 	vpxor		%ymm2,%ymm1,%ymm1
-	vmovdqa		%ymm1,%ymm6
-	vpslld		$12,%ymm6,%ymm6
+	vpslld		$12,%ymm1,%ymm6
 	vpsrld		$20,%ymm1,%ymm1
 	vpor		%ymm6,%ymm1,%ymm1
 
@@ -120,8 +117,7 @@ SYM_FUNC_START(chacha_1to2blocks_avx2)
 	# x2 += x3, x1 = rotl32(x1 ^ x2, 7)
 	vpaddd		%ymm3,%ymm2,%ymm2
 	vpxor		%ymm2,%ymm1,%ymm1
-	vmovdqa		%ymm1,%ymm7
-	vpslld		$7,%ymm7,%ymm7
+	vpslld		$7,%ymm1,%ymm7
 	vpsrld		$25,%ymm1,%ymm1
 	vpor		%ymm7,%ymm1,%ymm1
 
@@ -251,15 +247,13 @@ SYM_FUNC_START(chacha_2to4blocks_avx2)
 	# x2 += x3, x1 = rotl32(x1 ^ x2, 12)
 	vpaddd		%ymm3,%ymm2,%ymm2
 	vpxor		%ymm2,%ymm1,%ymm1
-	vmovdqa		%ymm1,%ymm10
-	vpslld		$12,%ymm10,%ymm10
+	vpslld		$12,%ymm1,%ymm10
 	vpsrld		$20,%ymm1,%ymm1
 	vpor		%ymm10,%ymm1,%ymm1
 
 	vpaddd		%ymm7,%ymm6,%ymm6
 	vpxor		%ymm6,%ymm5,%ymm5
-	vmovdqa		%ymm5,%ymm10
-	vpslld		$12,%ymm10,%ymm10
+	vpslld		$12,%ymm5,%ymm10
 	vpsrld		$20,%ymm5,%ymm5
 	vpor		%ymm10,%ymm5,%ymm5
 
@@ -275,15 +269,13 @@ SYM_FUNC_START(chacha_2to4blocks_avx2)
 	# x2 += x3, x1 = rotl32(x1 ^ x2, 7)
 	vpaddd		%ymm3,%ymm2,%ymm2
 	vpxor		%ymm2,%ymm1,%ymm1
-	vmovdqa		%ymm1,%ymm10
-	vpslld		$7,%ymm10,%ymm10
+	vpslld		$7,%ymm1,%ymm10
 	vpsrld		$25,%ymm1,%ymm1
 	vpor		%ymm10,%ymm1,%ymm1
 
 	vpaddd		%ymm7,%ymm6,%ymm6
 	vpxor		%ymm6,%ymm5,%ymm5
-	vmovdqa		%ymm5,%ymm10
-	vpslld		$7,%ymm10,%ymm10
+	vpslld		$7,%ymm5,%ymm10
 	vpsrld		$25,%ymm5,%ymm5
 	vpor		%ymm10,%ymm5,%ymm5
 
@@ -309,15 +301,13 @@ SYM_FUNC_START(chacha_2to4blocks_avx2)
 	# x2 += x3, x1 = rotl32(x1 ^ x2, 12)
 	vpaddd		%ymm3,%ymm2,%ymm2
 	vpxor		%ymm2,%ymm1,%ymm1
-	vmovdqa		%ymm1,%ymm10
-	vpslld		$12,%ymm10,%ymm10
+	vpslld		$12,%ymm1,%ymm10
 	vpsrld		$20,%ymm1,%ymm1
 	vpor		%ymm10,%ymm1,%ymm1
 
 	vpaddd		%ymm7,%ymm6,%ymm6
 	vpxor		%ymm6,%ymm5,%ymm5
-	vmovdqa		%ymm5,%ymm10
-	vpslld		$12,%ymm10,%ymm10
+	vpslld		$12,%ymm5,%ymm10
 	vpsrld		$20,%ymm5,%ymm5
 	vpor		%ymm10,%ymm5,%ymm5
 
@@ -333,15 +323,13 @@ SYM_FUNC_START(chacha_2to4blocks_avx2)
 	# x2 += x3, x1 = rotl32(x1 ^ x2, 7)
 	vpaddd		%ymm3,%ymm2,%ymm2
 	vpxor		%ymm2,%ymm1,%ymm1
-	vmovdqa		%ymm1,%ymm10
-	vpslld		$7,%ymm10,%ymm10
+	vpslld		$7,%ymm1,%ymm10
 	vpsrld		$25,%ymm1,%ymm1
 	vpor		%ymm10,%ymm1,%ymm1
 
 	vpaddd		%ymm7,%ymm6,%ymm6
 	vpxor		%ymm6,%ymm5,%ymm5
-	vmovdqa		%ymm5,%ymm10
-	vpslld		$7,%ymm10,%ymm10
+	vpslld		$7,%ymm5,%ymm10
 	vpsrld		$25,%ymm5,%ymm5
 	vpor		%ymm10,%ymm5,%ymm5
 

base-commit: 33b4f31b1aea3f62b61c892daefc0e6a77859032
-- 
2.56.1

Re: [PATCH] lib/crypto: x86/chacha: Drop redundant moves from AVX2 code

[Antonio Quartulli]

Hi,

On 11/11/2026 23:34, Eric Biggers wrote:
> The AVX2 optimized ChaCha code has many instances of moving one register
> to another, then left-shifting the destination register with 'vpslld'.
> This pattern likely was copied from the SSSE3 code, which uses 'pslld'.
> However, like all the other AVX instructions, 'vpslld' has separate
> source and destination registers.  Just utilize that instead.

Tested on a Ryzen 9 9950X in a KVM guest with AVX-512 hidden, so that
the AVX2 code is used: the ChaCha and ChaCha20-Poly1305 KUnit tests
pass.
The ChaCha benchmark shows no difference beyond noise (+/-0.6% at every
length, 3 runs each).

Reviewed-by: Antonio Quartulli <antonio@mandelbit.com>
Tested-by: Antonio Quartulli <antonio@mandelbit.com>

-- 
Antonio Quartulli

CEO and Co-Founder
Mandelbit Srl
https://www.mandelbit.com