From dbcfb0123374d83fbf44fcb80d9282969c702759 Mon Sep 17 00:00:00 2001
From: Solar Designer <solar@openwall.com>
Date: Sun, 9 Mar 2025 19:53:00 +0100
Subject: [PATCH] yescrypt: Use AVX512VL XOP-like bit rotates for faster
 Salsa20

---
 lib/alg-yescrypt-opt.c | 9 ++++++++-
 1 file changed, 8 insertions(+), 1 deletion(-)

diff --git a/lib/alg-yescrypt-opt.c b/lib/alg-yescrypt-opt.c
index a33c6e40..fc6e95e5 100644
--- a/lib/alg-yescrypt-opt.c
+++ b/lib/alg-yescrypt-opt.c
@@ -1,6 +1,6 @@
 /*-
  * Copyright 2009 Colin Percival
- * Copyright 2012-2018 Alexander Peslyak
+ * Copyright 2012-2025 Alexander Peslyak
  * All rights reserved.
  *
  * Redistribution and use in source and binary forms, with or without
@@ -47,6 +47,8 @@
 #if 0
 #ifdef __XOP__
 #warning "Note: XOP is enabled.  That's great."
+#elif defined(__AVX512VL__)
+#warning "Note: AVX512VL is enabled.  That's great."
 #elif defined(__AVX__)
 #warning "Note: AVX is enabled, which is great for classic scrypt and YESCRYPT_WORM, but is sometimes slightly slower than plain SSE2 for YESCRYPT_RW"
 #elif defined(__SSE2__)
@@ -81,6 +83,8 @@
 #include <emmintrin.h>
 #ifdef __XOP__
 #include <x86intrin.h>
+#elif defined(__AVX512VL__)
+#include <immintrin.h>
 #endif
 #elif defined(__SSE__)
 #include <xmmintrin.h>
@@ -179,6 +183,9 @@ static inline void salsa20_simd_unshuffle(const salsa20_blk_t *Bin,
 #ifdef __XOP__
 #define ARX(out, in1, in2, s) \
 	out = _mm_xor_si128(out, _mm_roti_epi32(_mm_add_epi32(in1, in2), s));
+#elif defined(__AVX512VL__)
+#define ARX(out, in1, in2, s) \
+	out = _mm_xor_si128(out, _mm_rol_epi32(_mm_add_epi32(in1, in2), s));
 #else
 #define ARX(out, in1, in2, s) { \
 	__m128i tmp = _mm_add_epi32(in1, in2); \