in my defense, nobody noticed the lack of a v prefix either

Author: floodyberry

src/scryptjane/scrypt-jane-mix_chacha-avx.h

@@ -104,9 +104,9 @@ asm_naked_fn(scrypt_ChunkMix_avx)
 			a3(vpshufb xmm3,xmm3,xmm5)
 			a3(vpshufd xmm0,xmm0,0x39)
 			a3(vpaddd xmm2,xmm2,xmm3)
-			a3(pshufd xmm3,xmm3,0x4e)
+			a3(vpshufd xmm3,xmm3,0x4e)
 			a3(vpxor  xmm1,xmm1,xmm2)
-			a3(pshufd xmm2,xmm2,0x93)
+			a3(vpshufd xmm2,xmm2,0x93)
 			a3(vpsrld xmm6,xmm1,25)
 			a3(vpslld xmm1,xmm1,7)
 			a3(vpxor  xmm1,xmm1,xmm6)
@@ -226,9 +226,9 @@ asm_naked_fn(scrypt_ChunkMix_avx)
 			a3(vpshufb xmm3,xmm3,xmm5)
 			a3(vpshufd xmm0,xmm0,0x39)
 			a3(vpaddd xmm2,xmm2,xmm3)
-			a3(pshufd xmm3,xmm3,0x4e)
+			a3(vpshufd xmm3,xmm3,0x4e)
 			a3(vpxor  xmm1,xmm1,xmm2)
-			a3(pshufd xmm2,xmm2,0x93)
+			a3(vpshufd xmm2,xmm2,0x93)
 			a3(vpsrld xmm12,xmm1,25)
 			a3(vpslld xmm1,xmm1,7)
 			a3(vpxor  xmm1,xmm1,xmm12)

src/scryptjane/scrypt-jane-mix_salsa-avx.h

@@ -64,16 +64,16 @@ asm_naked_fn(scrypt_ChunkMix_avx)
 			a3(vpsrld xmm5, xmm4, 19)
 			a3(vpslld xmm4, xmm4, 13)
 			a3(vpxor xmm1, xmm1, xmm5)
-			a3(pshufd xmm3, xmm3, 0x93)
+			a3(vpshufd xmm3, xmm3, 0x93)
 			a3(vpxor xmm1, xmm1, xmm4)
 			a3(vpaddd xmm4, xmm2, xmm1)
 			a3(vpsrld xmm5, xmm4, 14)
 			a3(vpslld xmm4, xmm4, 18)
 			a3(vpxor xmm0, xmm0, xmm5)
-			a3(pshufd xmm2, xmm2, 0x4e)
+			a3(vpshufd xmm2, xmm2, 0x4e)
 			a3(vpxor xmm0, xmm0, xmm4)
 			a3(vpaddd xmm4, xmm3, xmm0)
-			a3(pshufd xmm1, xmm1, 0x39)
+			a3(vpshufd xmm1, xmm1, 0x39)
 			a3(vpsrld xmm5, xmm4, 25)
 			a3(vpslld xmm4, xmm4, 7)
 			a3(vpxor xmm1, xmm1, xmm5)
@@ -87,15 +87,15 @@ asm_naked_fn(scrypt_ChunkMix_avx)
 			a3(vpsrld xmm5, xmm4, 19)
 			a3(vpslld xmm4, xmm4, 13)
 			a3(vpxor xmm3, xmm3, xmm5)
-			a3(pshufd xmm1, xmm1, 0x93)
+			a3(vpshufd xmm1, xmm1, 0x93)
 			a3(vpxor xmm3, xmm3, xmm4)
 			a3(vpaddd xmm4, xmm2, xmm3)
 			a3(vpsrld xmm5, xmm4, 14)
 			a3(vpslld xmm4, xmm4, 18)
 			a3(vpxor xmm0, xmm0, xmm5)
-			a3(pshufd xmm2, xmm2, 0x4e)
+			a3(vpshufd xmm2, xmm2, 0x4e)
 			a3(vpxor xmm0, xmm0, xmm4)
-			a3(pshufd xmm3, xmm3, 0x39)
+			a3(vpshufd xmm3, xmm3, 0x39)
 			a2(sub eax, 2)
 			aj(ja scrypt_salsa_avx_loop)
 		a3(vpaddd xmm0,xmm0,[esp+0])
@@ -184,16 +184,16 @@ asm_naked_fn(scrypt_ChunkMix_avx)
 			a3(vpsrld xmm5, xmm4, 19)
 			a3(vpslld xmm4, xmm4, 13)
 			a3(vpxor xmm1, xmm1, xmm5)
-			a3(pshufd xmm3, xmm3, 0x93)
+			a3(vpshufd xmm3, xmm3, 0x93)
 			a3(vpxor xmm1, xmm1, xmm4)
 			a3(vpaddd xmm4, xmm2, xmm1)
 			a3(vpsrld xmm5, xmm4, 14)
 			a3(vpslld xmm4, xmm4, 18)
 			a3(vpxor xmm0, xmm0, xmm5)
-			a3(pshufd xmm2, xmm2, 0x4e)
+			a3(vpshufd xmm2, xmm2, 0x4e)
 			a3(vpxor xmm0, xmm0, xmm4)
 			a3(vpaddd xmm4, xmm3, xmm0)
-			a3(pshufd xmm1, xmm1, 0x39)
+			a3(vpshufd xmm1, xmm1, 0x39)
 			a3(vpsrld xmm5, xmm4, 25)
 			a3(vpslld xmm4, xmm4, 7)
 			a3(vpxor xmm1, xmm1, xmm5)
@@ -207,15 +207,15 @@ asm_naked_fn(scrypt_ChunkMix_avx)
 			a3(vpsrld xmm5, xmm4, 19)
 			a3(vpslld xmm4, xmm4, 13)
 			a3(vpxor xmm3, xmm3, xmm5)
-			a3(pshufd xmm1, xmm1, 0x93)
+			a3(vpshufd xmm1, xmm1, 0x93)
 			a3(vpxor xmm3, xmm3, xmm4)
 			a3(vpaddd xmm4, xmm2, xmm3)
 			a3(vpsrld xmm5, xmm4, 14)
 			a3(vpslld xmm4, xmm4, 18)
 			a3(vpxor xmm0, xmm0, xmm5)
-			a3(pshufd xmm2, xmm2, 0x4e)
+			a3(vpshufd xmm2, xmm2, 0x4e)
 			a3(vpxor xmm0, xmm0, xmm4)
-			a3(pshufd xmm3, xmm3, 0x39)
+			a3(vpshufd xmm3, xmm3, 0x39)
 			a2(sub rax, 2)
 			aj(ja scrypt_salsa_avx_loop)
 		a3(vpaddd xmm0,xmm0,xmm8)