Merge pull request lucasjones#4 from wolf9466/master

Minor optimizations
Gizzywhop · May 22, 2014 · 5db4950 · 5db4950
2 parents 834d79c + 194148b
commit 5db4950
Show file tree

Hide file tree

Showing 2 changed files with 15 additions and 17 deletions.
diff --git a/aesb-x64.S b/aesb-x64.S
@@ -11,13 +11,13 @@
 fast_aesb_single_round:
 _fast_aesb_single_round:
 #if defined(_WIN64) || defined(__CYGWIN__)
-    movdqu (%rcx), %xmm1
+    movdqa (%rcx), %xmm1
     aesenc (%r8), %xmm1
-    movdqu %xmm1, (%rdx)
+    movdqa %xmm1, (%rdx)
 #else
-    movdqu (%rdi), %xmm1
+    movdqa (%rdi), %xmm1
     aesenc (%rdx), %xmm1
-    movdqu %xmm1, (%rsi)
+    movdqa %xmm1, (%rsi)
 #endif
     ret
 
@@ -28,30 +28,28 @@ _fast_aesb_single_round:
 fast_aesb_pseudo_round_mut:
 _fast_aesb_pseudo_round_mut:
 #if defined(_WIN64) || defined(__CYGWIN__)
-    mov $0, %r9
-    mov $10, %r10
-    movdqu (%rcx), %xmm1
+    mov %rdx, %r9
+    add $0xA0, %r9
+    movdqa (%rcx), %xmm1
 
     .LOOP:
             aesenc (%rdx), %xmm1
             add $0x10, %rdx
-            inc %r9
-            cmp %r10, %r9
+			cmp %r9, %rdx
             jl .LOOP
 
-    movdqu %xmm1, (%rcx)
+    movdqa %xmm1, (%rcx)
 #else
-    mov $0, %r9
-    mov $10, %r10
-    movdqu (%rdi), %xmm1
+    mov %rsi, %r9
+    add $0xA0, %r9
+    movdqa (%rdi), %xmm1
 
     .LOOP:
             aesenc (%rsi), %xmm1
             add $0x10, %rsi
-            inc %r9
-            cmp %r10, %r9
+            cmp %r9, %rsi
             jl .LOOP
 
-    movdqu %xmm1, (%rdi)
+    movdqa %xmm1, (%rdi)
 #endif
     ret
diff --git a/cryptonight.c b/cryptonight.c
@@ -123,7 +123,7 @@ static inline void xor_blocks_dst(const uint8_t* a, const uint8_t* b, uint8_t* d
 struct cryptonight_ctx {
     uint8_t long_state[MEMORY] __attribute((aligned(16)));
     union cn_slow_hash_state state;
-    uint8_t text[INIT_SIZE_BYTE];
+    uint8_t text[INIT_SIZE_BYTE] __attribute((aligned(16)));
     uint8_t a[AES_BLOCK_SIZE] __attribute__((aligned(16)));
     uint8_t b[AES_BLOCK_SIZE] __attribute__((aligned(16)));
     uint8_t c[AES_BLOCK_SIZE] __attribute__((aligned(16)));