Merge branch 'master' of github.com:LucasJones/cpuminer-multi

allen16881 · May 23, 2014 · 9bebbb3 · 9bebbb3
2 parents 5e287fa + adce8dc
commit 9bebbb3
Show file tree

Hide file tree

Showing 3 changed files with 25 additions and 19 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/crypto/oaes_lib.c b/crypto/oaes_lib.c
@@ -36,7 +36,9 @@ static const char _NR[] = {
 #include <stddef.h>
 #include <time.h> 
 #include <sys/timeb.h>
+#if !((defined(__FreeBSD__) && __FreeBSD__ >= 10) || defined(__APPLE__))
 #include <malloc.h>
+#endif
 #include <string.h>
 #include <stdlib.h>
 #include <stdio.h>

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)));
@@ -194,7 +194,9 @@ void cryptonight_hash_ctx(void* output, const void* input, size_t len, struct cr
 }
 
 void cryptonight_hash(void* output, const void* input, size_t len) {
-    cryptonight_hash_ctx(output, input, len, alloca(sizeof(struct cryptonight_ctx)));
+    struct cryptonight_ctx *ctx = (struct cryptonight_ctx*)malloc(sizeof(struct cryptonight_ctx));
+    cryptonight_hash_ctx(output, input, len, ctx);
+    free(ctx);
 }
 
 void cryptonight_hash_ctx_aes_ni(void* output, const void* input, size_t len, struct cryptonight_ctx* ctx) {
@@ -269,14 +271,15 @@ int scanhash_cryptonight(int thr_id, uint32_t *pdata, const uint32_t *ptarget,
     const uint32_t Htarg = ptarget[7];
     uint32_t hash[HASH_SIZE / 4] __attribute__((aligned(32)));
 
-    struct cryptonight_ctx *ctx = alloca(sizeof(struct cryptonight_ctx));
+    struct cryptonight_ctx *ctx = (struct cryptonight_ctx*)malloc(sizeof(struct cryptonight_ctx));
 
     if (aes_ni) {
         do {
             *nonceptr = ++n;
             cryptonight_hash_ctx_aes_ni(hash, pdata, 76, ctx);
             if (unlikely(hash[7] < ptarget[7])) {
                 *hashes_done = n - first_nonce + 1;
+                free(ctx);
                 return true;
             }
         } while (likely((n <= max_nonce && !work_restart[thr_id].restart)));
@@ -286,10 +289,13 @@ int scanhash_cryptonight(int thr_id, uint32_t *pdata, const uint32_t *ptarget,
             cryptonight_hash_ctx(hash, pdata, 76, ctx);
             if (unlikely(hash[7] < ptarget[7])) {
                 *hashes_done = n - first_nonce + 1;
+                free(ctx);
                 return true;
             }
         } while (likely((n <= max_nonce && !work_restart[thr_id].restart)));
     }
+
+    free(ctx);
     *hashes_done = n - first_nonce + 1;
     return 0;
 }