X-Git-Url: http://wagner.pp.ru/gitweb/?a=blobdiff_plain;f=gosthash2012.c;h=9548bfb5492318313b83e631256479bbc4d43cc8;hb=1a2714eb8f2c526517a01c90bb02ccd0e22f4c08;hp=ddd35fad45c7cf32faefc58d933f88378e098a13;hpb=31138d02d9c3698c039154b3078c16f205e25902;p=openssl-gost%2Fengine.git

diff --git a/gosthash2012.c b/gosthash2012.c
index ddd35fa..9548bfb 100644
--- a/gosthash2012.c
+++ b/gosthash2012.c
@@ -9,6 +9,13 @@
  */
 
 #include "gosthash2012.h"
+#if defined(__x86_64__) || defined(__e2k__)
+# ifdef _MSC_VER
+#  include <intrin.h>
+# else
+#  include <x86intrin.h>
+# endif
+#endif
 
 #if defined(_WIN32) || defined(_WINDOWS)
 # define INLINE __inline
@@ -57,10 +64,13 @@ static INLINE void add512(union uint512_u * RESTRICT x,
                           const union uint512_u * RESTRICT y)
 {
 #ifndef __GOST3411_BIG_ENDIAN__
-    unsigned int CF;
+    unsigned int CF = 0;
     unsigned int i;
 
-    CF = 0;
+# ifdef HAVE_ADDCARRY_U64
+    for (i = 0; i < 8; i++)
+	CF = _addcarry_u64(CF, x->QWORD[i] , y->QWORD[i], &(x->QWORD[i]));
+# else
     for (i = 0; i < 8; i++) {
 	const unsigned long long left = x->QWORD[i];
 	unsigned long long sum;
@@ -82,7 +92,8 @@ static INLINE void add512(union uint512_u * RESTRICT x,
 	    CF = (sum < left);
 	x->QWORD[i] = sum;
     }
-#else
+# endif /* !__x86_64__ */
+#else /* __GOST3411_BIG_ENDIAN__ */
     const unsigned char *yp;
     unsigned char *xp;
     unsigned int i;
@@ -96,7 +107,7 @@ static INLINE void add512(union uint512_u * RESTRICT x,
         buf = xp[i] + yp[i] + (buf >> 8);
         xp[i] = (unsigned char)buf & 0xFF;
     }
-#endif
+#endif /* __GOST3411_BIG_ENDIAN__ */
 }
 
 static void g(union uint512_u *h, const union uint512_u * RESTRICT N,
@@ -110,7 +121,7 @@ static void g(union uint512_u *h, const union uint512_u * RESTRICT N,
     LOAD(N, xmm0, xmm2, xmm4, xmm6);
     XLPS128M(h, xmm0, xmm2, xmm4, xmm6);
 
-    LOAD(m, xmm1, xmm3, xmm5, xmm7);
+    ULOAD(m, xmm1, xmm3, xmm5, xmm7);
     XLPS128R(xmm0, xmm2, xmm4, xmm6, xmm1, xmm3, xmm5, xmm7);
 
     for (i = 0; i < 11; i++)
@@ -120,12 +131,15 @@ static void g(union uint512_u *h, const union uint512_u * RESTRICT N,
     X128R(xmm0, xmm2, xmm4, xmm6, xmm1, xmm3, xmm5, xmm7);
 
     X128M(h, xmm0, xmm2, xmm4, xmm6);
-    X128M(m, xmm0, xmm2, xmm4, xmm6);
-
-    UNLOAD(h, xmm0, xmm2, xmm4, xmm6);
+    ULOAD(m, xmm1, xmm3, xmm5, xmm7);
+    X128R(xmm0, xmm2, xmm4, xmm6, xmm1, xmm3, xmm5, xmm7);
 
+    STORE(h, xmm0, xmm2, xmm4, xmm6);
+# ifndef __i386__
     /* Restore the Floating-point status on the CPU */
+    /* This is only required on MMX, but EXTRACT32 is using MMX */
     _mm_empty();
+# endif
 #else
     union uint512_u Ki, data;
     unsigned int i;
@@ -181,7 +195,6 @@ static INLINE void stage3(gost2012_hash_ctx * CTX)
 void gost2012_hash_block(gost2012_hash_ctx * CTX,
                          const unsigned char *data, size_t len)
 {
-    register size_t chunksize;
     register size_t bufsize = CTX->bufsize;
 
     if (bufsize == 0) {
@@ -194,7 +207,7 @@ void gost2012_hash_block(gost2012_hash_ctx * CTX,
     }
 
     while (len) {
-        chunksize = 64 - bufsize;
+        register size_t chunksize = 64 - bufsize;
         if (chunksize > len)
             chunksize = len;