gosthash2012: Provide cmake test for _addcarry_u64

[openssl-gost/engine.git] / gosthash2012.c

diff --git a/gosthash2012.c b/gosthash2012.c
index ddd35fad45c7cf32faefc58d933f88378e098a13..5a4c64e488d0830beefa34c36c7192114d8ba99f 100644 (file)
--- a/gosthash2012.c
+++ b/gosthash2012.c
@@ -9,6 +9,13 @@
  */
 
 #include "gosthash2012.h"
+#ifdef __x86_64__
+# 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 +131,10 @@ 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);
 
-    /* Restore the Floating-point status on the CPU */
-    _mm_empty();
+    STORE(h, xmm0, xmm2, xmm4, xmm6);
 #else
     union uint512_u Ki, data;
     unsigned int i;