X-Git-Url: http://wagner.pp.ru/gitweb/?a=blobdiff_plain;f=gosthash2012.c;h=9548bfb5492318313b83e631256479bbc4d43cc8;hb=3b31dea89630ecd3f741171ed3e1d4259adabf0e;hp=d6cde2154a7143371057a530473c057c3f2dbed9;hpb=b3dd924c96575efc13392530295e0c30254780b1;p=openssl-gost%2Fengine.git

diff --git a/gosthash2012.c b/gosthash2012.c
index d6cde21..9548bfb 100644
--- a/gosthash2012.c
+++ b/gosthash2012.c
@@ -9,9 +9,12 @@
  */
 
 #include "gosthash2012.h"
-#ifdef __x86_64__
-# include <immintrin.h>
-# include <x86intrin.h>
+#if defined(__x86_64__) || defined(__e2k__)
+# ifdef _MSC_VER
+#  include <intrin.h>
+# else
+#  include <x86intrin.h>
+# endif
 #endif
 
 #if defined(_WIN32) || defined(_WINDOWS)
@@ -64,7 +67,7 @@ static INLINE void add512(union uint512_u * RESTRICT x,
     unsigned int CF = 0;
     unsigned int i;
 
-# ifdef __x86_64__
+# ifdef HAVE_ADDCARRY_U64
     for (i = 0; i < 8; i++)
 	CF = _addcarry_u64(CF, x->QWORD[i] , y->QWORD[i], &(x->QWORD[i]));
 # else
@@ -118,7 +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++)
@@ -128,12 +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;
@@ -189,7 +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) {
@@ -202,7 +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;