Fix SHA512 and optimize BN for mingw64.

[openssl.git] / crypto / bn / bn_lcl.h

diff --git a/crypto/bn/bn_lcl.h b/crypto/bn/bn_lcl.h
index 45e19221aafe2280bcb57432ca8d87ad077b8b4a..8e5e98e3f2b87fd80a1d33511df3ba495e60414a 100644 (file)
--- a/crypto/bn/bn_lcl.h
+++ b/crypto/bn/bn_lcl.h
@@ -163,6 +163,45 @@ extern "C" {
 
 
 
+/* BN_mod_exp_mont_conttime is based on the assumption that the
+ * L1 data cache line width of the target processor is at least
+ * the following value.
+ */
+#define MOD_EXP_CTIME_MIN_CACHE_LINE_WIDTH     ( 64 )
+#define MOD_EXP_CTIME_MIN_CACHE_LINE_MASK      (MOD_EXP_CTIME_MIN_CACHE_LINE_WIDTH - 1)
+
+/* Window sizes optimized for fixed window size modular exponentiation
+ * algorithm (BN_mod_exp_mont_consttime).
+ *
+ * To achieve the security goals of BN_mode_exp_mont_consttime, the
+ * maximum size of the window must not exceed
+ * log_2(MOD_EXP_CTIME_MIN_CACHE_LINE_WIDTH). 
+ *
+ * Window size thresholds are defined for cache line sizes of 32 and 64,
+ * cache line sizes where log_2(32)=5 and log_2(64)=6 respectively. A
+ * window size of 7 should only be used on processors that have a 128
+ * byte or greater cache line size.
+ */
+#if MOD_EXP_CTIME_MIN_CACHE_LINE_WIDTH == 64
+
+#  define BN_window_bits_for_ctime_exponent_size(b) \
+               ((b) > 937 ? 6 : \
+                (b) > 306 ? 5 : \
+                (b) >  89 ? 4 : \
+                (b) >  22 ? 3 : 1)
+#  define BN_MAX_WINDOW_BITS_FOR_CTIME_EXPONENT_SIZE   (6)
+
+#elif MOD_EXP_CTIME_MIN_CACHE_LINE_WIDTH == 32
+
+#  define BN_window_bits_for_ctime_exponent_size(b) \
+               ((b) > 306 ? 5 : \
+                (b) >  89 ? 4 : \
+                (b) >  22 ? 3 : 1)
+#  define BN_MAX_WINDOW_BITS_FOR_CTIME_EXPONENT_SIZE   (5)
+
+#endif
+
+
 /* Pentium pro 16,16,16,32,64 */
 /* Alpha       16,16,16,16.64 */
 #define BN_MULL_SIZE_NORMAL                    (16) /* 32 */
@@ -216,7 +255,8 @@ extern "C" {
             : "r"(a), "r"(b));         \
        ret;                    })
 #  endif       /* compiler */
-# elif defined(__x86_64) && defined(SIXTY_FOUR_BIT_LONG)
+# elif (defined(__x86_64) || defined(__x86_64__)) && \
+       (defined(SIXTY_FOUR_BIT_LONG) || defined(SIXTY_FOUR_BIT))
 #  if defined(__GNUC__)
 #   define BN_UMULT_HIGH(a,b)  ({      \
        register BN_ULONG ret,discard;  \
@@ -231,6 +271,15 @@ extern "C" {
                : "a"(a),"g"(b)         \
                : "cc");
 #  endif
+# elif (defined(_M_AMD64) || defined(_M_X64)) && defined(SIXTY_FOUR_BIT)
+#  if defined(_MSC_VER) && _MSC_VER>=1400
+    unsigned __int64 __umulh   (unsigned __int64 a,unsigned __int64 b);
+    unsigned __int64 _umul128  (unsigned __int64 a,unsigned __int64 b,
+                                unsigned __int64 *h);
+#   pragma intrinsic(__umulh,_umul128)
+#   define BN_UMULT_HIGH(a,b)          __umulh((a),(b))
+#   define BN_UMULT_LOHI(low,high,a,b) ((low)=_umul128((a),(b),&(high)))
+#  endif
 # endif                /* cpu */
 #endif         /* OPENSSL_NO_ASM */
 
@@ -274,6 +323,33 @@ extern "C" {
        (r1)=Hw(t); \
        }
 
+#elif defined(BN_UMULT_LOHI)
+#define mul_add(r,a,w,c) {             \
+       BN_ULONG high,low,ret,tmp=(a);  \
+       ret =  (r);                     \
+       BN_UMULT_LOHI(low,high,w,tmp);  \
+       ret += (c);                     \
+       (c) =  (ret<(c))?1:0;           \
+       (c) += high;                    \
+       ret += low;                     \
+       (c) += (ret<low)?1:0;           \
+       (r) =  ret;                     \
+       }
+
+#define mul(r,a,w,c)   {               \
+       BN_ULONG high,low,ret,ta=(a);   \
+       BN_UMULT_LOHI(low,high,w,ta);   \
+       ret =  low + (c);               \
+       (c) =  high;                    \
+       (c) += (ret<low)?1:0;           \
+       (r) =  ret;                     \
+       }
+
+#define sqr(r0,r1,a)   {               \
+       BN_ULONG tmp=(a);               \
+       BN_UMULT_LOHI(r0,r1,tmp,tmp);   \
+       }
+
 #elif defined(BN_UMULT_HIGH)
 #define mul_add(r,a,w,c) {             \
        BN_ULONG high,low,ret,tmp=(a);  \
@@ -406,6 +482,7 @@ BN_ULONG bn_add_part_words(BN_ULONG *r, const BN_ULONG *a, const BN_ULONG *b,
        int cl, int dl);
 BN_ULONG bn_sub_part_words(BN_ULONG *r, const BN_ULONG *a, const BN_ULONG *b,
        int cl, int dl);
+int bn_mul_mont(BN_ULONG *rp, const BN_ULONG *ap, const BN_ULONG *bp, const BN_ULONG *np,const BN_ULONG *n0, int num);
 
 #ifdef  __cplusplus
 }