bn_nist.c update from HEAD.

PR: 1593

1 files changed, 195 insertions, 234 deletions

diff --git a/crypto/bn/bn_nist.c b/crypto/bn/bn_nist.c
index fd15c28777..e14232fdbb 100644
--- a/crypto/bn/bn_nist.c
+++ b/crypto/bn/bn_nist.c
@@ -99,41 +99,6 @@ static const BN_ULONG _nist_p_521[] = {0xFFFFFFFF,0xFFFFFFFF,0xFFFFFFFF,
 	0xFFFFFFFF,0xFFFFFFFF,0xFFFFFFFF,0xFFFFFFFF,0xFFFFFFFF,0xFFFFFFFF,
 	0xFFFFFFFF,0xFFFFFFFF,0xFFFFFFFF,0xFFFFFFFF,0xFFFFFFFF,0xFFFFFFFF,
 	0xFFFFFFFF,0x000001FF};
-#elif BN_BITS2 == 16
-static const BN_ULONG _nist_p_192[] = {0xFFFF,0xFFFF,0xFFFF,0xFFFF,0xFFFE,
-	0xFFFF,0xFFFF,0xFFFF,0xFFFF,0xFFFF,0xFFFF,0xFFFF};
-static const BN_ULONG _nist_p_224[] = {0x0001,0x0000,0x0000,0x0000,0x0000,
-	0x0000,0xFFFF,0xFFFF,0xFFFF,0xFFFF,0xFFFF,0xFFFF,0xFFFF,0xFFFF};
-static const BN_ULONG _nist_p_256[] = {0xFFFF,0xFFFF,0xFFFF,0xFFFF,0xFFFF,
-	0xFFFF,0x0000,0x0000,0x0000,0x0000,0x0000,0x0000,0x0001,0x0000,0xFFFF,
-	0xFFFF};
-static const BN_ULONG _nist_p_384[] = {0xFFFF,0xFFFF,0x0000,0x0000,0x0000,
-	0x0000,0xFFFF,0xFFFF,0xFFFE,0xFFFF,0xFFFF,0xFFFF,0xFFFF,0xFFFF,0xFFFF,
-	0xFFFF,0xFFFF,0xFFFF,0xFFFF,0xFFFF,0xFFFF,0xFFFF,0xFFFF,0xFFFF};
-static const BN_ULONG _nist_p_521[] = {0xFFFF,0xFFFF,0xFFFF,0xFFFF,0xFFFF,
-	0xFFFF,0xFFFF,0xFFFF,0xFFFF,0xFFFF,0xFFFF,0xFFFF,0xFFFF,0xFFFF,0xFFFF,
-	0xFFFF,0xFFFF,0xFFFF,0xFFFF,0xFFFF,0xFFFF,0xFFFF,0xFFFF,0xFFFF,0xFFFF,
-	0xFFFF,0xFFFF,0xFFFF,0xFFFF,0xFFFF,0xFFFF,0xFFFF,0x01FF};
-#elif BN_BITS2 == 8
-static const BN_ULONG _nist_p_192[] = {0xFF,0xFF,0xFF,0xFF,0xFF,0xFF,0xFF,0xFF,
-	0xFE,0xFF,0xFF,0xFF,0xFF,0xFF,0xFF,0xFF,0xFF,0xFF,0xFF,0xFF,0xFF,0xFF,
-	0xFF,0xFF};
-static const BN_ULONG _nist_p_224[] = {0x01,0x00,0x00,0x00,0x00,0x00,0x00,0x00,
-	0x00,0x00,0x00,0x00,0xFF,0xFF,0xFF,0xFF,0xFF,0xFF,0xFF,0xFF,0xFF,0xFF,
-	0xFF,0xFF,0xFF,0xFF,0xFF,0xFF};
-static const BN_ULONG _nist_p_256[] = {0xFF,0xFF,0xFF,0xFF,0xFF,0xFF,0xFF,0xFF,
-	0xFF,0xFF,0xFF,0xFF,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00,
-	0x00,0x00,0x01,0x00,0x00,0x00,0xFF,0xFF,0xFF,0xFF};
-static const BN_ULONG _nist_p_384[] = {0xFF,0xFF,0xFF,0xFF,0x00,0x00,0x00,0x00,
-	0x00,0x00,0x00,0x00,0xFF,0xFF,0xFF,0xFF,0xFE,0xFF,0xFF,0xFF,0xFF,0xFF,
-	0xFF,0xFF,0xFF,0xFF,0xFF,0xFF,0xFF,0xFF,0xFF,0xFF,0xFF,0xFF,0xFF,0xFF,
-	0xFF,0xFF,0xFF,0xFF,0xFF,0xFF,0xFF,0xFF,0xFF,0xFF,0xFF,0xFF};
-static const BN_ULONG _nist_p_521[] = {0xFF,0xFF,0xFF,0xFF,0xFF,0xFF,0xFF,0xFF,
-	0xFF,0xFF,0xFF,0xFF,0xFF,0xFF,0xFF,0xFF,0xFF,0xFF,0xFF,0xFF,0xFF,0xFF,
-	0xFF,0xFF,0xFF,0xFF,0xFF,0xFF,0xFF,0xFF,0xFF,0xFF,0xFF,0xFF,0xFF,0xFF,
-	0xFF,0xFF,0xFF,0xFF,0xFF,0xFF,0xFF,0xFF,0xFF,0xFF,0xFF,0xFF,0xFF,0xFF,
-	0xFF,0xFF,0xFF,0xFF,0xFF,0xFF,0xFF,0xFF,0xFF,0xFF,0xFF,0xFF,0xFF,0xFF,
-	0xFF,0x01};
 #endif
 
 const BIGNUM *BN_get0_nist_prime_192(void)
@@ -171,17 +136,6 @@ const BIGNUM *BN_get0_nist_prime_521(void)
 	return &const_nist_521;
 	}
 
-/* some misc internal functions */
-#if BN_BITS2 != 64
-static BN_ULONG _256_data[BN_NIST_256_TOP*6];
-static int _is_set_256_data = 0;
-static void _init_256_data(void);
-
-static BN_ULONG _384_data[BN_NIST_384_TOP*8];
-static int _is_set_384_data = 0;
-static void _init_384_data(void);
-#endif
-
 #define BN_NIST_ADD_ONE(a)	while (!(*(a)=(*(a)+1)&BN_MASK2)) ++(a);
 
 static void nist_cp_bn_0(BN_ULONG *buf, BN_ULONG *a, int top, int max)
@@ -203,10 +157,10 @@ static void nist_cp_bn(BN_ULONG *buf, BN_ULONG *a, int top)
         }
 
 #if BN_BITS2 == 64
-#define bn_cp_64(to, n, from, m)	(to)[n] = (from)[m];
+#define bn_cp_64(to, n, from, m)	(to)[n] = (m>=0)?((from)[m]):0;
 #define bn_64_set_0(to, n)		(to)[n] = (BN_ULONG)0;
 /* TBD */
-#define bn_cp_32(to, n, from, m)	(to)[n] = (from)[m];
+#define bn_cp_32(to, n, from, m)	(to)[n] = (m>=0)?((from)[m]):0;
 #define bn_32_set_0(to, n)		(to)[n] = (BN_ULONG)0;
 #else
 #define bn_cp_64(to, n, from, m) \
@@ -220,26 +174,8 @@ static void nist_cp_bn(BN_ULONG *buf, BN_ULONG *a, int top)
 	bn_32_set_0(to, (n)*2+1); \
 	}
 #if BN_BITS2 == 32
-#define bn_cp_32(to, n, from, m)	(to)[n] = (from)[m];
+#define bn_cp_32(to, n, from, m)	(to)[n] = (m>=0)?((from)[m]):0;
 #define bn_32_set_0(to, n)		(to)[n] = (BN_ULONG)0;
-#elif BN_BITS2 == 16
-#define bn_cp_32(to, n, from, m) \
-	{ \
-	(to)[(n)*2]   = (from)[(m)*2];  \
-	(to)[(n)*2+1] = (from)[(m)*2+1];\
-	}
-#define bn_32_set_0(to, n) { (to)[(n)*2] = 0; (to)[(n)*2+1] = 0; }
-#elif BN_BITS2 == 8
-#define bn_cp_32(to, n, from, m) \
-	{ \
-	(to)[(n)*4]   = (from)[(m)*4];  \
-	(to)[(n)*4+1] = (from)[(m)*4+1];\
-	(to)[(n)*4+2] = (from)[(m)*4+2];\
-	(to)[(n)*4+3] = (from)[(m)*4+3];\
-	}
-#define bn_32_set_0(to, n) \
-	{ (to)[(n)*4]   = (BN_ULONG)0; (to)[(n)*4+1] = (BN_ULONG)0; \
-	  (to)[(n)*4+2] = (BN_ULONG)0; (to)[(n)*4+3] = (BN_ULONG)0; }
 #endif
 #endif /* BN_BITS2 != 64 */
 
@@ -255,9 +191,13 @@ int BN_nist_mod_192(BIGNUM *r, const BIGNUM *a, const BIGNUM *field,
 	BN_CTX *ctx)
 	{
 	int      top = a->top, i;
+	int      carry;
 	register BN_ULONG *r_d, *a_d = a->d;
 	BN_ULONG t_d[BN_NIST_192_TOP],
-	         buf[BN_NIST_192_TOP];
+	         buf[BN_NIST_192_TOP],
+		 c_d[BN_NIST_192_TOP],
+		*res;
+	size_t   mask;
 
 	i = BN_ucmp(field, a);
 	if (i == 0)
@@ -283,36 +223,28 @@ int BN_nist_mod_192(BIGNUM *r, const BIGNUM *a, const BIGNUM *field,
 
 	nist_cp_bn_0(buf, a_d + BN_NIST_192_TOP, top - BN_NIST_192_TOP, BN_NIST_192_TOP);
 
-#if defined(OPENSSL_SYS_VMS) && defined(__DECC)
-# pragma message save
-# pragma message disable BADSUBSCRIPT
-#endif
-
 	nist_set_192(t_d, buf, 0, 3, 3);
-	if (bn_add_words(r_d, r_d, t_d, BN_NIST_192_TOP))
-		bn_sub_words(r_d, r_d, _nist_p_192, BN_NIST_192_TOP);
-
+	carry = bn_add_words(r_d, r_d, t_d, BN_NIST_192_TOP);
+	mask = 0-(size_t)bn_sub_words(c_d,r_d,_nist_p_192,BN_NIST_192_TOP);
+	mask = ~mask | (0-(size_t)carry);
+	res = (BN_ULONG *)(((size_t)c_d&mask) | ((size_t)r_d&~mask));
+	
 	nist_set_192(t_d, buf, 4, 4, 0);
-	if (bn_add_words(r_d, r_d, t_d, BN_NIST_192_TOP))
-		bn_sub_words(r_d, r_d, _nist_p_192, BN_NIST_192_TOP);
-
-#if defined(OPENSSL_SYS_VMS) && defined(__DECC)
-# pragma message restore
-#endif
+	carry = bn_add_words(r_d, res, t_d, BN_NIST_192_TOP);
+	mask = 0-(size_t)bn_sub_words(c_d,r_d,_nist_p_192,BN_NIST_192_TOP);
+	mask = ~mask | (0-(size_t)carry);
+	res = (BN_ULONG *)(((size_t)c_d&mask) | ((size_t)r_d&~mask));
 
 	nist_set_192(t_d, buf, 5, 5, 5)
-	if (bn_add_words(r_d, r_d, t_d, BN_NIST_192_TOP))
-		bn_sub_words(r_d, r_d, _nist_p_192, BN_NIST_192_TOP);
+	carry = bn_add_words(r_d, res, t_d, BN_NIST_192_TOP);
+	mask = 0-(size_t)bn_sub_words(c_d,r_d,_nist_p_192,BN_NIST_192_TOP);
+	mask = ~mask | (0-(size_t)carry);
+	res = (BN_ULONG *)(((size_t)c_d&mask) | ((size_t)r_d&~mask));
 
+	nist_cp_bn(r_d, res, BN_NIST_192_TOP);
 	r->top = BN_NIST_192_TOP;
 	bn_correct_top(r);
-	if (BN_ucmp(r, field) >= 0)
-		{
-		bn_sub_words(r_d, r_d, _nist_p_192, BN_NIST_192_TOP);
-		bn_correct_top(r);
-		}
 
-	bn_check_top(r);
 	return 1;
 	}
 
@@ -332,9 +264,13 @@ int BN_nist_mod_224(BIGNUM *r, const BIGNUM *a, const BIGNUM *field,
 	{
 #if BN_BITS2 == 32
 	int	top = a->top, i;
+	int	carry;
 	BN_ULONG *r_d, *a_d = a->d;
 	BN_ULONG t_d[BN_NIST_224_TOP],
-	         buf[BN_NIST_224_TOP];
+	         buf[BN_NIST_224_TOP],
+		 c_d[BN_NIST_224_TOP],
+		*res;
+	size_t   mask;
 
 	i = BN_ucmp(field, a);
 	if (i == 0)
@@ -361,52 +297,48 @@ int BN_nist_mod_224(BIGNUM *r, const BIGNUM *a, const BIGNUM *field,
 	nist_cp_bn_0(buf, a_d + BN_NIST_224_TOP, top - BN_NIST_224_TOP, BN_NIST_224_TOP);
 
 	nist_set_224(t_d, buf, 10, 9, 8, 7, 0, 0, 0);
-	if (bn_add_words(r_d, r_d, t_d, BN_NIST_224_TOP))
-		bn_sub_words(r_d,r_d,_nist_p_224,BN_NIST_224_TOP);
+	carry = bn_add_words(r_d, r_d, t_d, BN_NIST_224_TOP);
+	mask = 0-(size_t)bn_sub_words(c_d,r_d,_nist_p_224,BN_NIST_224_TOP);
+	mask = ~mask | (0-(size_t)carry);
+	res = (BN_ULONG *)(((size_t)c_d&mask) | ((size_t)r_d&~mask));
+	
 	nist_set_224(t_d, buf, 0, 13, 12, 11, 0, 0, 0);
-	if (bn_add_words(r_d, r_d, t_d, BN_NIST_224_TOP))
-		bn_sub_words(r_d,r_d,_nist_p_224,BN_NIST_224_TOP);
+	carry = bn_add_words(r_d, res, t_d, BN_NIST_224_TOP);
+	mask = 0-(size_t)bn_sub_words(c_d,r_d,_nist_p_224,BN_NIST_224_TOP);
+	mask = ~mask | (0-(size_t)carry);
+	res = (BN_ULONG *)(((size_t)c_d&mask) | ((size_t)r_d&~mask));
+
 	nist_set_224(t_d, buf, 13, 12, 11, 10, 9, 8, 7);
-	if (bn_sub_words(r_d, r_d, t_d, BN_NIST_224_TOP))
+#if BRANCH_FREE
+	carry = bn_sub_words(r_d, res, t_d, BN_NIST_224_TOP);
+	bn_add_words(c_d,r_d,_nist_p_224,BN_NIST_224_TOP);
+	mask = 0-(size_t)carry;
+	res = (BN_ULONG *)(((size_t)c_d&mask) | ((size_t)r_d&~mask));
+#else
+	if (bn_sub_words(r_d, res, t_d, BN_NIST_224_TOP))
 		bn_add_words(r_d,r_d,_nist_p_224,BN_NIST_224_TOP);
+#endif
 	nist_set_224(t_d, buf, 0, 0, 0, 0, 13, 12, 11);
+#if BRANCH_FREE
+	carry = bn_sub_words(r_d, res, t_d, BN_NIST_224_TOP);
+	bn_add_words(c_d,r_d,_nist_p_224,BN_NIST_224_TOP);
+	mask = 0-(size_t)carry;
+	res = (BN_ULONG *)(((size_t)c_d&mask) | ((size_t)r_d&~mask));
+
+	nist_cp_bn(r_d, res, BN_NIST_224_TOP);
+#else
 	if (bn_sub_words(r_d, r_d, t_d, BN_NIST_224_TOP))
 		bn_add_words(r_d,r_d,_nist_p_224,BN_NIST_224_TOP);
-
+#endif
 	r->top = BN_NIST_224_TOP;
 	bn_correct_top(r);
-	if (BN_ucmp(r, field) >= 0)
-		{
-		bn_sub_words(r_d, r_d, _nist_p_224, BN_NIST_224_TOP);
-		bn_correct_top(r);
-		}
-	bn_check_top(r);
+
 	return 1;
-#else
+#else	/* BN_BITS!=32 */
 	return 0;
 #endif
 	}
 
-#if BN_BITS2 == 32
-static void _init_256_data(void)
-	{
-	int	i;
-	BN_ULONG *tmp1 = _256_data;
-	const BN_ULONG *tmp2 = tmp1;
-
-	memcpy(tmp1, _nist_p_256, BN_NIST_256_TOP * sizeof(BN_ULONG));
-	tmp1 += BN_NIST_256_TOP;
-
-	for (i=0; i<5; i++)
-		{
-		bn_add_words(tmp1, _nist_p_256, tmp2, BN_NIST_256_TOP);
-		tmp2  = tmp1;
-		tmp1 += BN_NIST_256_TOP;
-		}
-	_is_set_256_data = 1;
-	}
-#endif
-
 #define nist_set_256(to, from, a1, a2, a3, a4, a5, a6, a7, a8) \
 	{ \
 	if (a8 != 0) bn_cp_32(to, 0, from, (a8) - 8) else bn_32_set_0(to, 0)\
@@ -427,19 +359,11 @@ int BN_nist_mod_256(BIGNUM *r, const BIGNUM *a, const BIGNUM *field,
 	int	carry = 0;
 	register BN_ULONG *a_d = a->d, *r_d;
 	BN_ULONG t_d[BN_NIST_256_TOP],
-	         t_d2[BN_NIST_256_TOP],
-	         buf[BN_NIST_256_TOP];
+	         buf[BN_NIST_256_TOP],
+		 c_d[BN_NIST_256_TOP],
+		*res;
+	size_t   mask;
 
-	if (!_is_set_256_data)
-		{
-		CRYPTO_w_lock(CRYPTO_LOCK_BN);
-		
-		if (!_is_set_256_data)
-			_init_256_data();
-		
-		CRYPTO_w_unlock(CRYPTO_LOCK_BN);
-		}
-	
 	i = BN_ucmp(field, a);
 	if (i == 0)
 		{
@@ -467,86 +391,91 @@ int BN_nist_mod_256(BIGNUM *r, const BIGNUM *a, const BIGNUM *field,
 	/*S1*/
 	nist_set_256(t_d, buf, 15, 14, 13, 12, 11, 0, 0, 0);
 	/*S2*/
-	nist_set_256(t_d2,buf, 0, 15, 14, 13, 12, 0, 0, 0);
-	carry = bn_add_words(t_d, t_d, t_d2, BN_NIST_256_TOP);
-	/* left shift */
-		{
-		register BN_ULONG *ap,t,c;
-		ap = t_d;
-		c=0;
-		for (i = BN_NIST_256_TOP; i != 0; --i)
-			{
-			t= *ap;
-			*(ap++)=((t<<1)|c)&BN_MASK2;
-			c=(t & BN_TBIT)?1:0;
-			}
-		carry <<= 1;
-		carry  |= c;
-		}
+	nist_set_256(c_d,buf, 0, 15, 14, 13, 12, 0, 0, 0);
+	carry = bn_add_words(t_d, t_d, c_d, BN_NIST_256_TOP);
+	mask = 0-(size_t)bn_sub_words(c_d,t_d,_nist_p_256,BN_NIST_256_TOP);
+	mask = ~mask | (0-(size_t)carry);
+	res = (BN_ULONG *)(((size_t)c_d&mask) | ((size_t)t_d&~mask));
+
+	carry = bn_add_words(t_d, res, res, BN_NIST_256_TOP);
+	mask = 0-(size_t)bn_sub_words(c_d,t_d,_nist_p_256,BN_NIST_256_TOP);
+	mask = ~mask | (0-(size_t)carry);
+	res = (BN_ULONG *)(((size_t)c_d&mask) | ((size_t)t_d&~mask));
+
+	carry = bn_add_words(r_d, r_d, res, BN_NIST_256_TOP);
+	mask = 0-(size_t)bn_sub_words(c_d,r_d,_nist_p_256,BN_NIST_256_TOP);
+	mask = ~mask | (0-(size_t)carry);
+	res = (BN_ULONG *)(((size_t)c_d&mask) | ((size_t)r_d&~mask));
 
-	carry += bn_add_words(r_d, r_d, t_d, BN_NIST_256_TOP);
 	/*S3*/
 	nist_set_256(t_d, buf, 15, 14, 0, 0, 0, 10, 9, 8);
-	carry += bn_add_words(r_d, r_d, t_d, BN_NIST_256_TOP);
+	carry = bn_add_words(r_d, res, t_d, BN_NIST_256_TOP);
+	mask = 0-(size_t)bn_sub_words(c_d,r_d,_nist_p_256,BN_NIST_256_TOP);
+	mask = ~mask | (0-(size_t)carry);
+	res = (BN_ULONG *)(((size_t)c_d&mask) | ((size_t)r_d&~mask));
+
 	/*S4*/
 	nist_set_256(t_d, buf, 8, 13, 15, 14, 13, 11, 10, 9);
-	carry += bn_add_words(r_d, r_d, t_d, BN_NIST_256_TOP);
-
-	if (carry)
-		bn_sub_words(r_d, r_d, _256_data + BN_NIST_256_TOP *
-				(carry-1), BN_NIST_256_TOP);
+	carry = bn_add_words(r_d, res, t_d, BN_NIST_256_TOP);
+	mask = 0-(size_t)bn_sub_words(c_d,r_d,_nist_p_256,BN_NIST_256_TOP);
+	mask = ~mask | (0-(size_t)carry);
+	res = (BN_ULONG *)(((size_t)c_d&mask) | ((size_t)r_d&~mask));
 
 	/*D1*/
 	nist_set_256(t_d, buf, 10, 8, 0, 0, 0, 13, 12, 11);
-	carry = bn_sub_words(r_d, r_d, t_d, BN_NIST_256_TOP);
+#if BRANCH_FREE
+	carry = bn_sub_words(r_d, res, t_d, BN_NIST_256_TOP);
+	bn_add_words(c_d,r_d,_nist_p_256,BN_NIST_256_TOP);
+	mask = 0-(size_t)carry;
+	res = (BN_ULONG *)(((size_t)c_d&mask) | ((size_t)r_d&~mask));	
+#else
+	if (bn_sub_words(r_d, res, t_d, BN_NIST_256_TOP))
+		bn_add_words(r_d,r_d,_nist_p_256,BN_NIST_256_TOP);
+#endif
 	/*D2*/
 	nist_set_256(t_d, buf, 11, 9, 0, 0, 15, 14, 13, 12);
-	carry += bn_sub_words(r_d, r_d, t_d, BN_NIST_256_TOP);
+#if BRANCH_FREE
+	carry = bn_sub_words(r_d, res, t_d, BN_NIST_256_TOP);
+	bn_add_words(c_d,r_d,_nist_p_256,BN_NIST_256_TOP);
+	mask = 0-(size_t)carry;
+	res = (BN_ULONG *)(((size_t)c_d&mask) | ((size_t)r_d&~mask));	
+#else
+	if (bn_sub_words(r_d, r_d, t_d, BN_NIST_256_TOP))
+		bn_add_words(r_d,r_d,_nist_p_256,BN_NIST_256_TOP);
+#endif
 	/*D3*/
 	nist_set_256(t_d, buf, 12, 0, 10, 9, 8, 15, 14, 13);
-	carry += bn_sub_words(r_d, r_d, t_d, BN_NIST_256_TOP);
+#if BRANCH_FREE
+	carry = bn_sub_words(r_d, res, t_d, BN_NIST_256_TOP);
+	bn_add_words(c_d,r_d,_nist_p_256,BN_NIST_256_TOP);
+	mask = 0-(size_t)carry;
+	res = (BN_ULONG *)(((size_t)c_d&mask) | ((size_t)r_d&~mask));	
+#else
+	if (bn_sub_words(r_d, r_d, t_d, BN_NIST_256_TOP))
+		bn_add_words(r_d,r_d,_nist_p_256,BN_NIST_256_TOP);
+#endif
 	/*D4*/
 	nist_set_256(t_d, buf, 13, 0, 11, 10, 9, 0, 15, 14);
-	carry += bn_sub_words(r_d, r_d, t_d, BN_NIST_256_TOP);
-	
-	if (carry)
-		bn_add_words(r_d, r_d, _256_data + BN_NIST_256_TOP *
-				(carry-1), BN_NIST_256_TOP);
+#if BRANCH_FREE
+	carry = bn_sub_words(r_d, res, t_d, BN_NIST_256_TOP);
+	bn_add_words(c_d,r_d,_nist_p_256,BN_NIST_256_TOP);
+	mask = 0-(size_t)carry;
+	res = (BN_ULONG *)(((size_t)c_d&mask) | ((size_t)r_d&~mask));	
 
+	nist_cp_bn(r_d, res, BN_NIST_384_TOP);
+#else
+	if (bn_sub_words(r_d, r_d, t_d, BN_NIST_256_TOP))
+		bn_add_words(r_d,r_d,_nist_p_256,BN_NIST_256_TOP);
+#endif
 	r->top = BN_NIST_256_TOP;
 	bn_correct_top(r);
-	if (BN_ucmp(r, field) >= 0)
-		{
-		bn_sub_words(r_d, r_d, _nist_p_256, BN_NIST_256_TOP);
-		bn_correct_top(r);
-		}
-	bn_check_top(r);
+
 	return 1;
-#else
+#else	/* BN_BITS!=32 */
 	return 0;
 #endif
 	}
 
-#if BN_BITS2 == 32
-static void _init_384_data(void)
-	{
-	int	i;
-	BN_ULONG *tmp1 = _384_data;
-	const BN_ULONG *tmp2 = tmp1;
-
-	memcpy(tmp1, _nist_p_384, BN_NIST_384_TOP * sizeof(BN_ULONG));
-	tmp1 += BN_NIST_384_TOP;
-
-	for (i=0; i<7; i++)
-		{
-		bn_add_words(tmp1, _nist_p_384, tmp2, BN_NIST_384_TOP);
-		tmp2  = tmp1;
-		tmp1 += BN_NIST_384_TOP;
-		}
-	_is_set_384_data = 1;
-	}
-#endif
-
 #define nist_set_384(to,from,a1,a2,a3,a4,a5,a6,a7,a8,a9,a10,a11,a12) \
 	{ \
 	if (a12 != 0) bn_cp_32(to, 0, from,  (a12) - 12) else bn_32_set_0(to, 0)\
@@ -571,17 +500,10 @@ int BN_nist_mod_384(BIGNUM *r, const BIGNUM *a, const BIGNUM *field,
 	int	carry = 0;
 	register BN_ULONG *r_d, *a_d = a->d;
 	BN_ULONG t_d[BN_NIST_384_TOP],
-	         buf[BN_NIST_384_TOP];
-
-	if (!_is_set_384_data)
-		{
-		CRYPTO_w_lock(CRYPTO_LOCK_BN);
-		
-		if (!_is_set_384_data)
-			_init_384_data();
-
-		CRYPTO_w_unlock(CRYPTO_LOCK_BN);
-		}
+	         buf[BN_NIST_384_TOP],
+		 c_d[BN_NIST_384_TOP],
+		*res;
+	size_t	 mask;
 
 	i = BN_ucmp(field, a);
 	if (i == 0)
@@ -624,50 +546,93 @@ int BN_nist_mod_384(BIGNUM *r, const BIGNUM *a, const BIGNUM *field,
 		}
 	carry = bn_add_words(r_d+(128/BN_BITS2), r_d+(128/BN_BITS2), 
 		t_d, BN_NIST_256_TOP);
+	/*
+	 * we need if (result>=modulus) subtract(result,modulus);
+	 * in n-bit space this can be expressed as
+	 * if (carry || result>=modulus) subtract(result,modulus);
+	 * the catch is that comparison implies subtraction and
+	 * therefore one can write tmp=subtract(result,modulus);
+	 * and then if(carry || !borrow) result=tmp; this's what
+	 * happens below, but without explicit if:-) a.
+	 */
+	mask = 0-(size_t)bn_sub_words(c_d,r_d,_nist_p_384,BN_NIST_384_TOP);
+	mask = ~mask | (0-(size_t)carry);
+	res = (BN_ULONG *)(((size_t)c_d&mask) | ((size_t)r_d&~mask));
+
 	/*S2 */
-	carry += bn_add_words(r_d, r_d, buf, BN_NIST_384_TOP);
+	carry = bn_add_words(r_d, res, buf, BN_NIST_384_TOP);
+	mask = 0-(size_t)bn_sub_words(c_d,r_d,_nist_p_384,BN_NIST_384_TOP);
+	mask = ~mask | (0-(size_t)carry);
+	res = (BN_ULONG *)(((size_t)c_d&mask) | ((size_t)r_d&~mask));
+
 	/*S3*/
 	nist_set_384(t_d,buf,20,19,18,17,16,15,14,13,12,23,22,21);
-	carry += bn_add_words(r_d, r_d, t_d, BN_NIST_384_TOP);
+	carry = bn_add_words(r_d, res, t_d, BN_NIST_384_TOP);
+	mask = 0-(size_t)bn_sub_words(c_d,r_d,_nist_p_384,BN_NIST_384_TOP);
+	mask = ~mask | (0-(size_t)carry);
+	res = (BN_ULONG *)(((size_t)c_d&mask) | ((size_t)r_d&~mask));
+
 	/*S4*/
 	nist_set_384(t_d,buf,19,18,17,16,15,14,13,12,20,0,23,0);
-	carry += bn_add_words(r_d, r_d, t_d, BN_NIST_384_TOP);
+	carry = bn_add_words(r_d, res, t_d, BN_NIST_384_TOP);
+	mask = 0-(size_t)bn_sub_words(c_d,r_d,_nist_p_384,BN_NIST_384_TOP);
+	mask = ~mask | (0-(size_t)carry);
+	res = (BN_ULONG *)(((size_t)c_d&mask) | ((size_t)r_d&~mask));
+
 	/*S5*/
-	nist_set_256(t_d, buf, 0, 0, 0, 0, 23-4, 22-4, 21-4, 20-4);
-	carry += bn_add_words(r_d+(128/BN_BITS2), r_d+(128/BN_BITS2), 
-		t_d, BN_NIST_256_TOP);
+	nist_set_384(t_d, buf,0,0,0,0,23,22,21,20,0,0,0,0);
+	carry = bn_add_words(r_d, res, t_d, BN_NIST_384_TOP);
+	mask = 0-(size_t)bn_sub_words(c_d,r_d,_nist_p_384,BN_NIST_384_TOP);
+	mask = ~mask | (0-(size_t)carry);
+	res = (BN_ULONG *)(((size_t)c_d&mask) | ((size_t)r_d&~mask));
+
 	/*S6*/
 	nist_set_384(t_d,buf,0,0,0,0,0,0,23,22,21,0,0,20);
-	carry += bn_add_words(r_d, r_d, t_d, BN_NIST_384_TOP);
-
-	if (carry)
-		bn_sub_words(r_d, r_d, _384_data + BN_NIST_384_TOP *
-				(carry-1), BN_NIST_384_TOP);
+	carry = bn_add_words(r_d, res, t_d, BN_NIST_384_TOP);
+	mask = 0-(size_t)bn_sub_words(c_d,r_d,_nist_p_384,BN_NIST_384_TOP);
+	mask = ~mask | (0-(size_t)carry);
+	res = (BN_ULONG *)(((size_t)c_d&mask) | ((size_t)r_d&~mask));
 
 	/*D1*/
 	nist_set_384(t_d,buf,22,21,20,19,18,17,16,15,14,13,12,23);
-	carry = bn_sub_words(r_d, r_d, t_d, BN_NIST_384_TOP);
+#if BRANCH_FREE
+	carry = bn_sub_words(r_d, res, t_d, BN_NIST_384_TOP);
+	bn_add_words(c_d,r_d,_nist_p_384,BN_NIST_384_TOP);
+	mask = 0-(size_t)carry;
+	res = (BN_ULONG *)(((size_t)c_d&mask) | ((size_t)r_d&~mask));
+#else
+	if (bn_sub_words(r_d, res, t_d, BN_NIST_384_TOP))
+		bn_add_words(r_d,r_d,_nist_p_384,BN_NIST_384_TOP);
+#endif
 	/*D2*/
 	nist_set_384(t_d,buf,0,0,0,0,0,0,0,23,22,21,20,0);
-	carry += bn_sub_words(r_d, r_d, t_d, BN_NIST_384_TOP);
+#if BRANCH_FREE
+	carry = bn_sub_words(r_d, res, t_d, BN_NIST_384_TOP);
+	bn_add_words(c_d,r_d,_nist_p_384,BN_NIST_384_TOP);
+	mask = 0-(size_t)carry;
+	res = (BN_ULONG *)(((size_t)c_d&mask) | ((size_t)r_d&~mask));
+#else
+	if (bn_sub_words(r_d, r_d, t_d, BN_NIST_384_TOP))
+		bn_add_words(r_d,r_d,_nist_p_384,BN_NIST_384_TOP);
+#endif
 	/*D3*/
 	nist_set_384(t_d,buf,0,0,0,0,0,0,0,23,23,0,0,0);
-	carry += bn_sub_words(r_d, r_d, t_d, BN_NIST_384_TOP);
-
-	if (carry)
-		bn_add_words(r_d, r_d, _384_data + BN_NIST_384_TOP *
-				(carry-1), BN_NIST_384_TOP);
+#if BRANCH_FREE
+	carry = bn_sub_words(r_d, res, t_d, BN_NIST_384_TOP);
+	bn_add_words(c_d,r_d,_nist_p_384,BN_NIST_384_TOP);
+	mask = 0-(size_t)carry;
+	res = (BN_ULONG *)(((size_t)c_d&mask) | ((size_t)r_d&~mask));
 
+	nist_cp_bn(r_d, res, BN_NIST_384_TOP);
+#else
+	if (bn_sub_words(r_d, r_d, t_d, BN_NIST_384_TOP))
+		bn_add_words(r_d,r_d,_nist_p_384,BN_NIST_384_TOP);
+#endif
 	r->top = BN_NIST_384_TOP;
 	bn_correct_top(r);
-	if (BN_ucmp(r, field) >= 0)
-		{
-		bn_sub_words(r_d, r_d, _nist_p_384, BN_NIST_384_TOP);
-		bn_correct_top(r);
-		}
-	bn_check_top(r);
+
 	return 1;
-#else
+#else	/* BN_BITS!=32 */
 	return 0;
 #endif
 	}
@@ -679,10 +644,6 @@ int BN_nist_mod_521(BIGNUM *r, const BIGNUM *a, const BIGNUM *field,
 #define BN_NIST_521_TOP_MASK	(BN_ULONG)0x1FF
 #elif BN_BITS2 == 32
 #define BN_NIST_521_TOP_MASK	(BN_ULONG)0x1FF
-#elif BN_BITS2 == 16
-#define BN_NIST_521_TOP_MASK	(BN_ULONG)0x1FF
-#elif BN_BITS2 == 8
-#define BN_NIST_521_TOP_MASK	(BN_ULONG)0x1
 #endif
 	int	top, ret = 0;
 	BN_ULONG *r_d;