Correctly connects to millipede.

Location keyword on android.cfg isn't supported, EIP corresponding code
has been commented out. I think we should support it in ics-openvpn, so
that we can show the location instead of the server name.

I've updated all opensssl, openvpn, etc. subprojects from rev 813 of
ics-openvpn, and jni too.

1 files changed, 326 insertions, 61 deletions

diff --git a/app/openssl/crypto/bn/bn_nist.c b/app/openssl/crypto/bn/bn_nist.c
index c6de0326..e22968d4 100644
--- a/app/openssl/crypto/bn/bn_nist.c
+++ b/app/openssl/crypto/bn/bn_nist.c
@@ -286,26 +286,25 @@ const BIGNUM *BN_get0_nist_prime_521(void)
 	}
 
 
-static void nist_cp_bn_0(BN_ULONG *buf, BN_ULONG *a, int top, int max)
+static void nist_cp_bn_0(BN_ULONG *dst, const BN_ULONG *src, int top, int max)
 	{
 	int i;
-	BN_ULONG *_tmp1 = (buf), *_tmp2 = (a);
 
 #ifdef BN_DEBUG
 	OPENSSL_assert(top <= max);
 #endif
-	for (i = (top); i != 0; i--)
-		*_tmp1++ = *_tmp2++;
-	for (i = (max) - (top); i != 0; i--)
-		*_tmp1++ = (BN_ULONG) 0;
+	for (i = 0; i < top; i++)
+		dst[i] = src[i];
+	for (; i < max; i++)
+		dst[i] = 0;
 	}
 
-static void nist_cp_bn(BN_ULONG *buf, BN_ULONG *a, int top)
+static void nist_cp_bn(BN_ULONG *dst, const BN_ULONG *src, int top)
 	{ 
 	int i;
-	BN_ULONG *_tmp1 = (buf), *_tmp2 = (a);
-	for (i = (top); i != 0; i--)
-		*_tmp1++ = *_tmp2++;
+
+	for (i = 0; i < top; i++)
+		dst[i] = src[i];
 	}
 
 #if BN_BITS2 == 64
@@ -319,6 +318,13 @@ static void nist_cp_bn(BN_ULONG *buf, BN_ULONG *a, int top)
 						:(to[(n)/2] =((m)&1)?(from[(m)/2]>>32):(from[(m)/2]&BN_MASK2l)))
 #define bn_32_set_0(to, n)		(((n)&1)?(to[(n)/2]&=BN_MASK2l):(to[(n)/2]=0));
 #define bn_cp_32(to,n,from,m)		((m)>=0)?bn_cp_32_naked(to,n,from,m):bn_32_set_0(to,n)
+# if defined(L_ENDIAN)
+#  if defined(__arch64__)
+#   define NIST_INT64 long
+#  else
+#   define NIST_INT64 long long
+#  endif
+# endif
 #else
 #define bn_cp_64(to, n, from, m) \
 	{ \
@@ -330,13 +336,15 @@ static void nist_cp_bn(BN_ULONG *buf, BN_ULONG *a, int top)
 	bn_32_set_0(to, (n)*2); \
 	bn_32_set_0(to, (n)*2+1); \
 	}
-#if BN_BITS2 == 32
 #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;
-#endif
+# if defined(_WIN32) && !defined(__GNUC__)
+#  define NIST_INT64 __int64
+# elif defined(BN_LLONG)
+#  define NIST_INT64 long long
+# endif
 #endif /* BN_BITS2 != 64 */
 
-
 #define nist_set_192(to, from, a1, a2, a3) \
 	{ \
 	bn_cp_64(to, 0, from, (a3) - 3) \
@@ -350,9 +358,11 @@ int BN_nist_mod_192(BIGNUM *r, const BIGNUM *a, const BIGNUM *field,
 	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],
-		 c_d[BN_NIST_192_TOP],
+	union	{
+		BN_ULONG	bn[BN_NIST_192_TOP];
+		unsigned int	ui[BN_NIST_192_TOP*sizeof(BN_ULONG)/sizeof(unsigned int)];
+		} buf;
+	BN_ULONG c_d[BN_NIST_192_TOP],
 		*res;
 	PTR_SIZE_INT mask;
 	static const BIGNUM _bignum_nist_p_192_sqr = {
@@ -385,15 +395,48 @@ int BN_nist_mod_192(BIGNUM *r, const BIGNUM *a, const BIGNUM *field,
 	else
 		r_d = a_d;
 
-	nist_cp_bn_0(buf, a_d + BN_NIST_192_TOP, top - BN_NIST_192_TOP, BN_NIST_192_TOP);
+	nist_cp_bn_0(buf.bn, a_d + BN_NIST_192_TOP, top - BN_NIST_192_TOP, BN_NIST_192_TOP);
+
+#if defined(NIST_INT64)
+	{
+	NIST_INT64		acc;	/* accumulator */
+	unsigned int		*rp=(unsigned int *)r_d;
+	const unsigned int	*bp=(const unsigned int *)buf.ui;
+
+	acc  = rp[0];	acc += bp[3*2-6];
+			acc += bp[5*2-6]; rp[0] = (unsigned int)acc; acc >>= 32;
+
+	acc += rp[1];	acc += bp[3*2-5];
+			acc += bp[5*2-5]; rp[1] = (unsigned int)acc; acc >>= 32;
+
+	acc += rp[2];	acc += bp[3*2-6];
+			acc += bp[4*2-6];
+			acc += bp[5*2-6]; rp[2] = (unsigned int)acc; acc >>= 32;
+
+	acc += rp[3];	acc += bp[3*2-5];
+			acc += bp[4*2-5];
+			acc += bp[5*2-5]; rp[3] = (unsigned int)acc; acc >>= 32;
+
+	acc += rp[4];	acc += bp[4*2-6];
+			acc += bp[5*2-6]; rp[4] = (unsigned int)acc; acc >>= 32;
+
+	acc += rp[5];	acc += bp[4*2-5];
+			acc += bp[5*2-5]; rp[5] = (unsigned int)acc;
+
+	carry = (int)(acc>>32);
+	}
+#else
+	{
+	BN_ULONG t_d[BN_NIST_192_TOP];
 
-	nist_set_192(t_d, buf, 0, 3, 3);
+	nist_set_192(t_d, buf.bn, 0, 3, 3);
 	carry = (int)bn_add_words(r_d, r_d, t_d, BN_NIST_192_TOP);
-	nist_set_192(t_d, buf, 4, 4, 0);
+	nist_set_192(t_d, buf.bn, 4, 4, 0);
 	carry += (int)bn_add_words(r_d, r_d, t_d, BN_NIST_192_TOP);
-	nist_set_192(t_d, buf, 5, 5, 5)
+	nist_set_192(t_d, buf.bn, 5, 5, 5)
 	carry += (int)bn_add_words(r_d, r_d, t_d, BN_NIST_192_TOP);
-
+	}
+#endif
 	if (carry > 0)
 		carry = (int)bn_sub_words(r_d,r_d,_nist_p_192[carry-1],BN_NIST_192_TOP);
 	else
@@ -407,8 +450,9 @@ int BN_nist_mod_192(BIGNUM *r, const BIGNUM *a, const BIGNUM *field,
 	 */
 	mask  = 0-(PTR_SIZE_INT)bn_sub_words(c_d,r_d,_nist_p_192[0],BN_NIST_192_TOP);
 	mask &= 0-(PTR_SIZE_INT)carry;
+	res   = c_d;
 	res   = (BN_ULONG *)
-	 (((PTR_SIZE_INT)c_d&~mask) | ((PTR_SIZE_INT)r_d&mask));
+	 (((PTR_SIZE_INT)res&~mask) | ((PTR_SIZE_INT)r_d&mask));
 	nist_cp_bn(r_d, res, BN_NIST_192_TOP);
 	r->top = BN_NIST_192_TOP;
 	bn_correct_top(r);
@@ -435,9 +479,11 @@ int BN_nist_mod_224(BIGNUM *r, const BIGNUM *a, const BIGNUM *field,
 	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],
-		 c_d[BN_NIST_224_TOP],
+	union	{
+		BN_ULONG	bn[BN_NIST_224_TOP];
+		unsigned int	ui[BN_NIST_224_TOP*sizeof(BN_ULONG)/sizeof(unsigned int)];
+		} buf;
+	BN_ULONG c_d[BN_NIST_224_TOP],
 		*res;
 	PTR_SIZE_INT mask;
 	union { bn_addsub_f f; PTR_SIZE_INT p; } u;
@@ -474,26 +520,68 @@ int BN_nist_mod_224(BIGNUM *r, const BIGNUM *a, const BIGNUM *field,
 
 #if BN_BITS2==64
 	/* copy upper 256 bits of 448 bit number ... */
-	nist_cp_bn_0(t_d, a_d + (BN_NIST_224_TOP-1), top - (BN_NIST_224_TOP-1), BN_NIST_224_TOP);
+	nist_cp_bn_0(c_d, a_d + (BN_NIST_224_TOP-1), top - (BN_NIST_224_TOP-1), BN_NIST_224_TOP);
 	/* ... and right shift by 32 to obtain upper 224 bits */
-	nist_set_224(buf, t_d, 14, 13, 12, 11, 10, 9, 8);
+	nist_set_224(buf.bn, c_d, 14, 13, 12, 11, 10, 9, 8);
 	/* truncate lower part to 224 bits too */
 	r_d[BN_NIST_224_TOP-1] &= BN_MASK2l;
 #else
-	nist_cp_bn_0(buf, a_d + BN_NIST_224_TOP, top - BN_NIST_224_TOP, BN_NIST_224_TOP);
+	nist_cp_bn_0(buf.bn, a_d + BN_NIST_224_TOP, top - BN_NIST_224_TOP, BN_NIST_224_TOP);
 #endif
-	nist_set_224(t_d, buf, 10, 9, 8, 7, 0, 0, 0);
+
+#if defined(NIST_INT64) && BN_BITS2!=64
+	{
+	NIST_INT64		acc;	/* accumulator */
+	unsigned int		*rp=(unsigned int *)r_d;
+	const unsigned int	*bp=(const unsigned int *)buf.ui;
+
+	acc  = rp[0];	acc -= bp[7-7];
+			acc -= bp[11-7]; rp[0] = (unsigned int)acc; acc >>= 32;
+
+	acc += rp[1];	acc -= bp[8-7];
+			acc -= bp[12-7]; rp[1] = (unsigned int)acc; acc >>= 32;
+
+	acc += rp[2];	acc -= bp[9-7];
+			acc -= bp[13-7]; rp[2] = (unsigned int)acc; acc >>= 32;
+
+	acc += rp[3];	acc += bp[7-7];
+			acc += bp[11-7];
+			acc -= bp[10-7]; rp[3] = (unsigned int)acc; acc>>= 32;
+
+	acc += rp[4];	acc += bp[8-7];
+			acc += bp[12-7];
+			acc -= bp[11-7]; rp[4] = (unsigned int)acc; acc >>= 32;
+
+	acc += rp[5];	acc += bp[9-7];
+			acc += bp[13-7];
+			acc -= bp[12-7]; rp[5] = (unsigned int)acc; acc >>= 32;
+
+	acc += rp[6];	acc += bp[10-7];
+			acc -= bp[13-7]; rp[6] = (unsigned int)acc;
+
+	carry = (int)(acc>>32);
+# if BN_BITS2==64
+	rp[7] = carry;
+# endif
+	}	
+#else
+	{
+	BN_ULONG t_d[BN_NIST_224_TOP];
+
+	nist_set_224(t_d, buf.bn, 10, 9, 8, 7, 0, 0, 0);
 	carry = (int)bn_add_words(r_d, r_d, t_d, BN_NIST_224_TOP);
-	nist_set_224(t_d, buf, 0, 13, 12, 11, 0, 0, 0);
+	nist_set_224(t_d, buf.bn, 0, 13, 12, 11, 0, 0, 0);
 	carry += (int)bn_add_words(r_d, r_d, t_d, BN_NIST_224_TOP);
-	nist_set_224(t_d, buf, 13, 12, 11, 10, 9, 8, 7);
+	nist_set_224(t_d, buf.bn, 13, 12, 11, 10, 9, 8, 7);
 	carry -= (int)bn_sub_words(r_d, r_d, t_d, BN_NIST_224_TOP);
-	nist_set_224(t_d, buf, 0, 0, 0, 0, 13, 12, 11);
+	nist_set_224(t_d, buf.bn, 0, 0, 0, 0, 13, 12, 11);
 	carry -= (int)bn_sub_words(r_d, r_d, t_d, BN_NIST_224_TOP);
 
 #if BN_BITS2==64
 	carry = (int)(r_d[BN_NIST_224_TOP-1]>>32);
 #endif
+	}
+#endif
 	u.f = bn_sub_words;
 	if (carry > 0)
 		{
@@ -521,7 +609,8 @@ int BN_nist_mod_224(BIGNUM *r, const BIGNUM *a, const BIGNUM *field,
 	/* otherwise it's effectively same as in BN_nist_mod_192... */
 	mask  = 0-(PTR_SIZE_INT)(*u.f)(c_d,r_d,_nist_p_224[0],BN_NIST_224_TOP);
 	mask &= 0-(PTR_SIZE_INT)carry;
-	res   = (BN_ULONG *)(((PTR_SIZE_INT)c_d&~mask) |
+	res   = c_d;
+	res   = (BN_ULONG *)(((PTR_SIZE_INT)res&~mask) |
 	 ((PTR_SIZE_INT)r_d&mask));
 	nist_cp_bn(r_d, res, BN_NIST_224_TOP);
 	r->top = BN_NIST_224_TOP;
@@ -548,9 +637,11 @@ int BN_nist_mod_256(BIGNUM *r, const BIGNUM *a, const BIGNUM *field,
 	int	i, top = a->top;
 	int	carry = 0;
 	register BN_ULONG *a_d = a->d, *r_d;
-	BN_ULONG t_d[BN_NIST_256_TOP],
-	         buf[BN_NIST_256_TOP],
-		 c_d[BN_NIST_256_TOP],
+	union	{
+		BN_ULONG bn[BN_NIST_256_TOP];
+		unsigned int ui[BN_NIST_256_TOP*sizeof(BN_ULONG)/sizeof(unsigned int)];
+		} buf;
+	BN_ULONG c_d[BN_NIST_256_TOP],
 		*res;
 	PTR_SIZE_INT mask;
 	union { bn_addsub_f f; PTR_SIZE_INT p; } u;
@@ -584,12 +675,87 @@ int BN_nist_mod_256(BIGNUM *r, const BIGNUM *a, const BIGNUM *field,
 	else
 		r_d = a_d;
 
-	nist_cp_bn_0(buf, a_d + BN_NIST_256_TOP, top - BN_NIST_256_TOP, BN_NIST_256_TOP);
+	nist_cp_bn_0(buf.bn, a_d + BN_NIST_256_TOP, top - BN_NIST_256_TOP, BN_NIST_256_TOP);
+
+#if defined(NIST_INT64)
+	{
+	NIST_INT64		acc;	/* accumulator */
+	unsigned int		*rp=(unsigned int *)r_d;
+	const unsigned int	*bp=(const unsigned int *)buf.ui;
+
+	acc = rp[0];	acc += bp[8-8];
+			acc += bp[9-8];
+			acc -= bp[11-8];
+			acc -= bp[12-8];
+			acc -= bp[13-8];
+			acc -= bp[14-8]; rp[0] = (unsigned int)acc; acc >>= 32;
+
+	acc += rp[1];	acc += bp[9-8];
+			acc += bp[10-8];
+			acc -= bp[12-8];
+			acc -= bp[13-8];
+			acc -= bp[14-8];
+			acc -= bp[15-8]; rp[1] = (unsigned int)acc; acc >>= 32;
+
+	acc += rp[2];	acc += bp[10-8];
+			acc += bp[11-8];
+			acc -= bp[13-8];
+			acc -= bp[14-8];
+			acc -= bp[15-8]; rp[2] = (unsigned int)acc; acc >>= 32;
+
+	acc += rp[3];	acc += bp[11-8];
+			acc += bp[11-8];
+			acc += bp[12-8];
+			acc += bp[12-8];
+			acc += bp[13-8];
+			acc -= bp[15-8];
+			acc -= bp[8-8];
+			acc -= bp[9-8];  rp[3] = (unsigned int)acc; acc >>= 32;
+
+	acc += rp[4];	acc += bp[12-8];
+			acc += bp[12-8];
+			acc += bp[13-8];
+			acc += bp[13-8];
+			acc += bp[14-8];
+			acc -= bp[9-8];
+			acc -= bp[10-8]; rp[4] = (unsigned int)acc; acc >>= 32;
+
+	acc += rp[5];	acc += bp[13-8];
+			acc += bp[13-8];
+			acc += bp[14-8];
+			acc += bp[14-8];
+			acc += bp[15-8];
+			acc -= bp[10-8];
+			acc -= bp[11-8]; rp[5] = (unsigned int)acc; acc >>= 32;
+
+	acc += rp[6];	acc += bp[14-8];
+			acc += bp[14-8];
+			acc += bp[15-8];
+			acc += bp[15-8];
+			acc += bp[14-8];
+			acc += bp[13-8];
+			acc -= bp[8-8];
+			acc -= bp[9-8];  rp[6] = (unsigned int)acc; acc >>= 32;
+
+	acc += rp[7];	acc += bp[15-8];
+			acc += bp[15-8];
+			acc += bp[15-8];
+			acc += bp[8 -8];
+			acc -= bp[10-8];
+			acc -= bp[11-8];
+			acc -= bp[12-8];
+			acc -= bp[13-8]; rp[7] = (unsigned int)acc;
+
+	carry = (int)(acc>>32);
+	}
+#else
+	{
+	BN_ULONG t_d[BN_NIST_256_TOP];
 
 	/*S1*/
-	nist_set_256(t_d, buf, 15, 14, 13, 12, 11, 0, 0, 0);
+	nist_set_256(t_d, buf.bn, 15, 14, 13, 12, 11, 0, 0, 0);
 	/*S2*/
-	nist_set_256(c_d, buf, 0, 15, 14, 13, 12, 0, 0, 0);
+	nist_set_256(c_d, buf.bn, 0, 15, 14, 13, 12, 0, 0, 0);
 	carry = (int)bn_add_words(t_d, t_d, c_d, BN_NIST_256_TOP);
 	/* left shift */
 		{
@@ -607,24 +773,26 @@ int BN_nist_mod_256(BIGNUM *r, const BIGNUM *a, const BIGNUM *field,
 		}
 	carry += (int)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);
+	nist_set_256(t_d, buf.bn, 15, 14, 0, 0, 0, 10, 9, 8);
 	carry += (int)bn_add_words(r_d, r_d, t_d, BN_NIST_256_TOP);
 	/*S4*/
-	nist_set_256(t_d, buf, 8, 13, 15, 14, 13, 11, 10, 9);
+	nist_set_256(t_d, buf.bn, 8, 13, 15, 14, 13, 11, 10, 9);
 	carry += (int)bn_add_words(r_d, r_d, t_d, BN_NIST_256_TOP);
 	/*D1*/
-	nist_set_256(t_d, buf, 10, 8, 0, 0, 0, 13, 12, 11);
+	nist_set_256(t_d, buf.bn, 10, 8, 0, 0, 0, 13, 12, 11);
 	carry -= (int)bn_sub_words(r_d, r_d, t_d, BN_NIST_256_TOP);
 	/*D2*/
-	nist_set_256(t_d, buf, 11, 9, 0, 0, 15, 14, 13, 12);
+	nist_set_256(t_d, buf.bn, 11, 9, 0, 0, 15, 14, 13, 12);
 	carry -= (int)bn_sub_words(r_d, r_d, t_d, BN_NIST_256_TOP);
 	/*D3*/
-	nist_set_256(t_d, buf, 12, 0, 10, 9, 8, 15, 14, 13);
+	nist_set_256(t_d, buf.bn, 12, 0, 10, 9, 8, 15, 14, 13);
 	carry -= (int)bn_sub_words(r_d, r_d, t_d, BN_NIST_256_TOP);
 	/*D4*/
-	nist_set_256(t_d, buf, 13, 0, 11, 10, 9, 0, 15, 14);
+	nist_set_256(t_d, buf.bn, 13, 0, 11, 10, 9, 0, 15, 14);
 	carry -= (int)bn_sub_words(r_d, r_d, t_d, BN_NIST_256_TOP);
 
+	}
+#endif
 	/* see BN_nist_mod_224 for explanation */
 	u.f = bn_sub_words;
 	if (carry > 0)
@@ -641,7 +809,8 @@ int BN_nist_mod_256(BIGNUM *r, const BIGNUM *a, const BIGNUM *field,
 
 	mask  = 0-(PTR_SIZE_INT)(*u.f)(c_d,r_d,_nist_p_256[0],BN_NIST_256_TOP);
 	mask &= 0-(PTR_SIZE_INT)carry;
-	res   = (BN_ULONG *)(((PTR_SIZE_INT)c_d&~mask) |
+	res   = c_d;
+	res   = (BN_ULONG *)(((PTR_SIZE_INT)res&~mask) |
 	 ((PTR_SIZE_INT)r_d&mask));
 	nist_cp_bn(r_d, res, BN_NIST_256_TOP);
 	r->top = BN_NIST_256_TOP;
@@ -672,9 +841,11 @@ int BN_nist_mod_384(BIGNUM *r, const BIGNUM *a, const BIGNUM *field,
 	int	i, top = a->top;
 	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],
-		 c_d[BN_NIST_384_TOP],
+	union	{
+		BN_ULONG bn[BN_NIST_384_TOP];
+		unsigned int ui[BN_NIST_384_TOP*sizeof(BN_ULONG)/sizeof(unsigned int)];
+		} buf;
+	BN_ULONG c_d[BN_NIST_384_TOP],
 		*res;
 	PTR_SIZE_INT mask;
 	union { bn_addsub_f f; PTR_SIZE_INT p; } u;
@@ -709,10 +880,100 @@ int BN_nist_mod_384(BIGNUM *r, const BIGNUM *a, const BIGNUM *field,
 	else
 		r_d = a_d;
 
-	nist_cp_bn_0(buf, a_d + BN_NIST_384_TOP, top - BN_NIST_384_TOP, BN_NIST_384_TOP);
+	nist_cp_bn_0(buf.bn, a_d + BN_NIST_384_TOP, top - BN_NIST_384_TOP, BN_NIST_384_TOP);
+
+#if defined(NIST_INT64)
+	{
+	NIST_INT64		acc;	/* accumulator */
+	unsigned int		*rp=(unsigned int *)r_d;
+	const unsigned int	*bp=(const unsigned int *)buf.ui;
+
+	acc = rp[0];	acc += bp[12-12];
+			acc += bp[21-12];
+			acc += bp[20-12];
+			acc -= bp[23-12]; rp[0] = (unsigned int)acc; acc >>= 32;
+
+	acc += rp[1];	acc += bp[13-12];
+			acc += bp[22-12];
+			acc += bp[23-12];
+			acc -= bp[12-12];
+			acc -= bp[20-12]; rp[1] = (unsigned int)acc; acc >>= 32;
+
+	acc += rp[2];	acc += bp[14-12];
+			acc += bp[23-12];
+			acc -= bp[13-12];
+			acc -= bp[21-12]; rp[2] = (unsigned int)acc; acc >>= 32;
+
+	acc += rp[3];	acc += bp[15-12];
+			acc += bp[12-12];
+			acc += bp[20-12];
+			acc += bp[21-12];
+			acc -= bp[14-12];
+			acc -= bp[22-12];
+			acc -= bp[23-12]; rp[3] = (unsigned int)acc; acc >>= 32;
+
+	acc += rp[4];	acc += bp[21-12];
+			acc += bp[21-12];
+			acc += bp[16-12];
+			acc += bp[13-12];
+			acc += bp[12-12];
+			acc += bp[20-12];
+			acc += bp[22-12];
+			acc -= bp[15-12];
+			acc -= bp[23-12];
+			acc -= bp[23-12]; rp[4] = (unsigned int)acc; acc >>= 32;
+
+	acc += rp[5];	acc += bp[22-12];
+			acc += bp[22-12];
+			acc += bp[17-12];
+			acc += bp[14-12];
+			acc += bp[13-12];
+			acc += bp[21-12];
+			acc += bp[23-12];
+			acc -= bp[16-12]; rp[5] = (unsigned int)acc; acc >>= 32;
+			
+	acc += rp[6];	acc += bp[23-12];
+			acc += bp[23-12];
+			acc += bp[18-12];
+			acc += bp[15-12];
+			acc += bp[14-12];
+			acc += bp[22-12];
+			acc -= bp[17-12]; rp[6] = (unsigned int)acc; acc >>= 32;
+			
+	acc += rp[7];	acc += bp[19-12];
+			acc += bp[16-12];
+			acc += bp[15-12];
+			acc += bp[23-12];
+			acc -= bp[18-12]; rp[7] = (unsigned int)acc; acc >>= 32;
+			
+	acc += rp[8];	acc += bp[20-12];
+			acc += bp[17-12];
+			acc += bp[16-12];
+			acc -= bp[19-12]; rp[8] = (unsigned int)acc; acc >>= 32;
+			
+	acc += rp[9];	acc += bp[21-12];
+			acc += bp[18-12];
+			acc += bp[17-12];
+			acc -= bp[20-12]; rp[9] = (unsigned int)acc; acc >>= 32;
+			
+	acc += rp[10];	acc += bp[22-12];
+			acc += bp[19-12];
+			acc += bp[18-12];
+			acc -= bp[21-12]; rp[10] = (unsigned int)acc; acc >>= 32;
+			
+	acc += rp[11];	acc += bp[23-12];
+			acc += bp[20-12];
+			acc += bp[19-12];
+			acc -= bp[22-12]; rp[11] = (unsigned int)acc;
+
+	carry = (int)(acc>>32);
+	}
+#else
+	{
+	BN_ULONG t_d[BN_NIST_384_TOP];
 
 	/*S1*/
-	nist_set_256(t_d, buf, 0, 0, 0, 0, 0, 23-4, 22-4, 21-4);
+	nist_set_256(t_d, buf.bn, 0, 0, 0, 0, 0, 23-4, 22-4, 21-4);
 		/* left shift */
 		{
 		register BN_ULONG *ap,t,c;
@@ -729,29 +990,31 @@ int BN_nist_mod_384(BIGNUM *r, const BIGNUM *a, const BIGNUM *field,
 	carry = (int)bn_add_words(r_d+(128/BN_BITS2), r_d+(128/BN_BITS2), 
 		t_d, BN_NIST_256_TOP);
 	/*S2 */
-	carry += (int)bn_add_words(r_d, r_d, buf, BN_NIST_384_TOP);
+	carry += (int)bn_add_words(r_d, r_d, buf.bn, BN_NIST_384_TOP);
 	/*S3*/
-	nist_set_384(t_d,buf,20,19,18,17,16,15,14,13,12,23,22,21);
+	nist_set_384(t_d,buf.bn,20,19,18,17,16,15,14,13,12,23,22,21);
 	carry += (int)bn_add_words(r_d, r_d, t_d, BN_NIST_384_TOP);
 	/*S4*/
-	nist_set_384(t_d,buf,19,18,17,16,15,14,13,12,20,0,23,0);
+	nist_set_384(t_d,buf.bn,19,18,17,16,15,14,13,12,20,0,23,0);
 	carry += (int)bn_add_words(r_d, r_d, t_d, BN_NIST_384_TOP);
 	/*S5*/
-	nist_set_384(t_d, buf,0,0,0,0,23,22,21,20,0,0,0,0);
+	nist_set_384(t_d, buf.bn,0,0,0,0,23,22,21,20,0,0,0,0);
 	carry += (int)bn_add_words(r_d, r_d, t_d, BN_NIST_384_TOP);
 	/*S6*/
-	nist_set_384(t_d,buf,0,0,0,0,0,0,23,22,21,0,0,20);
+	nist_set_384(t_d,buf.bn,0,0,0,0,0,0,23,22,21,0,0,20);
 	carry += (int)bn_add_words(r_d, r_d, t_d, BN_NIST_384_TOP);
 	/*D1*/
-	nist_set_384(t_d,buf,22,21,20,19,18,17,16,15,14,13,12,23);
+	nist_set_384(t_d,buf.bn,22,21,20,19,18,17,16,15,14,13,12,23);
 	carry -= (int)bn_sub_words(r_d, r_d, t_d, BN_NIST_384_TOP);
 	/*D2*/
-	nist_set_384(t_d,buf,0,0,0,0,0,0,0,23,22,21,20,0);
+	nist_set_384(t_d,buf.bn,0,0,0,0,0,0,0,23,22,21,20,0);
 	carry -= (int)bn_sub_words(r_d, r_d, t_d, BN_NIST_384_TOP);
 	/*D3*/
-	nist_set_384(t_d,buf,0,0,0,0,0,0,0,23,23,0,0,0);
+	nist_set_384(t_d,buf.bn,0,0,0,0,0,0,0,23,23,0,0,0);
 	carry -= (int)bn_sub_words(r_d, r_d, t_d, BN_NIST_384_TOP);
 
+	}
+#endif
 	/* see BN_nist_mod_224 for explanation */
 	u.f = bn_sub_words;
 	if (carry > 0)
@@ -768,7 +1031,8 @@ int BN_nist_mod_384(BIGNUM *r, const BIGNUM *a, const BIGNUM *field,
 
 	mask  = 0-(PTR_SIZE_INT)(*u.f)(c_d,r_d,_nist_p_384[0],BN_NIST_384_TOP);
 	mask &= 0-(PTR_SIZE_INT)carry;
-	res   = (BN_ULONG *)(((PTR_SIZE_INT)c_d&~mask) |
+	res   = c_d;
+	res   = (BN_ULONG *)(((PTR_SIZE_INT)res&~mask) |
 	 ((PTR_SIZE_INT)r_d&mask));
 	nist_cp_bn(r_d, res, BN_NIST_384_TOP);
 	r->top = BN_NIST_384_TOP;
@@ -834,7 +1098,8 @@ int BN_nist_mod_521(BIGNUM *r, const BIGNUM *a, const BIGNUM *field,
 
 	bn_add_words(r_d,r_d,t_d,BN_NIST_521_TOP);
 	mask = 0-(PTR_SIZE_INT)bn_sub_words(t_d,r_d,_nist_p_521,BN_NIST_521_TOP);
-	res  = (BN_ULONG *)(((PTR_SIZE_INT)t_d&~mask) |
+	res  = t_d;
+	res  = (BN_ULONG *)(((PTR_SIZE_INT)res&~mask) |
 	 ((PTR_SIZE_INT)r_d&mask));
 	nist_cp_bn(r_d,res,BN_NIST_521_TOP);
 	r->top = BN_NIST_521_TOP;