bn_gf2m.c: optimized BN_GF2m_mod_inv delivers sometimes 2x of ECDSA sign.

Exact improvement coefficients vary from one benchmark and platform to
another, e.g. it performs 70%-33% better on ARM, hereafter less for
longer keys, and 100%-90% better on x86_64.

1 files changed, 80 insertions, 16 deletions

diff --git a/crypto/bn/bn_gf2m.c b/crypto/bn/bn_gf2m.c
index e170fffc6c..6caf2888a6 100644
--- a/crypto/bn/bn_gf2m.c
+++ b/crypto/bn/bn_gf2m.c
@@ -364,21 +364,17 @@ int BN_GF2m_mod_arr(BIGNUM *r, const BIGNUM *a, const int p[])
 int	BN_GF2m_mod(BIGNUM *r, const BIGNUM *a, const BIGNUM *p)
 	{
 	int ret = 0;
-	const int max = BN_num_bits(p) + 1;
-	int *arr=NULL;
+	int arr[6];
 	bn_check_top(a);
 	bn_check_top(p);
-	if ((arr = (int *)OPENSSL_malloc(sizeof(int) * max)) == NULL) goto err;
-	ret = BN_GF2m_poly2arr(p, arr, max);
-	if (!ret || ret > max)
+	ret = BN_GF2m_poly2arr(p, arr, sizeof(arr)/sizeof(arr[0]));
+	if (!ret || ret > sizeof(arr)/sizeof(arr[0]))
 		{
 		BNerr(BN_F_BN_GF2M_MOD,BN_R_INVALID_LENGTH);
-		goto err;
+		return 0;
 		}
 	ret = BN_GF2m_mod_arr(r, a, arr);
 	bn_check_top(r);
-err:
-	if (arr) OPENSSL_free(arr);
 	return ret;
 	}
 
@@ -533,18 +529,18 @@ int BN_GF2m_mod_inv(BIGNUM *r, const BIGNUM *a, const BIGNUM *p, BN_CTX *ctx)
 
 	BN_CTX_start(ctx);
 	
-	b = BN_CTX_get(ctx);
-	c = BN_CTX_get(ctx);
-	u = BN_CTX_get(ctx);
-	v = BN_CTX_get(ctx);
-	if (v == NULL) goto err;
+	if ((b = BN_CTX_get(ctx))==NULL) goto err;
+	if ((c = BN_CTX_get(ctx))==NULL) goto err;
+	if ((u = BN_CTX_get(ctx))==NULL) goto err;
+	if ((v = BN_CTX_get(ctx))==NULL) goto err;
 
-	if (!BN_one(b)) goto err;
 	if (!BN_GF2m_mod(u, a, p)) goto err;
-	if (!BN_copy(v, p)) goto err;
-
 	if (BN_is_zero(u)) goto err;
 
+	if (!BN_copy(v, p)) goto err;
+#if 0
+	if (!BN_one(b)) goto err;
+
 	while (1)
 		{
 		while (!BN_is_odd(u))
@@ -568,7 +564,75 @@ int BN_GF2m_mod_inv(BIGNUM *r, const BIGNUM *a, const BIGNUM *p, BN_CTX *ctx)
 		if (!BN_GF2m_add(u, u, v)) goto err;
 		if (!BN_GF2m_add(b, b, c)) goto err;
 		}
+#else
+	{
+	int i,	ubits = BN_num_bits(u),
+		vbits = BN_num_bits(v),	/* v is copy of p */
+		top = p->top;
+	BN_ULONG *udp,*bdp,*vdp,*cdp;
+
+	bn_wexpand(u,top);	udp = u->d;
+				for (i=u->top;i<top;i++) udp[i] = 0;
+				u->top = top;
+	bn_wexpand(b,top);	bdp = b->d;
+				bdp[0] = 1;
+				for (i=1;i<top;i++) bdp[i] = 0;
+				b->top = top;
+	bn_wexpand(c,top);	cdp = c->d;
+				for (i=0;i<top;i++) cdp[i] = 0;
+				c->top = top;
+	vdp = v->d;	/* It pays off to "cache" *->d pointers, because
+			 * it allows optimizer to be more aggressive.
+			 * But we don't have to "cache" p->d, because *p
+			 * is declared 'const'... */
+	while (1)
+		{
+		while (ubits && !(udp[0]&1))
+			{
+			BN_ULONG u0,u1,b0,b1,mask;
+
+			u0   = udp[0];
+			b0   = bdp[0];
+			mask = (BN_ULONG)0-(b0&1);
+			b0  ^= p->d[0]&mask;
+			for (i=0;i<top-1;i++)
+				{
+				u1 = udp[i+1];
+				udp[i] = ((u0>>1)|(u1<<(BN_BITS2-1)))&BN_MASK2;
+				u0 = u1;
+				b1 = bdp[i+1]^(p->d[i+1]&mask);
+				bdp[i] = ((b0>>1)|(b1<<(BN_BITS2-1)))&BN_MASK2;
+				b0 = b1;
+				}
+			udp[i] = u0>>1;
+			bdp[i] = b0>>1;
+			ubits--;
+			}
 
+		if (ubits<=BN_BITS2 && udp[0]==1) break;
+
+		if (ubits<vbits)
+			{
+			i = ubits; ubits = vbits; vbits = i;
+			tmp = u; u = v; v = tmp;
+			tmp = b; b = c; c = tmp;
+			udp = vdp; vdp = v->d;
+			bdp = cdp; cdp = c->d;
+			}
+		for(i=0;i<top;i++)
+			{
+			udp[i] ^= vdp[i];
+			bdp[i] ^= cdp[i];
+			}
+		if (ubits==vbits)
+			{
+			bn_fix_top(u);
+			ubits = BN_num_bits(u);
+			}
+		}
+	bn_fix_top(b);
+	}
+#endif
 
 	if (!BN_copy(r, b)) goto err;
 	bn_check_top(r);