diff --git a/ecdsa.c b/ecdsa.c
index 0b012ecb0e..85a3fab241 100644
--- a/ecdsa.c
+++ b/ecdsa.c
@@ -195,53 +195,101 @@ int point_is_negative_of(const curve_point *p, const curve_point *q)
 
 #if USE_PRECOMPUTED_CP
 
+// Negate a (modulo prime) if cond is 0xffffffff, keep it if cond is 0.
+// The timing of this function does not depend on cond.
+static void conditional_negate(uint32_t cond, bignum256 *a, const bignum256 *prime)
+{
+	int j;
+	uint32_t tmp = 1;
+	for (j = 0; j < 8; j++) {
+		tmp += 0x3fffffff + prime->val[j] - a->val[j];
+		a->val[j] = ((tmp & 0x3fffffff) & cond) | (a->val[j] & ~cond);
+		tmp >>= 30;
+	}
+	tmp += 0x3fffffff + prime->val[j] - a->val[j];
+	a->val[j] = ((tmp & 0x3fffffff) & cond) | (a->val[j] & ~cond);
+}
+
 // res = k * G
+// k must be a normalized number with 0 <= k < order256k1
 void scalar_multiply(const bignum256 *k, curve_point *res)
 {
-	int i, j;
-	// result is zero
 	assert (bn_is_less(k, &order256k1));
-	bignum256 a = *k;
-	int sign = (a.val[0] & 1) ^ 1;
-	uint32_t lowbits;
-	// make number odd
-	if (sign) {
-		bn_subtract(&order256k1, &a, &a);
-	}
-	a.val[8] |= 0x10000;
-	assert((a.val[0] & 1) != 0);
-	assert((a.val[8] & 0x10000) != 0);
 
-	// now compute  res = a *G step by step.
-	// initial res
+	int i, j;
+	bignum256 a;
+	uint32_t is_even = (k->val[0] & 1) - 1;
+	uint32_t lowbits;
+
+	// is_even = 0xffffffff if k is even, 0 otherwise.
+
+	// add 2^256.
+	// make number odd: subtract order256k1 if even
+	uint32_t tmp = 1;
+	uint32_t is_non_zero = 0;
+	for (j = 0; j < 8; j++) {
+		is_non_zero |= k->val[j];
+		tmp += 0x3fffffff + k->val[j] - (order256k1.val[j] & is_even);
+		a.val[j] = tmp & 0x3fffffff;
+		tmp >>= 30;
+	}
+	is_non_zero |= k->val[j];
+	a.val[j] = tmp + 0xffff + k->val[j] - (order256k1.val[j] & is_even);
+	assert((a.val[0] & 1) != 0);
+
+	// special case 0*G:  just return zero. We don't care about constant time.
+	if (!is_non_zero) {
+		point_set_infinity(res);
+		return;
+	}
+
+	// Now a = k + 2^256 (mod order256k1) and a is odd.
+	//
+	// The idea is to bring the new a into the form.
+	// sum_{i=0..64} a[i] 16^i,  where |a[i]| < 16 and a[i] is odd.
+	// a[0] is odd, since a is odd.  If a[i] would be even, we can
+	// add 1 to it and subtract 16 from a[i-1].  Afterwards,
+	// a[64] = 1, which is the 2^256 that we added before.
+	//
+	// Since k = a - 2^256 (mod order256k1), we can compute
+	//   k*G = sum_{i=0..63} a[i] 16^i * G
+	//
+	// We have a big table secp256k1_cp that stores all possible
+	// values of |a[i]| 16^i * G.
+	// secp256k1_cp[i][j] = (2*j+1) * 16^i * G
+
+	// now compute  res = sum_{i=0..63} a[i] * 16^i * G step by step.
+	// initial res = |a[0]| * G.  Note that a[0] = a & 0xf if (a&0x10) != 0
+	// and - (16 - (a & 0xf)) otherwise.   We can compute this as
+	//   ((a ^ (((a >> 4) & 1) - 1)) & 0xf) >> 1
+	// since a is odd.
 	lowbits = a.val[0] & ((1 << 5) - 1);
 	lowbits ^= (lowbits >> 4) - 1;
 	lowbits &= 15;
 	*res = secp256k1_cp[0][lowbits >> 1];
 	for (i = 1; i < 64; i ++) {
-		// invariant res = abs((a % 2*16^i) - 16^i) * G
+		// invariant res = sign(a[i-1]) sum_{j=0..i-1} (a[j] * 16^j * G)
+		// Note that sign(a[i-1]
 
+		// shift a by 4 places.
 		for (j = 0; j < 8; j++) {
 			a.val[j] = (a.val[j] >> 4) | ((a.val[j + 1] & 0xf) << 26);
 		}
 		a.val[j] >>= 4;
+		// a = old(a)>>(4*i)
+		// a is even iff sign(a[i-1]) = -1
 
 		lowbits = a.val[0] & ((1 << 5) - 1);
 		lowbits ^= (lowbits >> 4) - 1;
 		lowbits &= 15;
-		if ((lowbits & 1) == 0) {
-			// negate last result to make signs of this round and the
-			// last round equal.
-			bn_subtract(&prime256k1, &res->y, &res->y);
-		}
+		// negate last result to make signs of this round and the
+		// last round equal.
+		conditional_negate((lowbits & 1) - 1, &res->y, &prime256k1);
 		
 		// add odd factor
 		point_add(&secp256k1_cp[i][lowbits >> 1], res);
 	}
-	if (sign) {
-		// negate 
-		bn_subtract(&prime256k1, &res->y, &res->y);
-	}
+	conditional_negate(((a.val[0] >> 4) & 1) - 1, &res->y, &prime256k1);
 }
 
 #else