X-Git-Url: https://www.bearssl.org/gitweb//home/git/?p=BearSSL;a=blobdiff_plain;f=src%2Fec%2Fec_p256_m15.c;h=6ce57e00bce378b40fff91a9fbd2e7de1c3efeda;hp=0d40aef5ba8b416c8f643fb1b8e4dce59e277ce2;hb=d592e999329b4a661e5660d56e3226b73a2ad364;hpb=3655193439d4e093bb80a2f0d9e02179d424de08

diff --git a/src/ec/ec_p256_m15.c b/src/ec/ec_p256_m15.c
index 0d40aef..6ce57e0 100644
--- a/src/ec/ec_p256_m15.c
+++ b/src/ec/ec_p256_m15.c
@@ -1101,18 +1101,20 @@ mul_f256(uint32_t *d, const uint32_t *a, const uint32_t *b)
 	}
 
 	/*
-	 * Propagate carries. Since the operation above really is a
-	 * truncature, followed by the addition of nonnegative values,
-	 * the result will be positive. Moreover, the carry cannot
-	 * exceed 5 bits (we performed 20 additions with values smaller
-	 * than 256 bits).
+	 * Propagate carries. This is a signed propagation, and the
+	 * result may be negative. The loop above may enlarge values,
+	 * but not two much: worst case is the chain involving t[i - 3],
+	 * in which a value may be added to itself up to 7 times. Since
+	 * starting values are 13-bit each, all words fit on 20 bits
+	 * (21 to account for the sign bit).
 	 */
 	cc = norm13(t, t, 20);
 
 	/*
 	 * Perform modular reduction again for the bits beyond 256 (the carry
-	 * and the bits 256..259). This time, we can simply inject full
-	 * word values.
+	 * and the bits 256..259). Since the largest shift below is by 10
+	 * bits, and the values fit on 21 bits, values fit in 32-bit words,
+	 * thereby allowing injecting full word values.
 	 */
 	cc = (cc << 4) | (t[19] >> 9);
 	t[19] &= 0x01FF;
@@ -1120,6 +1122,22 @@ mul_f256(uint32_t *d, const uint32_t *a, const uint32_t *b)
 	t[14] -= cc << 10;
 	t[7] -= cc << 5;
 	t[0] += cc;
+
+	/*
+	 * If the carry is negative, then after carry propagation, we may
+	 * end up with a value which is negative, and we don't want that.
+	 * Thus, in that case, we add the modulus. Note that the subtraction
+	 * result, when the carry is negative, is always smaller than the
+	 * modulus, so the extra addition will not make the value exceed
+	 * twice the modulus.
+	 */
+	cc >>= 31;
+	t[0] -= cc;
+	t[7] += cc << 5;
+	t[14] += cc << 10;
+	t[17] -= cc << 3;
+	t[19] += cc << 9;
+
 	norm13(d, t, 20);
 }
 
@@ -1172,18 +1190,20 @@ square_f256(uint32_t *d, const uint32_t *a)
 	}
 
 	/*
-	 * Propagate carries. Since the operation above really is a
-	 * truncature, followed by the addition of nonnegative values,
-	 * the result will be positive. Moreover, the carry cannot
-	 * exceed 5 bits (we performed 20 additions with values smaller
-	 * than 256 bits).
+	 * Propagate carries. This is a signed propagation, and the
+	 * result may be negative. The loop above may enlarge values,
+	 * but not two much: worst case is the chain involving t[i - 3],
+	 * in which a value may be added to itself up to 7 times. Since
+	 * starting values are 13-bit each, all words fit on 20 bits
+	 * (21 to account for the sign bit).
 	 */
 	cc = norm13(t, t, 20);
 
 	/*
 	 * Perform modular reduction again for the bits beyond 256 (the carry
-	 * and the bits 256..259). This time, we can simply inject full
-	 * word values.
+	 * and the bits 256..259). Since the largest shift below is by 10
+	 * bits, and the values fit on 21 bits, values fit in 32-bit words,
+	 * thereby allowing injecting full word values.
 	 */
 	cc = (cc << 4) | (t[19] >> 9);
 	t[19] &= 0x01FF;
@@ -1191,6 +1211,22 @@ square_f256(uint32_t *d, const uint32_t *a)
 	t[14] -= cc << 10;
 	t[7] -= cc << 5;
 	t[0] += cc;
+
+	/*
+	 * If the carry is negative, then after carry propagation, we may
+	 * end up with a value which is negative, and we don't want that.
+	 * Thus, in that case, we add the modulus. Note that the subtraction
+	 * result, when the carry is negative, is always smaller than the
+	 * modulus, so the extra addition will not make the value exceed
+	 * twice the modulus.
+	 */
+	cc >>= 31;
+	t[0] -= cc;
+	t[7] += cc << 5;
+	t[14] += cc << 10;
+	t[17] -= cc << 3;
+	t[19] += cc << 9;
+
 	norm13(d, t, 20);
 }