Merge #1033: Add _fe_half and use in _gej_add_ge and _gej_double

e848c37 Update sage files for new formulae (Peter Dettman)
d64bb5d Add fe_half tests for worst-case inputs (Peter Dettman)
4eb8b93 Further improve doubling formula using fe_half (Peter Dettman)
557b31f Doubling formula using fe_half (Pieter Wuille)
2cbb4b1 Run more iterations of run_field_misc (Pieter Wuille)
9cc5c25 Add test for secp256k1_fe_half (Pieter Wuille)
925f78d Add _fe_half and use in _gej_add_ge (Peter Dettman)

Pull request description:

  - Trades 1 _half for 3 _mul_int and 2 _normalize_weak

  Gives around 2-3% faster signing and ECDH, depending on compiler/platform.

ACKs for top commit:
  sipa:
    utACK e848c37
  jonasnick:
    ACK e848c37
  real-or-random:
    ACK e848c37

Tree-SHA512: 81a6c93b3d983f1b48ec8e8b6f262ba914215045a95415147f41ee6e85296aa4d0cbbad9f370cdf475571447baad861d2cc8e0b04a71202d48959cb8a098f584

Author: real-or-random

sage/prove_group_implementations.sage

@@ -8,25 +8,20 @@ load("weierstrass_prover.sage")
 def formula_secp256k1_gej_double_var(a):
   """libsecp256k1's secp256k1_gej_double_var, used by various addition functions"""
   rz = a.Z * a.Y
-  rz = rz * 2
-  t1 = a.X^2
-  t1 = t1 * 3
-  t2 = t1^2
-  t3 = a.Y^2
-  t3 = t3 * 2
-  t4 = t3^2
-  t4 = t4 * 2
-  t3 = t3 * a.X
-  rx = t3
-  rx = rx * 4
-  rx = -rx
-  rx = rx + t2
-  t2 = -t2
-  t3 = t3 * 6
-  t3 = t3 + t2
-  ry = t1 * t3
-  t2 = -t4
-  ry = ry + t2
+  s = a.Y^2
+  l = a.X^2
+  l = l * 3
+  l = l / 2
+  t = -s
+  t = t * a.X
+  rx = l^2
+  rx = rx + t
+  rx = rx + t
+  s = s^2
+  t = t + rx
+  ry = t * l
+  ry = ry + s
+  ry = -ry
   return jacobianpoint(rx, ry, rz)
 
 def formula_secp256k1_gej_add_var(branch, a, b):
@@ -197,7 +192,8 @@ def formula_secp256k1_gej_add_ge(branch, a, b):
     rr_alt = rr
     m_alt = m
   n = m_alt^2
-  q = n * t
+  q = -t
+  q = q * n
   n = n^2
   if degenerate:
     n = m
@@ -210,17 +206,14 @@ def formula_secp256k1_gej_add_ge(branch, a, b):
     zeroes.update({rz : 'r.z=0'})
   else:
     nonzeroes.update({rz : 'r.z!=0'})
-  rz = rz * 2
-  q = -q
   t = t + q
   rx = t
   t = t * 2
   t = t + q
   t = t * rr_alt
   t = t + n
   ry = -t
-  rx = rx * 4
-  ry = ry * 4
+  ry = ry / 2
   if a_infinity:
     rx = b.X
     ry = b.Y

src/bench_internal.c

@@ -140,6 +140,15 @@ void bench_scalar_inverse_var(void* arg, int iters) {
     CHECK(j <= iters);
 }
 
+void bench_field_half(void* arg, int iters) {
+    int i;
+    bench_inv *data = (bench_inv*)arg;
+
+    for (i = 0; i < iters; i++) {
+        secp256k1_fe_half(&data->fe[0]);
+    }
+}
+
 void bench_field_normalize(void* arg, int iters) {
     int i;
     bench_inv *data = (bench_inv*)arg;
@@ -354,6 +363,7 @@ int main(int argc, char **argv) {
     if (d || have_flag(argc, argv, "scalar") || have_flag(argc, argv, "inverse")) run_benchmark("scalar_inverse", bench_scalar_inverse, bench_setup, NULL, &data, 10, iters);
     if (d || have_flag(argc, argv, "scalar") || have_flag(argc, argv, "inverse")) run_benchmark("scalar_inverse_var", bench_scalar_inverse_var, bench_setup, NULL, &data, 10, iters);
 
+    if (d || have_flag(argc, argv, "field") || have_flag(argc, argv, "half")) run_benchmark("field_half", bench_field_half, bench_setup, NULL, &data, 10, iters*100);
     if (d || have_flag(argc, argv, "field") || have_flag(argc, argv, "normalize")) run_benchmark("field_normalize", bench_field_normalize, bench_setup, NULL, &data, 10, iters*100);
     if (d || have_flag(argc, argv, "field") || have_flag(argc, argv, "normalize")) run_benchmark("field_normalize_weak", bench_field_normalize_weak, bench_setup, NULL, &data, 10, iters*100);
     if (d || have_flag(argc, argv, "field") || have_flag(argc, argv, "sqr")) run_benchmark("field_sqr", bench_field_sqr, bench_setup, NULL, &data, 10, iters*10);

src/field.h

@@ -130,4 +130,13 @@ static void secp256k1_fe_storage_cmov(secp256k1_fe_storage *r, const secp256k1_f
 /** If flag is true, set *r equal to *a; otherwise leave it. Constant-time.  Both *r and *a must be initialized.*/
 static void secp256k1_fe_cmov(secp256k1_fe *r, const secp256k1_fe *a, int flag);
 
+/** Halves the value of a field element modulo the field prime. Constant-time.
+ *  For an input magnitude 'm', the output magnitude is set to 'floor(m/2) + 1'.
+ *  The output is not guaranteed to be normalized, regardless of the input. */
+static void secp256k1_fe_half(secp256k1_fe *r);
+
+/** Sets each limb of 'r' to its upper bound at magnitude 'm'. The output will also have its
+ *  magnitude set to 'm' and is normalized if (and only if) 'm' is zero. */
+static void secp256k1_fe_get_bounds(secp256k1_fe *r, int m);
+
 #endif /* SECP256K1_FIELD_H */

src/field_10x26_impl.h

@@ -49,6 +49,26 @@ static void secp256k1_fe_verify(const secp256k1_fe *a) {
 }
 #endif
 
+static void secp256k1_fe_get_bounds(secp256k1_fe *r, int m) {
+    VERIFY_CHECK(m >= 0);
+    VERIFY_CHECK(m <= 2048);
+    r->n[0] = 0x3FFFFFFUL * 2 * m;
+    r->n[1] = 0x3FFFFFFUL * 2 * m;
+    r->n[2] = 0x3FFFFFFUL * 2 * m;
+    r->n[3] = 0x3FFFFFFUL * 2 * m;
+    r->n[4] = 0x3FFFFFFUL * 2 * m;
+    r->n[5] = 0x3FFFFFFUL * 2 * m;
+    r->n[6] = 0x3FFFFFFUL * 2 * m;
+    r->n[7] = 0x3FFFFFFUL * 2 * m;
+    r->n[8] = 0x3FFFFFFUL * 2 * m;
+    r->n[9] = 0x03FFFFFUL * 2 * m;
+#ifdef VERIFY
+    r->magnitude = m;
+    r->normalized = (m == 0);
+    secp256k1_fe_verify(r);
+#endif
+}
+
 static void secp256k1_fe_normalize(secp256k1_fe *r) {
     uint32_t t0 = r->n[0], t1 = r->n[1], t2 = r->n[2], t3 = r->n[3], t4 = r->n[4],
              t5 = r->n[5], t6 = r->n[6], t7 = r->n[7], t8 = r->n[8], t9 = r->n[9];
@@ -1133,6 +1153,82 @@ static SECP256K1_INLINE void secp256k1_fe_cmov(secp256k1_fe *r, const secp256k1_
 #endif
 }
 
+static SECP256K1_INLINE void secp256k1_fe_half(secp256k1_fe *r) {
+    uint32_t t0 = r->n[0], t1 = r->n[1], t2 = r->n[2], t3 = r->n[3], t4 = r->n[4],
+             t5 = r->n[5], t6 = r->n[6], t7 = r->n[7], t8 = r->n[8], t9 = r->n[9];
+    uint32_t one = (uint32_t)1;
+    uint32_t mask = -(t0 & one) >> 6;
+
+#ifdef VERIFY
+    secp256k1_fe_verify(r);
+    VERIFY_CHECK(r->magnitude < 32);
+#endif
+
+    /* Bounds analysis (over the rationals).
+     *
+     * Let m = r->magnitude
+     *     C = 0x3FFFFFFUL * 2
+     *     D = 0x03FFFFFUL * 2
+     *
+     * Initial bounds: t0..t8 <= C * m
+     *                     t9 <= D * m
+     */
+
+    t0 += 0x3FFFC2FUL & mask;
+    t1 += 0x3FFFFBFUL & mask;
+    t2 += mask;
+    t3 += mask;
+    t4 += mask;
+    t5 += mask;
+    t6 += mask;
+    t7 += mask;
+    t8 += mask;
+    t9 += mask >> 4;
+
+    VERIFY_CHECK((t0 & one) == 0);
+
+    /* t0..t8: added <= C/2
+     *     t9: added <= D/2
+     *
+     * Current bounds: t0..t8 <= C * (m + 1/2)
+     *                     t9 <= D * (m + 1/2)
+     */
+
+    r->n[0] = (t0 >> 1) + ((t1 & one) << 25);
+    r->n[1] = (t1 >> 1) + ((t2 & one) << 25);
+    r->n[2] = (t2 >> 1) + ((t3 & one) << 25);
+    r->n[3] = (t3 >> 1) + ((t4 & one) << 25);
+    r->n[4] = (t4 >> 1) + ((t5 & one) << 25);
+    r->n[5] = (t5 >> 1) + ((t6 & one) << 25);
+    r->n[6] = (t6 >> 1) + ((t7 & one) << 25);
+    r->n[7] = (t7 >> 1) + ((t8 & one) << 25);
+    r->n[8] = (t8 >> 1) + ((t9 & one) << 25);
+    r->n[9] = (t9 >> 1);
+
+    /* t0..t8: shifted right and added <= C/4 + 1/2
+     *     t9: shifted right
+     *
+     * Current bounds: t0..t8 <= C * (m/2 + 1/2)
+     *                     t9 <= D * (m/2 + 1/4)
+     */
+
+#ifdef VERIFY
+    /* Therefore the output magnitude (M) has to be set such that:
+     *     t0..t8: C * M >= C * (m/2 + 1/2)
+     *         t9: D * M >= D * (m/2 + 1/4)
+     *
+     * It suffices for all limbs that, for any input magnitude m:
+     *     M >= m/2 + 1/2
+     *
+     * and since we want the smallest such integer value for M:
+     *     M == floor(m/2) + 1
+     */
+    r->magnitude = (r->magnitude >> 1) + 1;
+    r->normalized = 0;
+    secp256k1_fe_verify(r);
+#endif
+}
+
 static SECP256K1_INLINE void secp256k1_fe_storage_cmov(secp256k1_fe_storage *r, const secp256k1_fe_storage *a, int flag) {
     uint32_t mask0, mask1;
     VG_CHECK_VERIFY(r->n, sizeof(r->n));

src/field_5x52_impl.h

@@ -58,6 +58,21 @@ static void secp256k1_fe_verify(const secp256k1_fe *a) {
 }
 #endif
 
+static void secp256k1_fe_get_bounds(secp256k1_fe *r, int m) {
+    VERIFY_CHECK(m >= 0);
+    VERIFY_CHECK(m <= 2048);
+    r->n[0] = 0xFFFFFFFFFFFFFULL * 2 * m;
+    r->n[1] = 0xFFFFFFFFFFFFFULL * 2 * m;
+    r->n[2] = 0xFFFFFFFFFFFFFULL * 2 * m;
+    r->n[3] = 0xFFFFFFFFFFFFFULL * 2 * m;
+    r->n[4] = 0x0FFFFFFFFFFFFULL * 2 * m;
+#ifdef VERIFY
+    r->magnitude = m;
+    r->normalized = (m == 0);
+    secp256k1_fe_verify(r);
+#endif
+}
+
 static void secp256k1_fe_normalize(secp256k1_fe *r) {
     uint64_t t0 = r->n[0], t1 = r->n[1], t2 = r->n[2], t3 = r->n[3], t4 = r->n[4];
 
@@ -477,6 +492,71 @@ static SECP256K1_INLINE void secp256k1_fe_cmov(secp256k1_fe *r, const secp256k1_
 #endif
 }
 
+static SECP256K1_INLINE void secp256k1_fe_half(secp256k1_fe *r) {
+    uint64_t t0 = r->n[0], t1 = r->n[1], t2 = r->n[2], t3 = r->n[3], t4 = r->n[4];
+    uint64_t one = (uint64_t)1;
+    uint64_t mask = -(t0 & one) >> 12;
+
+#ifdef VERIFY
+    secp256k1_fe_verify(r);
+    VERIFY_CHECK(r->magnitude < 32);
+#endif
+
+    /* Bounds analysis (over the rationals).
+     *
+     * Let m = r->magnitude
+     *     C = 0xFFFFFFFFFFFFFULL * 2
+     *     D = 0x0FFFFFFFFFFFFULL * 2
+     *
+     * Initial bounds: t0..t3 <= C * m
+     *                     t4 <= D * m
+     */
+
+    t0 += 0xFFFFEFFFFFC2FULL & mask;
+    t1 += mask;
+    t2 += mask;
+    t3 += mask;
+    t4 += mask >> 4;
+
+    VERIFY_CHECK((t0 & one) == 0);
+
+    /* t0..t3: added <= C/2
+     *     t4: added <= D/2
+     *
+     * Current bounds: t0..t3 <= C * (m + 1/2)
+     *                     t4 <= D * (m + 1/2)
+     */
+
+    r->n[0] = (t0 >> 1) + ((t1 & one) << 51);
+    r->n[1] = (t1 >> 1) + ((t2 & one) << 51);
+    r->n[2] = (t2 >> 1) + ((t3 & one) << 51);
+    r->n[3] = (t3 >> 1) + ((t4 & one) << 51);
+    r->n[4] = (t4 >> 1);
+
+    /* t0..t3: shifted right and added <= C/4 + 1/2
+     *     t4: shifted right
+     *
+     * Current bounds: t0..t3 <= C * (m/2 + 1/2)
+     *                     t4 <= D * (m/2 + 1/4)
+     */
+
+#ifdef VERIFY
+    /* Therefore the output magnitude (M) has to be set such that:
+     *     t0..t3: C * M >= C * (m/2 + 1/2)
+     *         t4: D * M >= D * (m/2 + 1/4)
+     *
+     * It suffices for all limbs that, for any input magnitude m:
+     *     M >= m/2 + 1/2
+     *
+     * and since we want the smallest such integer value for M:
+     *     M == floor(m/2) + 1
+     */
+    r->magnitude = (r->magnitude >> 1) + 1;
+    r->normalized = 0;
+    secp256k1_fe_verify(r);
+#endif
+}
+
 static SECP256K1_INLINE void secp256k1_fe_storage_cmov(secp256k1_fe_storage *r, const secp256k1_fe_storage *a, int flag) {
     uint64_t mask0, mask1;
     VG_CHECK_VERIFY(r->n, sizeof(r->n));