mp_size_t
mpn_gcdext (mp_ptr gp, mp_ptr up, mp_size_t *usizep,
mp_ptr ap, mp_size_t an, mp_ptr bp, mp_size_t n)
{
mp_size_t talloc;
mp_size_t scratch;
mp_size_t matrix_scratch;
mp_size_t ualloc = n + 1;
mp_size_t un;
mp_ptr u0;
mp_ptr u1;
mp_ptr tp;
TMP_DECL;
ASSERT (an >= n);
ASSERT (n > 0);
TMP_MARK;
/* FIXME: Check for small sizes first, before setting up temporary
storage etc. */
talloc = MPN_GCDEXT_LEHMER_N_ITCH(n);
/* For initial division */
scratch = an - n + 1;
if (scratch > talloc)
talloc = scratch;
if (ABOVE_THRESHOLD (n, GCDEXT_DC_THRESHOLD))
{
/* For hgcd loop. */
mp_size_t hgcd_scratch;
mp_size_t update_scratch;
mp_size_t p1 = CHOOSE_P_1 (n);
mp_size_t p2 = CHOOSE_P_2 (n);
mp_size_t min_p = MIN(p1, p2);
mp_size_t max_p = MAX(p1, p2);
matrix_scratch = MPN_HGCD_MATRIX_INIT_ITCH (n - min_p);
hgcd_scratch = mpn_hgcd_itch (n - min_p);
update_scratch = max_p + n - 1;
scratch = matrix_scratch + MAX(hgcd_scratch, update_scratch);
if (scratch > talloc)
talloc = scratch;
/* Final mpn_gcdext_lehmer_n call. Need space for u and for
copies of a and b. */
scratch = MPN_GCDEXT_LEHMER_N_ITCH (GCDEXT_DC_THRESHOLD)
+ 3*GCDEXT_DC_THRESHOLD;
if (scratch > talloc)
talloc = scratch;
/* Cofactors u0 and u1 */
talloc += 2*(n+1);
}
tp = TMP_ALLOC_LIMBS(talloc);
if (an > n)
{
mpn_tdiv_qr (tp, ap, 0, ap, an, bp, n);
if (mpn_zero_p (ap, n))
{
MPN_COPY (gp, bp, n);
*usizep = 0;
TMP_FREE;
return n;
}
}
if (BELOW_THRESHOLD (n, GCDEXT_DC_THRESHOLD))
{
mp_size_t gn = mpn_gcdext_lehmer_n(gp, up, usizep, ap, bp, n, tp);
TMP_FREE;
return gn;
}
MPN_ZERO (tp, 2*ualloc);
u0 = tp; tp += ualloc;
u1 = tp; tp += ualloc;
{
/* For the first hgcd call, there are no u updates, and it makes
some sense to use a different choice for p. */
/* FIXME: We could trim use of temporary storage, since u0 and u1
are not used yet. For the hgcd call, we could swap in the u0
and u1 pointers for the relevant matrix elements. */
struct hgcd_matrix M;
mp_size_t p = CHOOSE_P_1 (n);
mp_size_t nn;
mpn_hgcd_matrix_init (&M, n - p, tp);
nn = mpn_hgcd (ap + p, bp + p, n - p, &M, tp + matrix_scratch);
if (nn > 0)
{
ASSERT (M.n <= (n - p - 1)/2);
ASSERT (M.n + p <= (p + n - 1) / 2);
/* Temporary storage 2 (p + M->n) <= p + n - 1 */
n = mpn_hgcd_matrix_adjust (&M, p + nn, ap, bp, p, tp + matrix_scratch);
MPN_COPY (u0, M.p[1][0], M.n);
MPN_COPY (u1, M.p[1][1], M.n);
un = M.n;
while ( (u0[un-1] | u1[un-1] ) == 0)
un--;
}
else
{
/* mpn_hgcd has failed. Then either one of a or b is very
small, or the difference is very small. Perform one
subtraction followed by one division. */
mp_size_t gn;
mp_size_t updated_un = 1;
u1[0] = 1;
/* Temporary storage 2n + 1 */
n = mpn_gcdext_subdiv_step (gp, &gn, up, usizep, ap, bp, n,
u0, u1, &updated_un, tp, tp + n);
if (n == 0)
{
TMP_FREE;
return gn;
}
un = updated_un;
ASSERT (un < ualloc);
}
}
while (ABOVE_THRESHOLD (n, GCDEXT_DC_THRESHOLD))
{
struct hgcd_matrix M;
mp_size_t p = CHOOSE_P_2 (n);
mp_size_t nn;
mpn_hgcd_matrix_init (&M, n - p, tp);
nn = mpn_hgcd (ap + p, bp + p, n - p, &M, tp + matrix_scratch);
if (nn > 0)
{
mp_ptr t0;
t0 = tp + matrix_scratch;
ASSERT (M.n <= (n - p - 1)/2);
ASSERT (M.n + p <= (p + n - 1) / 2);
/* Temporary storage 2 (p + M->n) <= p + n - 1 */
n = mpn_hgcd_matrix_adjust (&M, p + nn, ap, bp, p, t0);
/* By the same analysis as for mpn_hgcd_matrix_mul */
ASSERT (M.n + un <= ualloc);
/* FIXME: This copying could be avoided by some swapping of
* pointers. May need more temporary storage, though. */
MPN_COPY (t0, u0, un);
/* Temporary storage ualloc */
un = hgcd_mul_matrix_vector (&M, u0, t0, u1, un, t0 + un);
ASSERT (un < ualloc);
ASSERT ( (u0[un-1] | u1[un-1]) > 0);
}
else
{
/* mpn_hgcd has failed. Then either one of a or b is very
small, or the difference is very small. Perform one
subtraction followed by one division. */
mp_size_t gn;
mp_size_t updated_un = un;
/* Temporary storage 2n + 1 */
n = mpn_gcdext_subdiv_step (gp, &gn, up, usizep, ap, bp, n,
u0, u1, &updated_un, tp, tp + n);
if (n == 0)
{
TMP_FREE;
return gn;
}
un = updated_un;
ASSERT (un < ualloc);
}
}
if (UNLIKELY (mpn_cmp (ap, bp, n) == 0))
{
/* Must return the smallest cofactor, +u1 or -u0 */
int c;
MPN_COPY (gp, ap, n);
MPN_CMP (c, u0, u1, un);
ASSERT (c != 0);
if (c < 0)
{
MPN_NORMALIZE (u0, un);
MPN_COPY (up, u0, un);
*usizep = -un;
}
else
{
MPN_NORMALIZE_NOT_ZERO (u1, un);
MPN_COPY (up, u1, un);
*usizep = un;
}
TMP_FREE;
return n;
}
else if (mpn_zero_p (u0, un))
{
mp_size_t gn;
ASSERT (un == 1);
ASSERT (u1[0] == 1);
/* g = u a + v b = (u u1 - v u0) A + (...) B = u A + (...) B */
gn = mpn_gcdext_lehmer_n (gp, up, usizep, ap, bp, n, tp);
TMP_FREE;
return gn;
}
else
{
/* We have A = ... a + ... b
B = u0 a + u1 b
a = u1 A + ... B
b = -u0 A + ... B
with bounds
|u0|, |u1| <= B / min(a, b)
Compute g = u a + v b = (u u1 - v u0) A + (...) B
Here, u, v are bounded by
|u| <= b,
|v| <= a
*/
mp_size_t u0n;
mp_size_t u1n;
mp_size_t lehmer_un;
mp_size_t lehmer_vn;
mp_size_t gn;
mp_ptr lehmer_up;
mp_ptr lehmer_vp;
int negate;
lehmer_up = tp; tp += n;
/* Call mpn_gcdext_lehmer_n with copies of a and b. */
MPN_COPY (tp, ap, n);
MPN_COPY (tp + n, bp, n);
gn = mpn_gcdext_lehmer_n (gp, lehmer_up, &lehmer_un, tp, tp + n, n, tp + 2*n);
u0n = un;
MPN_NORMALIZE (u0, u0n);
if (lehmer_un == 0)
{
/* u == 0 ==> v = g / b == 1 ==> g = - u0 A + (...) B */
MPN_COPY (up, u0, u0n);
*usizep = -u0n;
TMP_FREE;
return gn;
}
lehmer_vp = tp;
/* Compute v = (g - u a) / b */
lehmer_vn = compute_v (lehmer_vp,
ap, bp, n, gp, gn, lehmer_up, lehmer_un, tp + n + 1);
if (lehmer_un > 0)
negate = 0;
else
{
lehmer_un = -lehmer_un;
negate = 1;
}
u1n = un;
MPN_NORMALIZE (u1, u1n);
/* It's possible that u0 = 1, u1 = 0 */
if (u1n == 0)
{
ASSERT (un == 1);
ASSERT (u0[0] == 1);
/* u1 == 0 ==> u u1 + v u0 = v */
MPN_COPY (up, lehmer_vp, lehmer_vn);
*usizep = negate ? lehmer_vn : - lehmer_vn;
TMP_FREE;
return gn;
}
ASSERT (lehmer_un + u1n <= ualloc);
ASSERT (lehmer_vn + u0n <= ualloc);
/* Now u0, u1, u are non-zero. We may still have v == 0 */
/* Compute u u0 */
if (lehmer_un <= u1n)
/* Should be the common case */
mpn_mul (up, u1, u1n, lehmer_up, lehmer_un);
else
mpn_mul (up, lehmer_up, lehmer_un, u1, u1n);
un = u1n + lehmer_un;
un -= (up[un - 1] == 0);
if (lehmer_vn > 0)
{
mp_limb_t cy;
/* Overwrites old u1 value */
if (lehmer_vn <= u0n)
/* Should be the common case */
mpn_mul (u1, u0, u0n, lehmer_vp, lehmer_vn);
else
mpn_mul (u1, lehmer_vp, lehmer_vn, u0, u0n);
u1n = u0n + lehmer_vn;
u1n -= (u1[u1n - 1] == 0);
if (u1n <= un)
{
cy = mpn_add (up, up, un, u1, u1n);
}
else
{
cy = mpn_add (up, u1, u1n, up, un);
un = u1n;
}
up[un] = cy;
un += (cy != 0);
ASSERT (un < ualloc);
}
*usizep = negate ? -un : un;
TMP_FREE;
return gn;
}
}
/* Temporary storage: Needs n limbs for the quotient, at qp. tp must
point to an area large enough for the resulting cofactor, plus one
limb extra. All in all, 2N + 1 if N is a bound for both inputs and
outputs. */
mp_size_t
mpn_gcdext_subdiv_step (mp_ptr gp, mp_size_t *gn, mp_ptr up, mp_size_t *usizep,
mp_ptr ap, mp_ptr bp, mp_size_t n,
mp_ptr u0, mp_ptr u1, mp_size_t *unp,
mp_ptr qp, mp_ptr tp)
{
mp_size_t an, bn, un;
mp_size_t qn;
mp_size_t u0n;
int swapped;
an = bn = n;
ASSERT (an > 0);
ASSERT (ap[an-1] > 0 || bp[an-1] > 0);
MPN_NORMALIZE (ap, an);
MPN_NORMALIZE (bp, bn);
un = *unp;
swapped = 0;
if (UNLIKELY (an == 0))
{
return_b:
MPN_COPY (gp, bp, bn);
*gn = bn;
MPN_NORMALIZE (u0, un);
MPN_COPY (up, u0, un);
*usizep = swapped ? un : -un;
return 0;
}
else if (UNLIKELY (bn == 0))
{
MPN_COPY (gp, ap, an);
*gn = an;
MPN_NORMALIZE (u1, un);
MPN_COPY (up, u1, un);
*usizep = swapped ? -un : un;
return 0;
}
/* Arrange so that a > b, subtract an -= bn, and maintain
normalization. */
if (an < bn)
{
MPN_PTR_SWAP (ap, an, bp, bn);
MP_PTR_SWAP (u0, u1);
swapped ^= 1;
}
else if (an == bn)
{
int c;
MPN_CMP (c, ap, bp, an);
if (UNLIKELY (c == 0))
{
MPN_COPY (gp, ap, an);
*gn = an;
/* Must return the smallest cofactor, +u1 or -u0 */
MPN_CMP (c, u0, u1, un);
ASSERT (c != 0 || (un == 1 && u0[0] == 1 && u1[0] == 1));
if (c < 0)
{
MPN_NORMALIZE (u0, un);
MPN_COPY (up, u0, un);
swapped ^= 1;
}
else
{
MPN_NORMALIZE_NOT_ZERO (u1, un);
MPN_COPY (up, u1, un);
}
*usizep = swapped ? -un : un;
return 0;
}
else if (c < 0)
{
MP_PTR_SWAP (ap, bp);
MP_PTR_SWAP (u0, u1);
swapped ^= 1;
}
}
/* Reduce a -= b, u1 += u0 */
ASSERT_NOCARRY (mpn_sub (ap, ap, an, bp, bn));
MPN_NORMALIZE (ap, an);
ASSERT (an > 0);
u1[un] = mpn_add_n (u1, u1, u0, un);
un += (u1[un] > 0);
/* Arrange so that a > b, and divide a = q b + r */
if (an < bn)
{
MPN_PTR_SWAP (ap, an, bp, bn);
MP_PTR_SWAP (u0, u1);
swapped ^= 1;
}
else if (an == bn)
{
int c;
MPN_CMP (c, ap, bp, an);
if (UNLIKELY (c == 0))
goto return_b;
else if (c < 0)
{
MP_PTR_SWAP (ap, bp);
MP_PTR_SWAP (u0, u1);
swapped ^= 1;
}
}
/* Reduce a -= q b, u1 += q u0 */
qn = an - bn + 1;
mpn_tdiv_qr (qp, ap, 0, ap, an, bp, bn);
if (mpn_zero_p (ap, bn))
goto return_b;
n = bn;
/* Update u1 += q u0 */
u0n = un;
MPN_NORMALIZE (u0, u0n);
if (u0n > 0)
{
qn -= (qp[qn - 1] == 0);
if (qn > u0n)
mpn_mul (tp, qp, qn, u0, u0n);
else
mpn_mul (tp, u0, u0n, qp, qn);
if (qn + u0n > un)
{
mp_size_t u1n = un;
un = qn + u0n;
un -= (tp[un-1] == 0);
u1[un] = mpn_add (u1, tp, un, u1, u1n);
}
else
{
u1[un] = mpn_add (u1, u1, un, tp, qn + u0n);
}
un += (u1[un] > 0);
}
*unp = un;
return n;
}
