mp_size_t
mpn_hgcd_step (mp_size_t n, mp_ptr ap, mp_ptr bp, mp_size_t s,
struct hgcd_matrix *M, mp_ptr tp)
{
struct hgcd_matrix1 M1;
mp_limb_t mask;
mp_limb_t ah, al, bh, bl;
ASSERT (n > s);
mask = ap[n-1] | bp[n-1];
ASSERT (mask > 0);
if (n == s + 1)
{
if (mask < 4)
goto subtract;
ah = ap[n-1]; al = ap[n-2];
bh = bp[n-1]; bl = bp[n-2];
}
else if (mask & GMP_NUMB_HIGHBIT)
{
ah = ap[n-1]; al = ap[n-2];
bh = bp[n-1]; bl = bp[n-2];
}
else
{
int shift;
count_leading_zeros (shift, mask);
ah = MPN_EXTRACT_NUMB (shift, ap[n-1], ap[n-2]);
al = MPN_EXTRACT_NUMB (shift, ap[n-2], ap[n-3]);
bh = MPN_EXTRACT_NUMB (shift, bp[n-1], bp[n-2]);
bl = MPN_EXTRACT_NUMB (shift, bp[n-2], bp[n-3]);
}
/* Try an mpn_hgcd2 step */
if (mpn_hgcd2 (ah, al, bh, bl, &M1))
{
/* Multiply M <- M * M1 */
mpn_hgcd_matrix_mul_1 (M, &M1, tp);
/* Can't swap inputs, so we need to copy. */
MPN_COPY (tp, ap, n);
/* Multiply M1^{-1} (a;b) */
return mpn_matrix22_mul1_inverse_vector (&M1, ap, tp, bp, n);
}
subtract:
return mpn_gcd_subdiv_step (ap, bp, n, s, hgcd_hook, M, tp);
}
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;
struct gcdext_ctx ctx;
mp_size_t un;
mp_ptr u0;
mp_ptr u1;
mp_ptr tp;
TMP_DECL;
ASSERT (an >= n);
ASSERT (n > 0);
ASSERT (bp[n-1] > 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;
ctx.gp = gp;
ctx.up = up;
ctx.usize = usizep;
{
/* 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. */
u1[0] = 1;
ctx.u0 = u0;
ctx.u1 = u1;
ctx.tp = tp + n; /* ualloc */
ctx.un = 1;
/* Temporary storage n */
n = mpn_gcd_subdiv_step (ap, bp, n, 0, mpn_gcdext_hook, &ctx, tp);
if (n == 0)
{
TMP_FREE;
return ctx.gn;
}
un = ctx.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. */
ctx.u0 = u0;
ctx.u1 = u1;
ctx.tp = tp + n; /* ualloc */
ctx.un = un;
/* Temporary storage n */
n = mpn_gcd_subdiv_step (ap, bp, n, 0, mpn_gcdext_hook, &ctx, tp);
if (n == 0)
{
TMP_FREE;
return ctx.gn;
}
un = ctx.un;
ASSERT (un < ualloc);
}
}
/* 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)
We always have u1 > 0, and u0 == 0 is possible only if u1 == 1,
in which case the only reduction done so far is a = A - k B for
some k.
Compute g = u a + v b = (u u1 - v u0) A + (...) B
Here, u, v are bounded by
|u| <= b,
|v| <= a
*/
ASSERT ( (ap[n-1] | bp[n-1]) > 0);
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);
/* c == 0 can happen only when A = (2k+1) G, B = 2 G. And in
this case we choose the cofactor + 1, corresponding to G = A
- k B, rather than -1, corresponding to G = - A + (k+1) B. */
ASSERT (c != 0 || (un == 1 && u0[0] == 1 && u1[0] == 1));
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 (UNLIKELY (u0[0] == 0) && un == 1)
{
mp_size_t gn;
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
{
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);
ASSERT (u0n > 0);
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);
ASSERT (u1n > 0);
ASSERT (lehmer_un + u1n <= ualloc);
ASSERT (lehmer_vn + u0n <= ualloc);
/* 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;
}
}
int
mpn_jacobi_n (mp_ptr ap, mp_ptr bp, mp_size_t n, unsigned bits)
{
mp_size_t scratch;
mp_size_t matrix_scratch;
mp_ptr tp;
TMP_DECL;
ASSERT (n > 0);
ASSERT ( (ap[n-1] | bp[n-1]) > 0);
ASSERT ( (bp[0] | ap[0]) & 1);
/* FIXME: Check for small sizes first, before setting up temporary
storage etc. */
scratch = MPN_GCD_SUBDIV_STEP_ITCH(n);
if (ABOVE_THRESHOLD (n, GCD_DC_THRESHOLD))
{
mp_size_t hgcd_scratch;
mp_size_t update_scratch;
mp_size_t p = CHOOSE_P (n);
mp_size_t dc_scratch;
matrix_scratch = MPN_HGCD_MATRIX_INIT_ITCH (n - p);
hgcd_scratch = mpn_hgcd_itch (n - p);
update_scratch = p + n - 1;
dc_scratch = matrix_scratch + MAX(hgcd_scratch, update_scratch);
if (dc_scratch > scratch)
scratch = dc_scratch;
}
TMP_MARK;
tp = TMP_ALLOC_LIMBS(scratch);
while (ABOVE_THRESHOLD (n, JACOBI_DC_THRESHOLD))
{
struct hgcd_matrix M;
mp_size_t p = 2*n/3;
mp_size_t matrix_scratch = MPN_HGCD_MATRIX_INIT_ITCH (n - p);
mp_size_t nn;
mpn_hgcd_matrix_init (&M, n - p, tp);
nn = mpn_hgcd_jacobi (ap + p, bp + p, n - p, &M, &bits,
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);
}
else
{
/* Temporary storage n */
n = mpn_gcd_subdiv_step (ap, bp, n, 0, jacobi_hook, &bits, tp);
if (!n)
{
TMP_FREE;
return bits == BITS_FAIL ? 0 : mpn_jacobi_finish (bits);
}
}
}
while (n > 2)
{
struct hgcd_matrix1 M;
mp_limb_t ah, al, bh, bl;
mp_limb_t mask;
mask = ap[n-1] | bp[n-1];
ASSERT (mask > 0);
if (mask & GMP_NUMB_HIGHBIT)
{
ah = ap[n-1]; al = ap[n-2];
bh = bp[n-1]; bl = bp[n-2];
}
else
{
int shift;
count_leading_zeros (shift, mask);
ah = MPN_EXTRACT_NUMB (shift, ap[n-1], ap[n-2]);
al = MPN_EXTRACT_NUMB (shift, ap[n-2], ap[n-3]);
bh = MPN_EXTRACT_NUMB (shift, bp[n-1], bp[n-2]);
bl = MPN_EXTRACT_NUMB (shift, bp[n-2], bp[n-3]);
}
/* Try an mpn_nhgcd2 step */
if (mpn_hgcd2_jacobi (ah, al, bh, bl, &M, &bits))
{
n = mpn_matrix22_mul1_inverse_vector (&M, tp, ap, bp, n);
MP_PTR_SWAP (ap, tp);
}
else
{
/* mpn_hgcd2 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. */
n = mpn_gcd_subdiv_step (ap, bp, n, 0, &jacobi_hook, &bits, tp);
if (!n)
{
TMP_FREE;
return bits == BITS_FAIL ? 0 : mpn_jacobi_finish (bits);
}
}
}
if (bits >= 16)
MP_PTR_SWAP (ap, bp);
ASSERT (bp[0] & 1);
if (n == 1)
{
mp_limb_t al, bl;
al = ap[0];
bl = bp[0];
TMP_FREE;
if (bl == 1)
return 1 - 2*(bits & 1);
else
return mpn_jacobi_base (al, bl, bits << 1);
}
else
{
int res = mpn_jacobi_2 (ap, bp, bits & 1);
TMP_FREE;
return res;
}
}
