/* Put in rp[0..n] the n+1 low limbs of {np, n} * {mp, n}.
Assume 2n limbs are allocated at rp. */
void
mpfr_mullow_n (mpfr_limb_ptr rp, mpfr_limb_srcptr np, mpfr_limb_srcptr mp,
mp_size_t n)
{
mp_size_t k;
MPFR_ASSERTN (MPFR_MULHIGH_TAB_SIZE >= 8); /* so that 3*(n/4) > n/2 */
k = MPFR_LIKELY (n < MPFR_MULHIGH_TAB_SIZE) ? mulhigh_ktab[n] : 3*(n/4);
MPFR_ASSERTD (k == -1 || k == 0 || (2 * k >= n && k < n));
if (k < 0)
mpn_mul_basecase (rp, np, n, mp, n);
else if (k == 0)
mpfr_mullow_n_basecase (rp, np, mp, n);
else if (n > MUL_FFT_THRESHOLD)
mpn_mul_n (rp, np, mp, n);
else
{
mp_size_t l = n - k;
mpn_mul_n (rp, np, mp, k); /* fills rp[0..2k] */
mpfr_mullow_n (rp + n, np + k, mp, l); /* fills rp[n..n+2l] */
mpn_add_n (rp + k, rp + k, rp + n, l + 1);
mpfr_mullow_n (rp + n, np, mp + k, l); /* fills rp[n..n+2l] */
mpn_add_n (rp + k, rp + k, rp + n, l + 1);
}
}
/* (rp, 2n) = (xp, n)*(yp, n) */
void
mpn_mulhigh_n (mp_ptr rp, mp_srcptr xp, mp_srcptr yp, mp_size_t n)
{
mp_limb_t t;
ASSERT(n > 0);
ASSERT_MPN(xp, n);
ASSERT_MPN(yp, n);
ASSERT(!MPN_OVERLAP_P(rp, 2 * n, xp, n));
ASSERT(!MPN_OVERLAP_P(rp, 2 * n, yp, n));
if (BELOW_THRESHOLD(n, MULHIGH_BASECASE_THRESHOLD))
{
mpn_mul_basecase(rp, xp, n, yp, n);
return;
}
if (ABOVE_THRESHOLD (n, MULHIGH_MUL_THRESHOLD))
{
mpn_mul_n(rp, xp, yp, n);
return;
}
mpn_mulshort_n(rp, xp, yp, n);
t = rp[n - 1] + n - 2;
if (UNLIKELY(t < n - 2))
mpn_mul_n(rp, xp, yp, n);
return;
}
/* Put in rp[n..2n-1] an approximation of the n high limbs
of {np, n} * {mp, n}. The error is less than n ulps of rp[n] (and the
approximation is always less or equal to the truncated full product).
Implements Algorithm ShortMul from [1].
*/
void
mpfr_mulhigh_n (mpfr_limb_ptr rp, mpfr_limb_srcptr np, mpfr_limb_srcptr mp,
mp_size_t n)
{
mp_size_t k;
MPFR_ASSERTN (MPFR_MULHIGH_TAB_SIZE >= 8); /* so that 3*(n/4) > n/2 */
k = MPFR_LIKELY (n < MPFR_MULHIGH_TAB_SIZE) ? mulhigh_ktab[n] : 3*(n/4);
/* Algorithm ShortMul from [1] requires k >= (n+3)/2, which translates
into k >= (n+4)/2 in the C language. */
MPFR_ASSERTD (k == -1 || k == 0 || (k >= (n+4)/2 && k < n));
if (k < 0)
mpn_mul_basecase (rp, np, n, mp, n); /* result is exact, no error */
else if (k == 0)
mpfr_mulhigh_n_basecase (rp, np, mp, n); /* basecase error < n ulps */
else if (n > MUL_FFT_THRESHOLD)
mpn_mul_n (rp, np, mp, n); /* result is exact, no error */
else
{
mp_size_t l = n - k;
mp_limb_t cy;
mpn_mul_n (rp + 2 * l, np + l, mp + l, k); /* fills rp[2l..2n-1] */
mpfr_mulhigh_n (rp, np + k, mp, l); /* fills rp[l-1..2l-1] */
cy = mpn_add_n (rp + n - 1, rp + n - 1, rp + l - 1, l + 1);
mpfr_mulhigh_n (rp, np, mp + k, l); /* fills rp[l-1..2l-1] */
cy += mpn_add_n (rp + n - 1, rp + n - 1, rp + l - 1, l + 1);
mpn_add_1 (rp + n + l, rp + n + l, k, cy); /* propagate carry */
}
}
void
mpn_sqr_n (mp_ptr p, mp_srcptr a, mp_size_t n)
{
ASSERT (n >= 1);
ASSERT (! MPN_OVERLAP_P (p, 2 * n, a, n));
#if 0
/* FIXME: Can this be removed? */
if (n == 0)
return;
#endif
if (BELOW_THRESHOLD (n, SQR_BASECASE_THRESHOLD))
{ /* mul_basecase is faster than sqr_basecase on small sizes sometimes */
mpn_mul_basecase (p, a, n, a, n);
}
else if (BELOW_THRESHOLD (n, SQR_KARATSUBA_THRESHOLD))
{
mpn_sqr_basecase (p, a, n);
}
else if (BELOW_THRESHOLD (n, SQR_TOOM3_THRESHOLD))
{
/* Allocate workspace of fixed size on stack: fast! */
mp_limb_t ws[MPN_KARA_SQR_N_TSIZE (SQR_TOOM3_THRESHOLD_LIMIT-1)];
ASSERT (SQR_TOOM3_THRESHOLD <= SQR_TOOM3_THRESHOLD_LIMIT);
mpn_kara_sqr_n (p, a, n, ws);
}
#if WANT_FFT || TUNE_PROGRAM_BUILD
else if (BELOW_THRESHOLD (n, SQR_FFT_THRESHOLD))
#else
else if (BELOW_THRESHOLD (n, MPN_TOOM3_MAX_N))
#endif
{
mp_ptr ws;
TMP_SDECL;
TMP_SMARK;
ws = TMP_SALLOC_LIMBS (MPN_TOOM3_SQR_N_TSIZE (n));
mpn_toom3_sqr_n (p, a, n, ws);
TMP_SFREE;
}
else
#if WANT_FFT || TUNE_PROGRAM_BUILD
{
/* The current FFT code allocates its own space. That should probably
change. */
mpn_mul_fft_full (p, a, n, a, n);
}
#else
{
/* Toom3 for large operands. Use workspace from the heap, as stack space
may be limited. Since n is at least MUL_TOOM3_THRESHOLD, multiplication
will take much longer than malloc()/free(). */
mp_ptr ws; mp_size_t ws_size;
ws_size = MPN_TOOM3_SQR_N_TSIZE (n);
ws = __GMP_ALLOCATE_FUNC_LIMBS (ws_size);
mpn_toom3_sqr_n (p, a, n, ws);
__GMP_FREE_FUNC_LIMBS (ws, ws_size);
}
#endif
}
void
mpn_sec_mul (mp_ptr rp,
mp_srcptr ap, mp_size_t an,
mp_srcptr bp, mp_size_t bn,
mp_ptr tp)
{
mpn_mul_basecase (rp, ap, an, bp, bn);
}
// (rp,2n)=(xp,n)*(yp,n) with temp space (tp,2*n+C)
void mpn_kara_mul_n(mp_ptr rp,mp_srcptr xp,mp_srcptr yp,mp_size_t n,mp_ptr tp)
{mp_size_t n2,n3;mp_srcptr xl,yl,xh,yh;mp_ptr dx,dy;int suboradd;mp_limb_t c;
n2=n>>1;suboradd=-1;
xl=xp;xh=xp+n2;yl=yp;yh=yp+n2;n3=n-n2;
dx=rp+2*n2;dy=dx+n3;
if((n&1)==0)
{if(mpn_cmp(xh,xl,n2)>=0){mpn_sub_n(dx,xh,xl,n2);}else{mpn_sub_n(dx,xl,xh,n2);suboradd=-suboradd;}
if(mpn_cmp(yh,yl,n2)>=0){mpn_sub_n(dy,yh,yl,n2);}else{mpn_sub_n(dy,yl,yh,n2);suboradd=-suboradd;}}
else
{if(xh[n2]!=0 || mpn_cmp(xh,xl,n2)>=0){c=mpn_sub_n(dx,xh,xl,n2);dx[n2]=xh[n2]-c;}else{mpn_sub_n(dx,xl,xh,n2);dx[n2]=0;suboradd=-suboradd;}
if(yh[n2]!=0 || mpn_cmp(yh,yl,n2)>=0){c=mpn_sub_n(dy,yh,yl,n2);dy[n2]=yh[n2]-c;}else{mpn_sub_n(dy,yl,yh,n2);dy[n2]=0;suboradd=-suboradd;}}
if(BELOW_THRESHOLD(n3,MUL_KARATSUBA_THRESHOLD))
{mpn_mul_basecase(rp,xl,n2,yl,n2);
mpn_mul_basecase(tp,dx,n3,dy,n3);
mpn_mul_basecase(rp+2*n2,xh,n3,yh,n3);}
else
{mpn_kara_mul_n(rp,xl,yl,n2,tp+2*n3);
mpn_kara_mul_n(tp,dx,dy,n3,tp+2*n3);
mpn_kara_mul_n(rp+2*n2,xh,yh,n3,tp+2*n3);}
if(suboradd==-1){mpn_karasub(rp,tp,n);}else{mpn_karaadd(rp,tp,n);}
return;}
void
mpn_sqr (mp_ptr p, mp_srcptr a, mp_size_t n)
{
ASSERT (n >= 1);
ASSERT (! MPN_OVERLAP_P (p, 2 * n, a, n));
#if 0
/* FIXME: Can this be removed? */
if (n == 0)
return;
#endif
if (BELOW_THRESHOLD (n, SQR_BASECASE_THRESHOLD))
{ /* mul_basecase is faster than sqr_basecase on small sizes sometimes */
mpn_mul_basecase (p, a, n, a, n);
}
else if (BELOW_THRESHOLD (n, SQR_KARATSUBA_THRESHOLD))
{
mpn_sqr_basecase (p, a, n);
}
else if (BELOW_THRESHOLD (n, SQR_TOOM3_THRESHOLD))
{
/* Allocate workspace of fixed size on stack: fast! */
mp_limb_t ws[MPN_KARA_SQR_N_TSIZE (SQR_TOOM3_THRESHOLD_LIMIT-1)];
ASSERT (SQR_TOOM3_THRESHOLD <= SQR_TOOM3_THRESHOLD_LIMIT);
mpn_kara_sqr_n (p, a, n, ws);
}
else if (BELOW_THRESHOLD (n, SQR_TOOM4_THRESHOLD))
{
mp_ptr ws;
TMP_SDECL;
TMP_SMARK;
ws = TMP_SALLOC_LIMBS (MPN_TOOM3_SQR_N_TSIZE (n));
mpn_toom3_sqr_n (p, a, n, ws);
TMP_SFREE;
}
else if (BELOW_THRESHOLD (n, SQR_TOOM8_THRESHOLD))
{
mpn_toom4_sqr_n (p, a, n);
}
#if WANT_FFT || TUNE_PROGRAM_BUILD
else if (BELOW_THRESHOLD (n, SQR_FFT_FULL_THRESHOLD))
#else
else
#endif
{
mpn_toom8_sqr_n (p, a, n);
}
#if WANT_FFT || TUNE_PROGRAM_BUILD
else
{
void
mpn_sec_sqr (mp_ptr rp,
mp_srcptr ap, mp_size_t an,
mp_ptr tp)
{
#ifndef SQR_BASECASE_LIM
/* If SQR_BASECASE_LIM is now not defined, use mpn_sqr_basecase for any operand
size. */
mpn_sqr_basecase (rp, ap, an);
#else
/* Else use mpn_sqr_basecase for its allowed sizes, else mpn_mul_basecase. */
mpn_mul_basecase (rp, ap, an, ap, an);
#endif
}
void
mpn_mul_n (mp_ptr p, mp_srcptr a, mp_srcptr b, mp_size_t n)
{
ASSERT (n >= 1);
ASSERT (! MPN_OVERLAP_P (p, 2 * n, a, n));
ASSERT (! MPN_OVERLAP_P (p, 2 * n, b, n));
if (BELOW_THRESHOLD (n, MUL_KARATSUBA_THRESHOLD))
{
mpn_mul_basecase (p, a, n, b, n);
}
else if (BELOW_THRESHOLD (n, MUL_TOOM3_THRESHOLD))
{
/* Allocate workspace of fixed size on stack: fast! */
mp_limb_t ws[MPN_KARA_MUL_N_TSIZE (MUL_TOOM3_THRESHOLD_LIMIT-1)];
ASSERT (MUL_TOOM3_THRESHOLD <= MUL_TOOM3_THRESHOLD_LIMIT);
mpn_kara_mul_n (p, a, b, n, ws);
}
else if (BELOW_THRESHOLD (n, MUL_TOOM4_THRESHOLD))
{
mp_ptr ws;
TMP_SDECL;
TMP_SMARK;
ws = TMP_SALLOC_LIMBS (MPN_TOOM3_MUL_N_TSIZE (n));
mpn_toom3_mul_n (p, a, b, n, ws);
TMP_SFREE;
}
else if (BELOW_THRESHOLD (n, MUL_TOOM8H_THRESHOLD))
{
mpn_toom4_mul_n (p, a, b, n);
}
#if WANT_FFT || TUNE_PROGRAM_BUILD
else if (BELOW_THRESHOLD (n, MUL_FFT_FULL_THRESHOLD))
{
mpn_toom8h_mul (p, a, n, b, n);
}
#endif
else
#if WANT_FFT || TUNE_PROGRAM_BUILD
{
mpn_mul_fft_main(p, a, n, b, n);
}
#else
{
/* Toom8 for large operands. */
mpn_toom8h_mul (p, a, n, b, n);
}
#endif
}
// (rp,2n)=(xp,n)^2 with temp space (tp,2*n+C)
void mpn_kara_sqr_n(mp_ptr rp,mp_srcptr xp,mp_size_t n,mp_ptr tp)
{mp_size_t n2,n3;mp_srcptr xl,xh;mp_ptr dx;mp_limb_t c;
n2=n>>1;
xl=xp;xh=xp+n2;n3=n-n2;
dx=rp+2*n2;
if((n&1)==0)
{if(mpn_cmp(xh,xl,n2)>=0){mpn_sub_n(dx,xh,xl,n2);}else{mpn_sub_n(dx,xl,xh,n2);}}
else
{if(xh[n2]!=0 || mpn_cmp(xh,xl,n2)>=0){c=mpn_sub_n(dx,xh,xl,n2);dx[n2]=xh[n2]-c;}else{mpn_sub_n(dx,xl,xh,n2);dx[n2]=0;}}
if(BELOW_THRESHOLD(n3,MUL_KARATSUBA_THRESHOLD))
{mpn_mul_basecase(rp,xl,n2,xl,n2);
mpn_mul_basecase(tp,dx,n3,dx,n3);
mpn_mul_basecase(rp+2*n2,xh,n3,xh,n3);}
else if(BELOW_THRESHOLD(n3,SQR_KARATSUBA_THRESHOLD))
{mpn_sqr_basecase(rp,xl,n2);
mpn_sqr_basecase(tp,dx,n3);
mpn_sqr_basecase(rp+2*n2,xh,n3);}
else
{mpn_kara_sqr_n(rp,xl,n2,tp+2*n3);
mpn_kara_sqr_n(tp,dx,n3,tp+2*n3);
mpn_kara_sqr_n(rp+2*n2,xh,n3,tp+2*n3);}
mpn_karasub(rp,tp,n);
return;}
void
mpn_sqr_n (mp_ptr prodp,
mp_srcptr up, mp_size_t un)
{
ASSERT (un >= 1);
ASSERT (! MPN_OVERLAP_P (prodp, 2*un, up, un));
/* FIXME: Can this be removed? */
if (un == 0)
return;
if (BELOW_THRESHOLD (un, SQR_BASECASE_THRESHOLD))
{ /* mul_basecase is faster than sqr_basecase on small sizes sometimes */
mpn_mul_basecase (prodp, up, un, up, un);
}
else if (BELOW_THRESHOLD (un, SQR_KARATSUBA_THRESHOLD))
{ /* plain schoolbook multiplication */
mpn_sqr_basecase (prodp, up, un);
}
else if (BELOW_THRESHOLD (un, SQR_TOOM3_THRESHOLD))
{ /* karatsuba multiplication */
mp_ptr tspace;
TMP_DECL (marker);
TMP_MARK (marker);
tspace = TMP_ALLOC_LIMBS (MPN_KARA_SQR_N_TSIZE (un));
mpn_kara_sqr_n (prodp, up, un, tspace);
TMP_FREE (marker);
}
#if WANT_FFT || TUNE_PROGRAM_BUILD
else if (BELOW_THRESHOLD (un, SQR_FFT_THRESHOLD))
#else
else
#endif
{ /* Toom3 multiplication.
Use workspace from the heap, as stack may be limited. Since n is
at least MUL_TOOM3_THRESHOLD, the multiplication will take much
longer than malloc()/free(). */
mp_ptr tspace;
mp_size_t tsize;
tsize = MPN_TOOM3_SQR_N_TSIZE (un);
tspace = __GMP_ALLOCATE_FUNC_LIMBS (tsize);
mpn_toom3_sqr_n (prodp, up, un, tspace);
__GMP_FREE_FUNC_LIMBS (tspace, tsize);
}
#if WANT_FFT || TUNE_PROGRAM_BUILD
else
{
void
mpn_mullow_n (mp_ptr rp, mp_srcptr xp, mp_srcptr yp, mp_size_t n)
{
if (BELOW_THRESHOLD (n, MULLOW_BASECASE_THRESHOLD))
{
/* Allocate workspace of fixed size on stack: fast! */
mp_limb_t ws[MUL_BASECASE_ALLOC];
mpn_mul_basecase (ws, xp, n, yp, n);
MPN_COPY (rp, ws, n);
}
else if (BELOW_THRESHOLD (n, MULLOW_DC_THRESHOLD))
{
mpn_mullow_basecase (rp, xp, yp, n);
}
else if (BELOW_THRESHOLD (n, MULLOW_MUL_N_THRESHOLD))
{
/* Divide-and-conquer */
mp_size_t n2 = n >> 1; /* floor(n/2) */
mp_size_t n1 = n - n2; /* ceil(n/2) */
mp_ptr tp;
TMP_SDECL;
TMP_SMARK;
tp = TMP_SALLOC_LIMBS (n1);
/* Split as x = x1 2^(n1 GMP_NUMB_BITS) + x0,
y = y1 2^(n2 GMP_NUMB_BITS) + y0 */
/* x0 * y0 */
mpn_mul_n (rp, xp, yp, n2);
if (n1 != n2)
rp[2 * n2] = mpn_addmul_1 (rp + n2, yp, n2, xp[n2]);
/* x1 * y0 * 2^(n1 GMP_NUMB_BITS) */
mpn_mullow_n (tp, xp + n1, yp, n2);
mpn_add_n (rp + n1, rp + n1, tp, n2);
/* x0 * y1 * 2^(n2 GMP_NUMB_BITS) */
mpn_mullow_n (tp, yp + n2, xp, n1);
mpn_add_n (rp + n2, rp + n2, tp, n1);
TMP_SFREE;
}
else
{
void
check (void)
{
mp_limb_t wp[100], xp[100], yp[100];
mp_size_t size = 100;
refmpn_zero (xp, size);
refmpn_zero (yp, size);
refmpn_zero (wp, size);
pre ("mpn_add_n");
mpn_add_n (wp, xp, yp, size);
post ();
#if HAVE_NATIVE_mpn_add_nc
pre ("mpn_add_nc");
mpn_add_nc (wp, xp, yp, size, CNST_LIMB(0));
post ();
#endif
#if HAVE_NATIVE_mpn_addlsh1_n
pre ("mpn_addlsh1_n");
mpn_addlsh1_n (wp, xp, yp, size);
post ();
#endif
#if HAVE_NATIVE_mpn_and_n
pre ("mpn_and_n");
mpn_and_n (wp, xp, yp, size);
post ();
#endif
#if HAVE_NATIVE_mpn_andn_n
pre ("mpn_andn_n");
mpn_andn_n (wp, xp, yp, size);
post ();
#endif
pre ("mpn_addmul_1");
mpn_addmul_1 (wp, xp, size, yp[0]);
post ();
#if HAVE_NATIVE_mpn_addmul_1c
pre ("mpn_addmul_1c");
mpn_addmul_1c (wp, xp, size, yp[0], CNST_LIMB(0));
post ();
#endif
#if HAVE_NATIVE_mpn_com_n
pre ("mpn_com_n");
mpn_com_n (wp, xp, size);
post ();
#endif
#if HAVE_NATIVE_mpn_copyd
pre ("mpn_copyd");
mpn_copyd (wp, xp, size);
post ();
#endif
#if HAVE_NATIVE_mpn_copyi
pre ("mpn_copyi");
mpn_copyi (wp, xp, size);
post ();
#endif
pre ("mpn_divexact_1");
mpn_divexact_1 (wp, xp, size, CNST_LIMB(123));
post ();
pre ("mpn_divexact_by3c");
mpn_divexact_by3c (wp, xp, size, CNST_LIMB(0));
post ();
pre ("mpn_divrem_1");
mpn_divrem_1 (wp, (mp_size_t) 0, xp, size, CNST_LIMB(123));
post ();
#if HAVE_NATIVE_mpn_divrem_1c
pre ("mpn_divrem_1c");
mpn_divrem_1c (wp, (mp_size_t) 0, xp, size, CNST_LIMB(123), CNST_LIMB(122));
post ();
#endif
pre ("mpn_gcd_1");
xp[0] |= 1;
notdead += (unsigned long) mpn_gcd_1 (xp, size, CNST_LIMB(123));
post ();
#if HAVE_NATIVE_mpn_gcd_finda
pre ("mpn_gcd_finda");
xp[0] |= 1;
xp[1] |= 1;
notdead += mpn_gcd_finda (xp);
post ();
#endif
pre ("mpn_hamdist");
notdead += mpn_hamdist (xp, yp, size);
post ();
#if HAVE_NATIVE_mpn_ior_n
pre ("mpn_ior_n");
mpn_ior_n (wp, xp, yp, size);
post ();
#endif
#if HAVE_NATIVE_mpn_iorn_n
pre ("mpn_iorn_n");
mpn_iorn_n (wp, xp, yp, size);
post ();
#endif
pre ("mpn_lshift");
mpn_lshift (wp, xp, size, 1);
post ();
pre ("mpn_mod_1");
notdead += mpn_mod_1 (xp, size, CNST_LIMB(123));
post ();
#if HAVE_NATIVE_mpn_mod_1c
pre ("mpn_mod_1c");
notdead += mpn_mod_1c (xp, size, CNST_LIMB(123), CNST_LIMB(122));
post ();
#endif
#if GMP_NUMB_BITS % 4 == 0
pre ("mpn_mod_34lsub1");
notdead += mpn_mod_34lsub1 (xp, size);
post ();
#endif
pre ("mpn_modexact_1_odd");
notdead += mpn_modexact_1_odd (xp, size, CNST_LIMB(123));
post ();
pre ("mpn_modexact_1c_odd");
notdead += mpn_modexact_1c_odd (xp, size, CNST_LIMB(123), CNST_LIMB(456));
post ();
pre ("mpn_mul_1");
mpn_mul_1 (wp, xp, size, yp[0]);
post ();
#if HAVE_NATIVE_mpn_mul_1c
pre ("mpn_mul_1c");
mpn_mul_1c (wp, xp, size, yp[0], CNST_LIMB(0));
post ();
#endif
#if HAVE_NATIVE_mpn_mul_2
pre ("mpn_mul_2");
mpn_mul_2 (wp, xp, size-1, yp);
post ();
#endif
pre ("mpn_mul_basecase");
mpn_mul_basecase (wp, xp, (mp_size_t) 3, yp, (mp_size_t) 3);
post ();
#if HAVE_NATIVE_mpn_nand_n
pre ("mpn_nand_n");
mpn_nand_n (wp, xp, yp, size);
post ();
#endif
#if HAVE_NATIVE_mpn_nior_n
pre ("mpn_nior_n");
mpn_nior_n (wp, xp, yp, size);
post ();
#endif
pre ("mpn_popcount");
notdead += mpn_popcount (xp, size);
post ();
pre ("mpn_preinv_mod_1");
notdead += mpn_preinv_mod_1 (xp, size, GMP_NUMB_MAX,
refmpn_invert_limb (GMP_NUMB_MAX));
post ();
#if USE_PREINV_DIVREM_1 || HAVE_NATIVE_mpn_preinv_divrem_1
pre ("mpn_preinv_divrem_1");
mpn_preinv_divrem_1 (wp, (mp_size_t) 0, xp, size, GMP_NUMB_MAX,
refmpn_invert_limb (GMP_NUMB_MAX), 0);
post ();
#endif
#if HAVE_NATIVE_mpn_rsh1add_n
pre ("mpn_rsh1add_n");
mpn_rsh1add_n (wp, xp, yp, size);
post ();
#endif
#if HAVE_NATIVE_mpn_rsh1sub_n
pre ("mpn_rsh1sub_n");
mpn_rsh1sub_n (wp, xp, yp, size);
post ();
#endif
pre ("mpn_rshift");
mpn_rshift (wp, xp, size, 1);
post ();
pre ("mpn_sqr_basecase");
mpn_sqr_basecase (wp, xp, (mp_size_t) 3);
post ();
pre ("mpn_submul_1");
mpn_submul_1 (wp, xp, size, yp[0]);
post ();
#if HAVE_NATIVE_mpn_submul_1c
pre ("mpn_submul_1c");
mpn_submul_1c (wp, xp, size, yp[0], CNST_LIMB(0));
post ();
#endif
pre ("mpn_sub_n");
mpn_sub_n (wp, xp, yp, size);
post ();
#if HAVE_NATIVE_mpn_sub_nc
pre ("mpn_sub_nc");
mpn_sub_nc (wp, xp, yp, size, CNST_LIMB(0));
post ();
#endif
#if HAVE_NATIVE_mpn_sublsh1_n
pre ("mpn_sublsh1_n");
mpn_sublsh1_n (wp, xp, yp, size);
post ();
#endif
#if HAVE_NATIVE_mpn_udiv_qrnnd
pre ("mpn_udiv_qrnnd");
mpn_udiv_qrnnd (&wp[0], CNST_LIMB(122), xp[0], CNST_LIMB(123));
post ();
#endif
#if HAVE_NATIVE_mpn_udiv_qrnnd_r
pre ("mpn_udiv_qrnnd_r");
mpn_udiv_qrnnd (CNST_LIMB(122), xp[0], CNST_LIMB(123), &wp[0]);
post ();
#endif
#if HAVE_NATIVE_mpn_umul_ppmm
pre ("mpn_umul_ppmm");
mpn_umul_ppmm (&wp[0], xp[0], yp[0]);
post ();
#endif
#if HAVE_NATIVE_mpn_umul_ppmm_r
pre ("mpn_umul_ppmm_r");
mpn_umul_ppmm_r (&wp[0], xp[0], yp[0]);
post ();
#endif
#if HAVE_NATIVE_mpn_xor_n
pre ("mpn_xor_n");
mpn_xor_n (wp, xp, yp, size);
post ();
#endif
#if HAVE_NATIVE_mpn_xnor_n
pre ("mpn_xnor_n");
mpn_xnor_n (wp, xp, yp, size);
post ();
#endif
}
void
mpn_kara_mul_n (mp_ptr p, mp_srcptr a, mp_srcptr b, mp_size_t n, mp_ptr ws)
{
mp_limb_t w, w0, w1;
mp_size_t n2;
mp_srcptr x, y;
mp_size_t i;
int sign;
n2 = n >> 1;
ASSERT (n2 > 0);
if ((n & 1) != 0)
{
/* Odd length. */
mp_size_t n1, n3, nm1;
n3 = n - n2;
sign = 0;
w = a[n2];
if (w != 0)
w -= mpn_sub_n (p, a, a + n3, n2);
else
{
i = n2;
do
{
--i;
w0 = a[i];
w1 = a[n3 + i];
}
while (w0 == w1 && i != 0);
if (w0 < w1)
{
x = a + n3;
y = a;
sign = ~0;
}
else
{
x = a;
y = a + n3;
}
mpn_sub_n (p, x, y, n2);
}
p[n2] = w;
w = b[n2];
if (w != 0)
w -= mpn_sub_n (p + n3, b, b + n3, n2);
else
{
i = n2;
do
{
--i;
w0 = b[i];
w1 = b[n3 + i];
}
while (w0 == w1 && i != 0);
if (w0 < w1)
{
x = b + n3;
y = b;
sign = ~sign;
}
else
{
x = b;
y = b + n3;
}
mpn_sub_n (p + n3, x, y, n2);
}
p[n] = w;
n1 = n + 1;
if (n2 < MUL_KARATSUBA_THRESHOLD)
{
if (n3 < MUL_KARATSUBA_THRESHOLD)
{
mpn_mul_basecase (ws, p, n3, p + n3, n3);
mpn_mul_basecase (p, a, n3, b, n3);
}
else
{
mpn_kara_mul_n (ws, p, p + n3, n3, ws + n1);
mpn_kara_mul_n (p, a, b, n3, ws + n1);
}
mpn_mul_basecase (p + n1, a + n3, n2, b + n3, n2);
}
else
{
mpn_kara_mul_n (ws, p, p + n3, n3, ws + n1);
mpn_kara_mul_n (p, a, b, n3, ws + n1);
mpn_kara_mul_n (p + n1, a + n3, b + n3, n2, ws + n1);
}
if (sign)
mpn_add_n (ws, p, ws, n1);
else
mpn_sub_n (ws, p, ws, n1);
nm1 = n - 1;
if (mpn_add_n (ws, p + n1, ws, nm1))
{
mp_limb_t x = (ws[nm1] + 1) & GMP_NUMB_MASK;
ws[nm1] = x;
if (x == 0)
ws[n] = (ws[n] + 1) & GMP_NUMB_MASK;
}
if (mpn_add_n (p + n3, p + n3, ws, n1))
{
mpn_incr_u (p + n1 + n3, 1);
}
}
else
{
/* Even length. */
i = n2;
do
{
--i;
w0 = a[i];
w1 = a[n2 + i];
}
while (w0 == w1 && i != 0);
sign = 0;
if (w0 < w1)
{
x = a + n2;
y = a;
sign = ~0;
}
else
{
x = a;
y = a + n2;
}
mpn_sub_n (p, x, y, n2);
i = n2;
do
{
--i;
w0 = b[i];
w1 = b[n2 + i];
}
while (w0 == w1 && i != 0);
if (w0 < w1)
{
x = b + n2;
y = b;
sign = ~sign;
}
else
{
x = b;
y = b + n2;
}
mpn_sub_n (p + n2, x, y, n2);
/* Pointwise products. */
if (n2 < MUL_KARATSUBA_THRESHOLD)
{
mpn_mul_basecase (ws, p, n2, p + n2, n2);
mpn_mul_basecase (p, a, n2, b, n2);
mpn_mul_basecase (p + n, a + n2, n2, b + n2, n2);
}
else
{
mpn_kara_mul_n (ws, p, p + n2, n2, ws + n);
mpn_kara_mul_n (p, a, b, n2, ws + n);
mpn_kara_mul_n (p + n, a + n2, b + n2, n2, ws + n);
}
/* Interpolate. */
if (sign)
w = mpn_add_n (ws, p, ws, n);
else
w = -mpn_sub_n (ws, p, ws, n);
w += mpn_add_n (ws, p + n, ws, n);
w += mpn_add_n (p + n2, p + n2, ws, n);
MPN_INCR_U (p + n2 + n, 2 * n - (n2 + n), w);
}
}
void
mpn_kara_sqr_n (mp_ptr p, mp_srcptr a, mp_size_t n, mp_ptr ws)
{
mp_limb_t w, w0, w1;
mp_size_t n2;
mp_srcptr x, y;
mp_size_t i;
n2 = n >> 1;
ASSERT (n2 > 0);
if ((n & 1) != 0)
{
/* Odd length. */
mp_size_t n1, n3, nm1;
n3 = n - n2;
w = a[n2];
if (w != 0)
w -= mpn_sub_n (p, a, a + n3, n2);
else
{
i = n2;
do
{
--i;
w0 = a[i];
w1 = a[n3 + i];
}
while (w0 == w1 && i != 0);
if (w0 < w1)
{
x = a + n3;
y = a;
}
else
{
x = a;
y = a + n3;
}
mpn_sub_n (p, x, y, n2);
}
p[n2] = w;
n1 = n + 1;
/* n2 is always either n3 or n3-1 so maybe the two sets of tests here
could be combined. But that's not important, since the tests will
take a miniscule amount of time compared to the function calls. */
if (BELOW_THRESHOLD (n3, SQR_BASECASE_THRESHOLD))
{
mpn_mul_basecase (ws, p, n3, p, n3);
mpn_mul_basecase (p, a, n3, a, n3);
}
else if (BELOW_THRESHOLD (n3, SQR_KARATSUBA_THRESHOLD))
{
mpn_sqr_basecase (ws, p, n3);
mpn_sqr_basecase (p, a, n3);
}
else
{
mpn_kara_sqr_n (ws, p, n3, ws + n1); /* (x-y)^2 */
mpn_kara_sqr_n (p, a, n3, ws + n1); /* x^2 */
}
if (BELOW_THRESHOLD (n2, SQR_BASECASE_THRESHOLD))
mpn_mul_basecase (p + n1, a + n3, n2, a + n3, n2);
else if (BELOW_THRESHOLD (n2, SQR_KARATSUBA_THRESHOLD))
mpn_sqr_basecase (p + n1, a + n3, n2);
else
mpn_kara_sqr_n (p + n1, a + n3, n2, ws + n1); /* y^2 */
/* Since x^2+y^2-(x-y)^2 = 2xy >= 0 there's no need to track the
borrow from mpn_sub_n. If it occurs then it'll be cancelled by a
carry from ws[n]. Further, since 2xy fits in n1 limbs there won't
be any carry out of ws[n] other than cancelling that borrow. */
mpn_sub_n (ws, p, ws, n1); /* x^2-(x-y)^2 */
nm1 = n - 1;
if (mpn_add_n (ws, p + n1, ws, nm1)) /* x^2+y^2-(x-y)^2 = 2xy */
{
mp_limb_t x = (ws[nm1] + 1) & GMP_NUMB_MASK;
ws[nm1] = x;
if (x == 0)
ws[n] = (ws[n] + 1) & GMP_NUMB_MASK;
}
if (mpn_add_n (p + n3, p + n3, ws, n1))
{
mpn_incr_u (p + n1 + n3, 1);
}
}
else
{
/* Even length. */
i = n2;
do
{
--i;
w0 = a[i];
w1 = a[n2 + i];
}
while (w0 == w1 && i != 0);
if (w0 < w1)
{
x = a + n2;
y = a;
}
else
{
x = a;
y = a + n2;
}
mpn_sub_n (p, x, y, n2);
/* Pointwise products. */
if (BELOW_THRESHOLD (n2, SQR_BASECASE_THRESHOLD))
{
mpn_mul_basecase (ws, p, n2, p, n2);
mpn_mul_basecase (p, a, n2, a, n2);
mpn_mul_basecase (p + n, a + n2, n2, a + n2, n2);
}
else if (BELOW_THRESHOLD (n2, SQR_KARATSUBA_THRESHOLD))
{
mpn_sqr_basecase (ws, p, n2);
mpn_sqr_basecase (p, a, n2);
mpn_sqr_basecase (p + n, a + n2, n2);
}
else
{
mpn_kara_sqr_n (ws, p, n2, ws + n);
mpn_kara_sqr_n (p, a, n2, ws + n);
mpn_kara_sqr_n (p + n, a + n2, n2, ws + n);
}
/* Interpolate. */
w = -mpn_sub_n (ws, p, ws, n);
w += mpn_add_n (ws, p + n, ws, n);
w += mpn_add_n (p + n2, p + n2, ws, n);
MPN_INCR_U (p + n2 + n, 2 * n - (n2 + n), w);
}
}
void
check_functions (void)
{
mp_limb_t wp[2], wp2[2], xp[2], yp[2], r;
int i;
memcpy (&__gmpn_cpuvec, &initial_cpuvec, sizeof (__gmpn_cpuvec));
for (i = 0; i < 2; i++)
{
xp[0] = 123;
yp[0] = 456;
mpn_add_n (wp, xp, yp, (mp_size_t) 1);
ASSERT_ALWAYS (wp[0] == 579);
}
memcpy (&__gmpn_cpuvec, &initial_cpuvec, sizeof (__gmpn_cpuvec));
for (i = 0; i < 2; i++)
{
xp[0] = 123;
wp[0] = 456;
r = mpn_addmul_1 (wp, xp, (mp_size_t) 1, CNST_LIMB(2));
ASSERT_ALWAYS (wp[0] == 702);
ASSERT_ALWAYS (r == 0);
}
#if HAVE_NATIVE_mpn_copyd
memcpy (&__gmpn_cpuvec, &initial_cpuvec, sizeof (__gmpn_cpuvec));
for (i = 0; i < 2; i++)
{
xp[0] = 123;
xp[1] = 456;
mpn_copyd (xp+1, xp, (mp_size_t) 1);
ASSERT_ALWAYS (xp[1] == 123);
}
#endif
#if HAVE_NATIVE_mpn_copyi
memcpy (&__gmpn_cpuvec, &initial_cpuvec, sizeof (__gmpn_cpuvec));
for (i = 0; i < 2; i++)
{
xp[0] = 123;
xp[1] = 456;
mpn_copyi (xp, xp+1, (mp_size_t) 1);
ASSERT_ALWAYS (xp[0] == 456);
}
#endif
memcpy (&__gmpn_cpuvec, &initial_cpuvec, sizeof (__gmpn_cpuvec));
for (i = 0; i < 2; i++)
{
xp[0] = 1605;
mpn_divexact_1 (wp, xp, (mp_size_t) 1, CNST_LIMB(5));
ASSERT_ALWAYS (wp[0] == 321);
}
memcpy (&__gmpn_cpuvec, &initial_cpuvec, sizeof (__gmpn_cpuvec));
for (i = 0; i < 2; i++)
{
xp[0] = 1296;
r = mpn_divexact_by3c (wp, xp, (mp_size_t) 1, CNST_LIMB(0));
ASSERT_ALWAYS (wp[0] == 432);
ASSERT_ALWAYS (r == 0);
}
memcpy (&__gmpn_cpuvec, &initial_cpuvec, sizeof (__gmpn_cpuvec));
for (i = 0; i < 2; i++)
{
xp[0] = 578;
r = mpn_divexact_byfobm1 (wp, xp, (mp_size_t) 1, CNST_LIMB(17),CNST_LIMB(-1)/CNST_LIMB(17));
ASSERT_ALWAYS (wp[0] == 34);
ASSERT_ALWAYS (r == 0);
}
memcpy (&__gmpn_cpuvec, &initial_cpuvec, sizeof (__gmpn_cpuvec));
for (i = 0; i < 2; i++)
{
xp[0] = 287;
r = mpn_divrem_1 (wp, (mp_size_t) 1, xp, (mp_size_t) 1, CNST_LIMB(7));
ASSERT_ALWAYS (wp[1] == 41);
ASSERT_ALWAYS (wp[0] == 0);
ASSERT_ALWAYS (r == 0);
}
memcpy (&__gmpn_cpuvec, &initial_cpuvec, sizeof (__gmpn_cpuvec));
for (i = 0; i < 2; i++)
{
xp[0] = 290;
r = mpn_divrem_euclidean_qr_1 (wp, 0, xp, (mp_size_t) 1, CNST_LIMB(7));
ASSERT_ALWAYS (wp[0] == 41);
ASSERT_ALWAYS (r == 3);
}
memcpy (&__gmpn_cpuvec, &initial_cpuvec, sizeof (__gmpn_cpuvec));
for (i = 0; i < 2; i++)
{
xp[0] = 12;
r = mpn_gcd_1 (xp, (mp_size_t) 1, CNST_LIMB(9));
ASSERT_ALWAYS (r == 3);
}
memcpy (&__gmpn_cpuvec, &initial_cpuvec, sizeof (__gmpn_cpuvec));
for (i = 0; i < 2; i++)
{
xp[0] = 0x1001;
mpn_lshift (wp, xp, (mp_size_t) 1, 1);
ASSERT_ALWAYS (wp[0] == 0x2002);
}
memcpy (&__gmpn_cpuvec, &initial_cpuvec, sizeof (__gmpn_cpuvec));
for (i = 0; i < 2; i++)
{
xp[0] = 14;
r = mpn_mod_1 (xp, (mp_size_t) 1, CNST_LIMB(4));
ASSERT_ALWAYS (r == 2);
}
#if (GMP_NUMB_BITS % 4) == 0
memcpy (&__gmpn_cpuvec, &initial_cpuvec, sizeof (__gmpn_cpuvec));
for (i = 0; i < 2; i++)
{
int bits = (GMP_NUMB_BITS / 4) * 3;
mp_limb_t mod = (CNST_LIMB(1) << bits) - 1;
mp_limb_t want = GMP_NUMB_MAX % mod;
xp[0] = GMP_NUMB_MAX;
r = mpn_mod_34lsub1 (xp, (mp_size_t) 1);
ASSERT_ALWAYS (r % mod == want);
}
#endif
// DECL_modexact_1c_odd ((*modexact_1c_odd));
memcpy (&__gmpn_cpuvec, &initial_cpuvec, sizeof (__gmpn_cpuvec));
for (i = 0; i < 2; i++)
{
xp[0] = 14;
r = mpn_mul_1 (wp, xp, (mp_size_t) 1, CNST_LIMB(4));
ASSERT_ALWAYS (wp[0] == 56);
ASSERT_ALWAYS (r == 0);
}
memcpy (&__gmpn_cpuvec, &initial_cpuvec, sizeof (__gmpn_cpuvec));
for (i = 0; i < 2; i++)
{
xp[0] = 5;
yp[0] = 7;
mpn_mul_basecase (wp, xp, (mp_size_t) 1, yp, (mp_size_t) 1);
ASSERT_ALWAYS (wp[0] == 35);
ASSERT_ALWAYS (wp[1] == 0);
}
#if HAVE_NATIVE_mpn_preinv_divrem_1 && GMP_NAIL_BITS == 0
memcpy (&__gmpn_cpuvec, &initial_cpuvec, sizeof (__gmpn_cpuvec));
for (i = 0; i < 2; i++)
{
xp[0] = 0x101;
r = mpn_preinv_divrem_1 (wp, (mp_size_t) 1, xp, (mp_size_t) 1,
GMP_LIMB_HIGHBIT,
refmpn_invert_limb (GMP_LIMB_HIGHBIT), 0);
ASSERT_ALWAYS (wp[0] == 0x202);
ASSERT_ALWAYS (wp[1] == 0);
ASSERT_ALWAYS (r == 0);
}
#endif
#if GMP_NAIL_BITS == 0
memcpy (&__gmpn_cpuvec, &initial_cpuvec, sizeof (__gmpn_cpuvec));
for (i = 0; i < 2; i++)
{
xp[0] = GMP_LIMB_HIGHBIT+123;
r = mpn_preinv_mod_1 (xp, (mp_size_t) 1, GMP_LIMB_HIGHBIT,
refmpn_invert_limb (GMP_LIMB_HIGHBIT));
ASSERT_ALWAYS (r == 123);
}
#endif
memcpy (&__gmpn_cpuvec, &initial_cpuvec, sizeof (__gmpn_cpuvec));
for (i = 0; i < 2; i++)
{
xp[0] = 5;
modlimb_invert(r,xp[0]);
r=-r;
yp[0]=43;
yp[1]=75;
mpn_redc_1 (wp, yp, xp, (mp_size_t) 1,r);
ASSERT_ALWAYS (wp[0] == 78);
}
memcpy (&__gmpn_cpuvec, &initial_cpuvec, sizeof (__gmpn_cpuvec));
for (i = 0; i < 2; i++)
{
xp[0]=5;
yp[0]=3;
mpn_sumdiff_n (wp, wp2,xp, yp,1);
ASSERT_ALWAYS (wp[0] == 8);
ASSERT_ALWAYS (wp2[0] == 2);
}
memcpy (&__gmpn_cpuvec, &initial_cpuvec, sizeof (__gmpn_cpuvec));
for (i = 0; i < 2; i++)
{
xp[0] = 0x8008;
mpn_rshift (wp, xp, (mp_size_t) 1, 1);
ASSERT_ALWAYS (wp[0] == 0x4004);
}
memcpy (&__gmpn_cpuvec, &initial_cpuvec, sizeof (__gmpn_cpuvec));
for (i = 0; i < 2; i++)
{
xp[0] = 5;
mpn_sqr_basecase (wp, xp, (mp_size_t) 1);
ASSERT_ALWAYS (wp[0] == 25);
ASSERT_ALWAYS (wp[1] == 0);
}
memcpy (&__gmpn_cpuvec, &initial_cpuvec, sizeof (__gmpn_cpuvec));
for (i = 0; i < 2; i++)
{
xp[0] = 999;
yp[0] = 666;
mpn_sub_n (wp, xp, yp, (mp_size_t) 1);
ASSERT_ALWAYS (wp[0] == 333);
}
memcpy (&__gmpn_cpuvec, &initial_cpuvec, sizeof (__gmpn_cpuvec));
for (i = 0; i < 2; i++)
{
xp[0] = 123;
wp[0] = 456;
r = mpn_submul_1 (wp, xp, (mp_size_t) 1, CNST_LIMB(2));
ASSERT_ALWAYS (wp[0] == 210);
ASSERT_ALWAYS (r == 0);
}
}
/* rp[n-1..0] = bp[bn-1..0] ^ ep[en-1..0] mod mp[n-1..0]
Requires that mp[n-1..0] is odd. FIXME: is this true?
Requires that ep[en-1..0] is > 1.
Uses scratch space at tp of 3n+1 limbs. */
void
mpn_powm_sec (mp_ptr rp, mp_srcptr bp, mp_size_t bn,
mp_srcptr ep, mp_size_t en,
mp_srcptr mp, mp_size_t n, mp_ptr tp)
{
mp_limb_t minv;
int cnt;
mp_bitcnt_t ebi;
int windowsize, this_windowsize;
mp_limb_t expbits;
mp_ptr pp, this_pp;
long i;
int cnd;
ASSERT (en > 1 || (en == 1 && ep[0] > 0));
ASSERT (n >= 1 && ((mp[0] & 1) != 0));
count_leading_zeros (cnt, ep[en - 1]);
ebi = (mp_bitcnt_t) en * GMP_LIMB_BITS - cnt;
windowsize = win_size (ebi);
binvert_limb (minv, mp[0]);
minv = -minv;
pp = tp + 4 * n;
this_pp = pp;
this_pp[n] = 1;
redcify (this_pp, this_pp + n, 1, mp, n, tp + 6 * n);
this_pp += n;
redcify (this_pp, bp, bn, mp, n, tp + 6 * n);
/* Precompute powers of b and put them in the temporary area at pp. */
for (i = (1 << windowsize) - 2; i > 0; i--)
{
mpn_mul_basecase (tp, this_pp, n, pp + n, n);
this_pp += n;
mpn_redc_1_sec (this_pp, tp, mp, n, minv);
}
expbits = getbits (ep, ebi, windowsize);
if (ebi < windowsize)
ebi = 0;
else
ebi -= windowsize;
#if WANT_CACHE_SECURITY
mpn_tabselect (rp, pp, n, 1 << windowsize, expbits);
#else
MPN_COPY (rp, pp + n * expbits, n);
#endif
while (ebi != 0)
{
expbits = getbits (ep, ebi, windowsize);
this_windowsize = windowsize;
if (ebi < windowsize)
{
this_windowsize -= windowsize - ebi;
ebi = 0;
}
else
ebi -= windowsize;
do
{
mpn_local_sqr (tp, rp, n, tp + 2 * n);
mpn_redc_1_sec (rp, tp, mp, n, minv);
this_windowsize--;
}
while (this_windowsize != 0);
#if WANT_CACHE_SECURITY
mpn_tabselect (tp + 2*n, pp, n, 1 << windowsize, expbits);
mpn_mul_basecase (tp, rp, n, tp + 2*n, n);
#else
mpn_mul_basecase (tp, rp, n, pp + n * expbits, n);
#endif
mpn_redc_1_sec (rp, tp, mp, n, minv);
}
MPN_COPY (tp, rp, n);
MPN_ZERO (tp + n, n);
mpn_redc_1_sec (rp, tp, mp, n, minv);
cnd = mpn_sub_n (tp, rp, mp, n); /* we need just retval */
mpn_subcnd_n (rp, rp, mp, n, !cnd);
}
mp_limb_t
mpn_mul (mp_ptr prodp,
mp_srcptr up, mp_size_t un,
mp_srcptr vp, mp_size_t vn)
{
mp_size_t l, k;
mp_limb_t c;
ASSERT (un >= vn);
ASSERT (vn >= 1);
ASSERT (! MPN_OVERLAP_P (prodp, un+vn, up, un));
ASSERT (! MPN_OVERLAP_P (prodp, un+vn, vp, vn));
if (un == vn)
{
if (up == vp)
{
mpn_sqr (prodp, up, un);
return prodp[2 * un - 1];
}
else
{
mpn_mul_n (prodp, up, vp, un);
return prodp[2 * un - 1];
}
}
if (vn < MUL_KARATSUBA_THRESHOLD)
{ /* plain schoolbook multiplication */
if (un <= MUL_BASECASE_MAX_UN)
mpn_mul_basecase (prodp, up, un, vp, vn);
else
{
/* We have un >> MUL_BASECASE_MAX_UN > vn. For better memory
locality, split up[] into MUL_BASECASE_MAX_UN pieces and multiply
these pieces with the vp[] operand. After each such partial
multiplication (but the last) we copy the most significant vn
limbs into a temporary buffer since that part would otherwise be
overwritten by the next multiplication. After the next
multiplication, we add it back. This illustrates the situation:
-->vn<--
| |<------- un ------->|
_____________________|
X /|
/XX__________________/ |
_____________________ |
X / |
/XX__________________/ |
_____________________ |
/ / |
/____________________/ |
==================================================================
The parts marked with X are the parts whose sums are copied into
the temporary buffer. */
mp_limb_t tp[MUL_KARATSUBA_THRESHOLD_LIMIT];
mp_limb_t cy;
ASSERT (MUL_KARATSUBA_THRESHOLD <= MUL_KARATSUBA_THRESHOLD_LIMIT);
mpn_mul_basecase (prodp, up, MUL_BASECASE_MAX_UN, vp, vn);
prodp += MUL_BASECASE_MAX_UN;
MPN_COPY (tp, prodp, vn); /* preserve high triangle */
up += MUL_BASECASE_MAX_UN;
un -= MUL_BASECASE_MAX_UN;
while (un > MUL_BASECASE_MAX_UN)
{
mpn_mul_basecase (prodp, up, MUL_BASECASE_MAX_UN, vp, vn);
cy = mpn_add_n (prodp, prodp, tp, vn); /* add back preserved triangle */
mpn_incr_u (prodp + vn, cy); /* safe? */
prodp += MUL_BASECASE_MAX_UN;
MPN_COPY (tp, prodp, vn); /* preserve high triangle */
up += MUL_BASECASE_MAX_UN;
un -= MUL_BASECASE_MAX_UN;
}
if (un > vn)
{
mpn_mul_basecase (prodp, up, un, vp, vn);
}
else
{
ASSERT_ALWAYS (un > 0);
mpn_mul_basecase (prodp, vp, vn, up, un);
}
cy = mpn_add_n (prodp, prodp, tp, vn); /* add back preserved triangle */
mpn_incr_u (prodp + vn, cy); /* safe? */
}
return prodp[un + vn - 1];
}
if (ABOVE_THRESHOLD (un + vn, 2*MUL_FFT_FULL_THRESHOLD)
&& ABOVE_THRESHOLD (3*vn, MUL_FFT_FULL_THRESHOLD))
{
mpn_mul_fft_main (prodp, up, un, vp, vn);
return prodp[un + vn - 1];
}
k = (un + 3)/4; // ceil(un/4)
#if GMP_NUMB_BITS == 32
if ((ABOVE_THRESHOLD (un + vn, 2*MUL_TOOM8H_THRESHOLD)) && (vn>=86) && (5*un <= 11*vn))
#else
if ((ABOVE_THRESHOLD (un + vn, 2*MUL_TOOM8H_THRESHOLD)) && (vn>=86) && (4*un <= 13*vn))
#endif
{
mpn_toom8h_mul(prodp, up, un, vp, vn);
return prodp[un + vn - 1];
}
if (ABOVE_THRESHOLD (un + vn, 2*MUL_TOOM4_THRESHOLD))
{
if (vn > 3*k)
{
mpn_toom4_mul(prodp, up, un, vp, vn);
return prodp[un + vn - 1];
} else
{
l = (un + 4)/5; // ceil(un/5)
if ((((vn > 9*k/4) && (un+vn <= 6*MUL_TOOM4_THRESHOLD))
|| ((vn > 2*l) && (un+vn > 6*MUL_TOOM4_THRESHOLD)))
&& (vn <= 3*l))
{
mpn_toom53_mul(prodp, up, un, vp, vn);
return prodp[un + vn - 1];
}
}
}
if (ABOVE_THRESHOLD (un + vn, 2*MUL_TOOM3_THRESHOLD) && (vn > k))
{
mp_ptr ws;
TMP_DECL;
TMP_MARK;
if (vn < 2*k) // un/2 >= vn > un/4
{
ws = TMP_ALLOC_LIMBS (MPN_TOOM3_MUL_TSIZE(un));
mpn_toom42_mul(prodp, up, un, vp, vn, ws);
TMP_FREE;
return prodp[un + vn - 1];
}
l = (un+2)/3; //ceil(u/3)
if (vn > 2*l) // un >= vn > 2un/3
{
ws = TMP_ALLOC_LIMBS (MPN_TOOM3_MUL_TSIZE(un));
mpn_toom3_mul(prodp, up, un, vp, vn, ws);
TMP_FREE;
return prodp[un + vn - 1];
} else // 2un/3 >= vn > un/3
{
ws = TMP_ALLOC_LIMBS (MPN_TOOM3_MUL_TSIZE(un));
mpn_toom32_mul(prodp, up, un, vp, vn, ws);
TMP_FREE;
return prodp[un + vn - 1];
}
}
mpn_mul_n (prodp, up, vp, vn);
if (un != vn)
{ mp_limb_t t;
mp_ptr ws;
TMP_DECL;
TMP_MARK;
prodp += vn;
l = vn;
up += vn;
un -= vn;
if (un < vn)
{
/* Swap u's and v's. */
MPN_SRCPTR_SWAP (up,un, vp,vn);
}
ws = TMP_ALLOC_LIMBS ((vn >= MUL_KARATSUBA_THRESHOLD ? vn : un) + vn);
t = 0;
while (vn >= MUL_KARATSUBA_THRESHOLD)
{
mpn_mul_n (ws, up, vp, vn);
if (l <= 2*vn)
{
t += mpn_add_n (prodp, prodp, ws, l);
if (l != 2*vn)
{
t = mpn_add_1 (prodp + l, ws + l, 2*vn - l, t);
l = 2*vn;
}
}
else
{
c = mpn_add_n (prodp, prodp, ws, 2*vn);
t += mpn_add_1 (prodp + 2*vn, prodp + 2*vn, l - 2*vn, c);
}
prodp += vn;
l -= vn;
up += vn;
un -= vn;
if (un < vn)
{
/* Swap u's and v's. */
MPN_SRCPTR_SWAP (up,un, vp,vn);
}
}
if (vn != 0)
{
mpn_mul_basecase (ws, up, un, vp, vn);
if (l <= un + vn)
{
t += mpn_add_n (prodp, prodp, ws, l);
if (l != un + vn)
t = mpn_add_1 (prodp + l, ws + l, un + vn - l, t);
}
else
{
c = mpn_add_n (prodp, prodp, ws, un + vn);
t += mpn_add_1 (prodp + un + vn, prodp + un + vn, l - un - vn, c);
}
}
TMP_FREE;
}
return prodp[un + vn - 1];
}
void
mpn_sqr (mp_ptr p, mp_srcptr a, mp_size_t n)
{
ASSERT (n >= 1);
ASSERT (! MPN_OVERLAP_P (p, 2 * n, a, n));
if (BELOW_THRESHOLD (n, SQR_BASECASE_THRESHOLD))
{ /* mul_basecase is faster than sqr_basecase on small sizes sometimes */
mpn_mul_basecase (p, a, n, a, n);
}
else if (BELOW_THRESHOLD (n, SQR_TOOM2_THRESHOLD))
{
mpn_sqr_basecase (p, a, n);
}
else if (BELOW_THRESHOLD (n, SQR_TOOM3_THRESHOLD))
{
/* Allocate workspace of fixed size on stack: fast! */
mp_limb_t ws[mpn_toom2_sqr_itch (SQR_TOOM3_THRESHOLD_LIMIT-1)];
ASSERT (SQR_TOOM3_THRESHOLD <= SQR_TOOM3_THRESHOLD_LIMIT);
mpn_toom2_sqr (p, a, n, ws);
}
else if (BELOW_THRESHOLD (n, SQR_TOOM4_THRESHOLD))
{
mp_ptr ws;
TMP_SDECL;
TMP_SMARK;
ws = TMP_SALLOC_LIMBS (mpn_toom3_sqr_itch (n));
mpn_toom3_sqr (p, a, n, ws);
TMP_SFREE;
}
else if (BELOW_THRESHOLD (n, SQR_TOOM6_THRESHOLD))
{
mp_ptr ws;
TMP_SDECL;
TMP_SMARK;
ws = TMP_SALLOC_LIMBS (mpn_toom4_sqr_itch (n));
mpn_toom4_sqr (p, a, n, ws);
TMP_SFREE;
}
else if (BELOW_THRESHOLD (n, SQR_TOOM8_THRESHOLD))
{
mp_ptr ws;
TMP_SDECL;
TMP_SMARK;
ws = TMP_SALLOC_LIMBS (mpn_toom6_sqr_itch (n));
mpn_toom6_sqr (p, a, n, ws);
TMP_SFREE;
}
else if (BELOW_THRESHOLD (n, SQR_FFT_THRESHOLD))
{
mp_ptr ws;
TMP_DECL;
TMP_MARK;
ws = TMP_ALLOC_LIMBS (mpn_toom8_sqr_itch (n));
mpn_toom8_sqr (p, a, n, ws);
TMP_FREE;
}
else
{
/* The current FFT code allocates its own space. That should probably
change. */
mpn_fft_mul (p, a, n, a, n);
}
}
/* (rp, 2n) = (xp, n)*(yp, n) */
static void
mpn_mulshort_n(mp_ptr rp, mp_srcptr xp, mp_srcptr yp, mp_size_t n)
{
mp_size_t m;
mp_limb_t t;
mp_ptr rpn2;
ASSERT(n >= 1);
ASSERT_MPN(xp, n);
ASSERT_MPN(yp, n);
ASSERT(!MPN_OVERLAP_P (rp, 2 * n, xp, n));
ASSERT(!MPN_OVERLAP_P (rp, 2 * n, yp, n));
if (BELOW_THRESHOLD(n, MULHIGH_BASECASE_THRESHOLD))
{
mpn_mul_basecase(rp, xp, n, yp, n);
return;
}
if (BELOW_THRESHOLD (n, MULHIGH_DC_THRESHOLD))
{
mpn_mulshort_n_basecase(rp, xp, yp, n);
return;
}
/* choose optimal m s.t. n + 2 <= 2m, m < n */
ASSERT (n >= 4);
m = 87 * n / 128;
if (2 * m < n + 2)
m = (n + 1) / 2 + 1;
if (m >= n)
m = n - 1;
ASSERT (n + 2 <= 2 * m);
ASSERT (m < n);
rpn2 = rp + n - 2;
mpn_mul_n (rp + n - m + n - m, xp + n - m, yp + n - m, m);
mpn_mulshort_n (rp, xp, yp + m, n - m);
ASSERT_NOCARRY (mpn_add (rpn2, rpn2, n + 2, rpn2 - m, n - m + 2));
mpn_mulshort_n (rp, xp + m, yp, n - m);
ASSERT_NOCARRY (mpn_add (rpn2, rpn2, n + 2, rpn2 - m, n - m + 2));
umul_ppmm (rp[1], t, xp[m - 1], yp[n - m - 1] << GMP_NAIL_BITS);
rp[0] = t >> GMP_NAIL_BITS;
ASSERT_NOCARRY (mpn_add (rpn2, rpn2, n + 2, rp, 2));
umul_ppmm (rp[1], t, xp[n - m - 1], yp[m - 1] << GMP_NAIL_BITS);
rp[0] = t >> GMP_NAIL_BITS;
ASSERT_NOCARRY (mpn_add (rpn2, rpn2, n + 2, rp, 2));
return;
}
